Accepted Submissions (đ€ = oral)
("#n" corresponds to the submission ID in CMT)
#3 | Freeform Preference Learning for Robotic Manipulation
Marcel Torne Villasevil (Stanford)*; Anubha Mahajan (Stanford); Abhijnya Bhat (Stanford); Chelsea Finn (Stanford)
Abstract: Reward design remains a central bottleneck for autonomous robot policy improvement, especially in long-horizon manipulation tasks where sparse success labels provide too little signal and binary preferences collapse many competing notions of quality into one ambiguous signal. We introduce Freeform Preference Learning (FPL), a method for learning robot policies from freeform human preferences. Rather than asking annotators which of two trajectories is better overall, FPL lets them define natural-language preference axes, such as speed, safety, quality of placement, or carefulness, and provide pairwise preferences along each axis. These annotations are used to learn a language-conditioned reward model that maps a trajectory and preference label to an axis-specific reward. We use this model to train a reward-conditioned policy that optimizes across the multiple human-specified dimensions. Across four real-world and two simulated long-horizon manipulation tasks, FPL improves over sparse-reward and binary-preference methods by 38 percentage points. Beyond improved performance, FPL learns dense progress signals without explicit subtask segmentation, shows compositionality of behavior not present in the data, and allows users to steer the policy towards different behaviors at test time without retraining.
#5 | Mechanistic Specialization Does Not Guarantee Performance: Evidence from Dual Attention Transformers
Jonathan Scott (University of California, Santa Cruz)*
Abstract: Dual Attention Transformers (DATs) extend decoder-only Transformers with a dedicated relational-attention stream, making them a natural architecture for abstract identity rules such as ABA and ABB. Surprisingly, we find that comparably sized GPT-2 models outperform DATs on these tasks. We investigate this gap with two complementary mechanistic analyses. First, causal mediation analysis shows that DATs exhibit stronger evidence of hypothesized symbolic mechanisms: symbol abstraction, symbol induction, and retrieval, than GPT-2. Second, a routing analysis shows why this specialization does not translate into better behavior: DATs make more wrong-copy errors, can attend to the correct source token while still predicting the wrong token, and show weak direct contribution from relational attention to the correct-versus-wrong output margin. Ablating positive-routing heads hurts performance, while amplifying those heads improves DAT more than matched controls. These results show that explicit relational inductive biases can shape internal organization without guaranteeing task success.
#6 | Behavior-Based User Segmentation via Listwise Learning-to-Rank Trees
Yang Liu (Meta)*
Abstract: We present Behavior-based User Segmentation (BUS), a novel tree-based data structure that hierarchically partitions users guided by product-specific engagement behaviors to generate interpretable user segments for recommendation. BUS reformulates user attribute-based recommendation as a listwise learning-to-rank problem, optimizing Normalized Discounted Cumulative Gain (NDCG) to ensure aggregated active-user behaviors effectively represent marginal (cold-start) users. A greedy construction algorithm with a novel regress operator filters irrelevant attributes at each iteration, guaranteeing monotonic loss reduction. We further extend BUS with connection-aware segmentation via social graphs to improve fairness and diversity. Deployed at industrial scale serving billions of users daily, BUS achieves significant online improvements in music ranking (+0.126% daily active producers) and email notification click-through rates (+2.95%). To our knowledge, BUS is the first listwise learning-to-rank framework for tree-based recommendation that integrates diverse categorical attributes with semantic interpretability at industrial scale.
#8 | Many Minds from One Model: Bayesian-Inspired Transformers for Population Diversity
Diji Yang (University of California Santa Cruz)*; Yi Zhang (University of California Santa Cruz)
Abstract: Despite their scale and success, modern transformers are usually trained as single-minded systems: optimization produces a deterministic set of parameters, representing a single hypothesis about the data. We propose Population Bayesian Transformers (B-Transformers), which enable sampling diverse yet coherent transformer large language model instances, referred to as "minds", from a single pre-trained LLM. B-Transformers introduces a Bayesian-inspired posterior proxy by injecting stochasticity directly into normalization layers, avoiding the prohibitive cost of full Bayesian neural networks or multi-model ensembles. During the generation of each response, we sample a single realization and hold it fixed, ensuring temporal consistency and reasoning coherence. Experiments on zero-shot generation and Reinforcement Learning with Verifiable Rewards (RLVR) show greater semantic diversity, additional valid reasoning paths, and consistently improved RLVR performance compared with fixed-parameter baselines.
#10 | Interpretable symptom vectors for depression in a large language model
Fangyi Zhu (Stanford Precision Neurotherapeutics Lab)*; Ajay Subramanian (Stanford Precision Neurotherapeutics Lab); Allison Constant (Kaiser Permanente); Camille Wang ( Stanford Precision Neurotherapeutics Lab); Ravish Gupta (BigCommerce); Corey Keller (Stanford Precision Neurotherapeutics Lab)
Abstract: Depression manifests in divergent symptom profiles, yet patients meeting the same diagnostic criteria are routinely reduced to a single severity score that discards this complexity. Large language models (LLMs) could recover this heterogeneity from patient speech, but how depressive symptoms are represented inside LLMs remains poorly understood, limiting clinical trust in their outputs. To examine the modelâs internal activations, we analyzed the residual stream of Gemma-3-27b-pt using mechanistic interpretability techniques. We recorded residual stream activations as the model processed symptom descriptions drawn from validated clinical instruments and showed geometrical separation between symptom groups peaked at layer 21 across several distance metrics. Using Semantic Projection, we then projected held-out naturalistic text onto symptom vectors built from those instruments. The resulting per-symptom coefficients preserved clinician-annotated rank ordering across mood, somatic, and suicidality axes. These results reveal a decorrelated, clinician-aligned symptom signal readable directly from model activations, offering a mechanistic foundation for interpretable, privacy-preserving depression-assessment tools.
#12 | Reduced-Rank Regression trees: a parsimonious model for nonlinear multi-output regression
Kuat Gazizov (UC Merced)*; Miguel Carreira-Perpiñån (UC Merced)
Abstract: Decision trees with linear leaves construct piecewise-linear mappings and have long been recognized as powerful tools for nonlinear regression. However, they produce large models with many parameters and can easily overfit. We propose the Reduced-Rank Regression tree (RRR tree), a model that explicitly enforces a low-rank structure on the linear mapping at each leaf. Besides reducing the number of parameters considerably, this has the advantage of introducing a form of regularization that improves the generalization ability of the model, and it makes it possible to learn deeper trees that adapt better to nonlinear data manifolds. It also capitalizes on the fact that the rank of the mapping can be globally high but locally low, which is a significant improvement over the traditional, linear RRR model. To train an RRR tree, we introduce an algorithm based on alternating optimization, which jointly learns the tree-structured partition of the space and the low-rank leaf parameters. This approach scales with respect to data dimensionality and sample size. In extensive experiments, we demonstrate that RRR trees achieve strong predictive performance while producing models that are orders of magnitude smaller and faster for inference than established baselines.
#13 | The SpecâTest Coverage Gap: Quantifying Test-Level Overspecification in Issue-Resolution Benchmarks
Avi Aggarwal (Purdue)*
Abstract: SWE-bench and similar benchmarks decide whether an agent fixed a bug by running a test patch that the maintainer wrote alongside the official fix. Because that test was written by someone who already knew how they intended to fix the bug, it often checks more than the issue asked for: a specific error string, a specific return value from an internal helper, or a code path the issue never mentions. An agent that fixes the reported bug in a reasonable way can still be scored as a failure. Prior work has studied issues that are too vague; we study the other side, tests that are too specific. We define the SpecâTest Coverage Gap (STCG), a per-task rating of how much of what a test checks could have been worked out from the issue text alone. STCG is a four-level rubric, running from behavior the issue states outright to assertions that lock in one particular implementation, applied by an LLM judge that reads the issue and the test patch but never the official code patch. On 12 hand-labeled SWE-bench tasks drawn from six repositories and stratified across the four levels, a Gemini-2.5-Pro judge reaches 0.71 quadratic weighted kappa and lands within one level of the expert label on 11 of 12 tasks. We treat this as a feasibility check, and set out what would be needed to use STCG to audit a benchmark.
#15 | An Expectation-Maximization Perspective on Reinforcement Learning for LLM Reasoning
Tianbing Xu (Waymo)*
Abstract: Reinforcement learning has become a central tool for improving the reasoning abilities of large language models, yet widely used methods such as PPO and GRPO rely on likelihood-ratio correction, clipping, and other stabilization mechanisms. We ask whether reasoning models can instead be trained from a simpler optimization principle. We show that reinforcement learning over latent reasoning trajectories naturally admits an Expectation-Maximization~\citep{em} formulation. The E-step constructs a reward-weighted distribution over sampled rationales, while the M-step performs reward-guided policy optimization. This perspective naturally yields a simple off-policy policy gradient that requires neither trajectory-level importance weighting nor clipping, and it admits a monotonic-improvement guarantee under exact E- and M-steps. On GSM8K and MATH Hard, EM Policy Gradient achieves performance comparable to or better than GRPO across multiple Qwen2.5 model scales. The results suggest that latent-variable optimization offers a simple and theoretically grounded alternative foundation for reinforcement learning in reasoning language models.
#16 | RideWay: Benchmarking Efficient Task Completion for Tool-Using Language Agents
Qingnuan Han (Didi Global); Boli Fang (Didi Labs)*; Mingzhi Hou (Didi Global); Claire Liu (Didi Global)
Abstract: Task success is necessary for service agents, but it is not sufficient: an agent that completes a request through avoidable questions, retries, or tool calls still imposes unnecessary cost. We introduce RideWay, a Chinese-language ride-hailing benchmark for evaluating efficient task completion. RideWay combines stateful tools, a fixed user simulator, rubric-based completion labels, human reference trajectories, and human preferences over successful agent trajectories. Its utility score gives no credit to failed attempts; among successful attempts, it rewards agents that reach the requested ride state with fewer avoidable assistant turns and tool calls. Experiments show that agents with similar success rates can differ substantially in interaction cost, and human preference studies support RideWay's metric as a practical way to evaluate not only whether agents complete service tasks, but how efficiently they do so.
#17 | User Journey Inference with LLMs
Manan Kalra (Pinterest)*; Lin Zhu (Pinterest)
Abstract: Personalizing for hundreds of millions of Pinterest users hinges on inferring their evolving, long-term goals â their "journeys". It is a critical challenge in modern recommendation systems, as traditional systems decompose the problem into keyword extraction, embedding, clustering, and separate attribute models, causing noise to compound and coherent goals to fragment. Our central idea is to treat journey inference as a single generative act: one Large Language Model (LLM) reads a user's raw, multi-source behavioral activity sequence and directly emits structured, multi-granularity journeys. By aggressively fine-tuning open-source LLMs (e.g., Qwen3-4B) and deploying them via a highly optimized Dynamo + vLLM serving architecture, we achieve scalable, real-time journey inference. Offline quality (BLEU/ROUGE, LLM-judge) saturates after only a few thousand examples, letting the small student rival its frontier teachers at production cost. Across online experiments, the system lifted push-notification opens by +1.3% and email click-throughs by +1.1%, with significant sitewide gains in user engagement and retention. A production recipe makes this viable at scale: automated prompt optimization (GEPA), Semantic IDs to shrink input tokens, and multilingual, region-aware serving. We further align the model using feedback-as-edit DPO pairs derived from YES/NO responses. The recipe is domain-agnostic: any platform with sequential user activity and structured intent targets (e-commerce, streaming, search) can adopt the same activity-to-intent formulation, collapsing classical multi-stage recommendation pipelines into a single generative model, making generative journey inference a general recipe for personalization beyond Pinterest.
#18 | GUIDE: Governed Unified Intelligence for Document-to-Artifact Generation in Enterprise Settings
Mohammadreza Sediqin (Centific)*; Shivali Dalmia (Centific); Sumukha Thoppanahalli (Centific); Srinivasa Karthikeya Reddy Kovvuri (Centific); Abhishek Mukherji (Centific)
Abstract: Converting enterprise guideline documents, which mix prose, complex tables, and embedded images, into operational artifacts remains manual expert work taking days per document. We treat this as a data management problem and present a multi-agent pipeline built around a versioned, schema-validated rule store with end-to-end provenance, in which agents parse multimodal content, normalize rules into a fixed schema, check consistency, evaluate quality in two stages, and escalate uncertain objects to persona-matched reviewers. On 120 enterprise guideline documents, the pipeline cuts an expert-estimated two to three days per document to one to two hours including review, extracting 3,896 rules with 3.2% hallucination and 71.4% approved without human edits, while removing the governance layers with the identical base model raises hallucination nearly fivefold and degrades every quality dimension at once. Our results show that reliability in document intelligence depends as much on governance infrastructure as on the model.
#19 | MCEK: Maintaining Consistency in Evolving Knowledge
Mohammadreza Sediqin (Centific)*; Shivali Dalmia (Centific); Sumukha Thoppanahalli (Centific); Abhishek Mukherji (Centific)
Abstract: Language models increasingly depend on external knowledge stores that outgrow the context window, yet as these stores evolve through revisions, reuploads, and extensions, the same knowledge reappears in new wording and newer entries quietly refine or contradict older ones. Retrieval-augmented and long-context methods query such stores assuming internal coherence, but nothing in the pipeline actually maintains it. To the best of our knowledge, we present the first framework to unify version tracking and rule-level consistency maintenance as a single memory-management problem: incoming documents are assigned to version families through centroid retrieval and bidirectional rule alignment, and their rules are then classified against family memory using similarity retrieval, bidirectional natural language inference, and Counterfactual Span Probing, a training-free method that recovers duplicates and refinements hidden inside the neutral class. Relation-specific policies handle routine updates automatically, routing only contradictions and refinements to human review. On an expert-annotated enterprise corpus, the framework reaches F1 0.97 for version grouping and 0.94 for rule consistency, fully locally without any large language model.
#21 | Learning Spatial Immune Architecture from Routine H&E Predicts Histologic Subtype and Survival in NSCLC
Prisha Bhanushali (None)*
Abstract: Predicting immunotherapy response in non-small cell lung cancer (NSCLC) is an open problem: only ~20% of patients respond, and the standard PD-L1 biomarker counts marker-positive cells while ignoring the spatial organization of the tumor immune microenvironment. We ask whether clinically meaningful spatial signal can be recovered from the H&E whole-slide images already collected for every patientâwithout transcriptomics or specialized staining. Our pipeline combines (1) deep-learning nucleus segmentation (StarDist) over ~500,000 cells across 395 regions from 82 NSCLC patients; (2) extraction of 19 per-cell morphological, intensity, and spatial features; (3) tumor/immune/stromal classification validated against expert annotations (up to 50.7% tumor co-localization); and (4) region-level spatial descriptorsâtumor-to-immune distance, local entropy, cell densityâplus a composite immune-architecture score. These features discriminate histologic subtypes: adenocarcinoma and squamous differ in cell density (p=0.001) and entropy (p=0.003), with squamous skewing immune-hot (composite p=0.006). Transferring the identical pipeline to an independent, outcome-linked cohort (TCGA LUAD/LUSC, n=14), Cox proportional-hazards modeling identifies tumor-to-immune distance as the only significant predictor of overall survival (HR=0.31, p=0.006, concordance=0.81), with greater distance predicting longer survival. This direction is opposite to the "immune-hot is favorable" prior, suggesting proximity may capture exhausted rather than active T cellsâa signal invisible to bulk transcriptomics. Interpretable spatial features from routine images thus yield subtype- and survival-relevant predictions: a low-cost, scalable direction for computational pathology. We note limitations (small survival cohort, rule-based classification) and next steps toward learned classifiers and larger validation.
#22 | From Fear to Fuel: Resolving the AI Paradox Through Managerial Oversight and Employee Resilience
Aurora Zamora (Saint Mary's College of California)*
Abstract: Organizations increasingly deploy artificial intelligence (AI) to augment knowledge work, yet employee responses remain highly variable despite similar governance practices. Existing computational approaches provide limited ability to model how employee perceptions evolve during AI adoption because they often lack contextual reasoning and adaptive social behavior. This research develops the Threat Crossover Framework, a theory-constrained large language model (LLM)-based organizational simulation that integrates organizational theory with generative agents to examine how perceived AI-related job displacement threat influences employee discomfort, trust, and strategic collaboration. The research follows a three-phase design. A completed agent-based simulation identified a Threat Crossover effect in which managerial oversight reduced employee discomfort and increased organizational trust when perceived job displacement threat was low but became substantially less effective as perceived threat intensified. Building on these findings, the proposed framework extends the simulation with theory-constrained LLM agents capable of contextual reasoning, organizational memory, and adaptive behavior while remaining grounded in organizational theory. A planned survey of 100 employees will provide an empirical assessment of the framework. The Threat Crossover Framework contributes a computational approach for modeling human-centered AI adoption by combining organizational theory with theory-constrained LLM agents to simulate complex socio-technical systems. The framework demonstrates how generative agents can complement traditional agent-based modeling while providing practical insight into the conditions under which managerial interventions succeed or fail during organizational AI adoption.
#23 | From Loop Detectors to Connected Vehicles: Hard-Braking Crash Prediction on Real Trajectories
Grigorios Grigoropoulos (Santa Clara University, Computer Science & Engineering)*; David Anastasiu (Santa Clara University, Computer Science & Engineering)
Abstract: Hard-braking events (HBEs) from connected vehicles are correlated with segment-level crash rates and have been proposed as an abundant surrogate for sparse crash data. We test this on two different inputs. In an initial experiment, a HBE proxy built from aggregated five-minute loop-detector speeds (METR-LA) could not establish the association: its coefficient is statistically almost zero even with a 40-fold increase in crashes. We then built a real-data mechanism: a reproducible pipeline that processes 80 million connected-vehicle readings from the Ann Arbor Safety Pilot to detect 111,354 true HBEs, links them to 101,014 police-reported crashes and 27,444 road segments through the state linear-referencing system and fits Highway Safety Manual Negative-Binomial models. The association is now clearly evident: segments with observed hard-braking carry 1.44x the crash rate of those without, and among HBE segments the crash rate rises with HBE frequency (incidence-rate ratio 1.28 per e-fold, p < 10â»Âł. It remains stable when a minimum amount of instrumented travel is required, and gets stronger on the injury-and-fatal subset. The result is clear: the modelling approach works on both datasets, but the HBE signal needs real kinematics that aggregated speeds cannot provide.
#24 | Deep Reinforcement Learning for ISAC Enabled Real Time Sporting Intelligence over Private 6G Networks
Derek Chui (Santa Clara University)*; Ramon Anaya (Santa Clara University); Collin Fiske (Santa Clara University); Monica Reyes (Santa Clara University); Pedro Aguirre-Salas (Santa Clara University); Krishna Murthy Kattiyan Ramamoorthy (Santa Clara University)
Abstract: Dense real-time environments, i.e., an ongoing stadium-hosted professional football match, usually flood the network with competing data streams. These data feeds can include player GPS, wearable telemetry, tracking cameras, aerial sensing, and more. While a conventional wireless system can allocate capacity by throughput or channel quality alone, it is blind to which stream is decision-relevant at a given instant. Thus, we investigate this as a learning problem for Pinching Antenna System (PASS)-enabled Non-Orthogonal Multiple Access (NOMA)-assisted Integrated Sensing and Communication (ISAC), where data-transmitting infrastructures must simultaneously communicate with and localize many fast-moving users while deciding transmission priority. Instead of a fixed heuristic, we propose a deep reinforcement learning-enabled allocation policy leveraging Deep Q Network (DQN/DDQN). Preliminary results on an analytical PASS-NOMA model reveal a congestion-dependent tradeoff. With ample spectrum, throughput-maximizing and relevance-aware allocation policies nearly coincide, but under congestion, exhaustive relevance delivers more decision-critical utility per served user. The learned policy recovers its grouping decisions in a single forward pass, with the remaining gap attributable to an undertrained power split head.
#25 | Assessing the LatencyâAccuracy Trade-off for LLM-driven ASR Correction
Alan Fung (Santa Clara University); Tokunbo Ogunfunmi (Santa Clara University)*
Abstract: LLM-based post-processing is a promising way to reduce ASR error rates without retraining the underlying speech model, and prior work shows its relative benefit grows as the baseline error rate grows â exactly the regime low-resource languages fall into. To study this trade-off, we use Whisper transcripts of FLEURS languages as a testbed, iteratively design a constrained correction prompt, and evaluate it both with a frontier model (Claude Sonnet 5) on high-error samples and with a locally-hosted 14B-parameter model (Qwen2.5-14B-Instruct) at a small scale. Careful prompt constraints can nearly halve CER on a high-erro Mandarin sample with the frontier model, but the same approach run locally shows no reliable net benefit and often violates output-format instructions, all while adding roughly 10Ă latency versus the base ASR pipeline. We argue this makes local LLM post-processing a questionable choice: the promised accuracy gains are largely gated behind frontier-scale models, which may not be available in the latency- and cost-aware environments where low-resource ASR is needed.
#26 | đ€ Neuro-symbolic Policy Iteration for Computer Agents
Minh Nguyen (Simular); Chih-Lun Lee (Simular); Lizhao Liu (Simular); Ang Li (Simular); Jiachen Yang (Simular)*
Abstract: Automating repetitive computer tasks requires a computer agent that succeeds reliably across many executions, at low latency and cost. However, agents that rely purely on large language models (LLM) for planning are stochastic, slow, and expensive, while robotic process automation scripts are fast and cheap but require engineering effort and are brittle to environment noise. We propose a neuro-symbolic policy iteration method that constructs and iteratively refines an executable code plan using LLM judges to optimize \passk{k} success rate. Evaluation on OSWorld and ScienceBoard shows that our method achieves significantly higher reliability than pure LLM-based agents, reduces cost by more than an order of magnitude, and speeds up task completion by $\sim$3x.
#27 | LLMON: An LLM-native Markup Language to Leverage Structure and Semantics at the LLM Interface
Michael Hind (IBM Research); Basel Shbita (IBM Research)*; Bo Wu (IBM Research); Farhan Ahmed (IBM Research); Chad DeLuca (IBM Research); Nathan Fulton ( IBM Research); David Cox (IBM Research); Dan Gutfreund (IBM Research)
Abstract: Textual Large Language Models (LLMs) provide a simple and familiar interface: a string of text is used for both input and output. However, the information conveyed to an LLM often has a richer structure and semantics, which is not conveyed in a string. For example, most prompts contain both instructions ("Summarize this paper into a paragraph") and data (the paper to summarize), but these are usually not distinguished when passed to the model. This can lead to model confusion and security risks, such as prompt injection attacks. We address this shortcoming by introducing an LLM-native markup language, LLMON (LLM Object Notation), that enables the structure and semantic metadata of the text to be communicated in a natural way to an LLM. Post-training three 3B base models on an 8.4B-token LLMON corpus improves a distractor-robustness benchmark by up to 40 points while matching or improving general capability. These results suggest that making prompt structure explicit is a practical lever for reliability at the LLM interface.
#28 | CoVA-SFT: A Large-Scale Dataset for\\Chain of Visual Abstractions
Tsung-Han Wu (UC Berkeley)*; Heekyung Lee (UC Berkeley); Anya Ji (UC Berkeley); Haoming Chen (UC Berkeley); Trevor Darrell (UC Berkeley); Joseph E. Gonzalez (UC Berkeley); David M. Chan (UC Berkeley)
Abstract: Long chain-of-thought (CoT) reasoning has produced large gains in mathematical problem solving, coding, and other complex tasks, becoming a foundational interface that allows large language models (LLMs) to decompose problems and write intermediate steps before producing an answer. While this interface is effective when the relevant intermediate state is naturally linguistic, it does not generalize easily to non-textual, structured states that are inefficient to serialize as text. Recent work has challenged the assumption that intermediate reasoning must be expressed only in language, extending it with sketches, generated images, code-rendered plots, and interleaved vision-language traces. Despite these advances, the community lacks a large-scale, multi-step, self-corrected dataset for teaching models to build internal visual workspaces. To address this gap, we introduce CoVA-SFT, a structured corpus for teaching multimodal large language models to interleave text and visual abstractions.
#32 | Operational Weather Forecasts Reverse the Ranking Between Time-Series Foundation Models and Specialist Load Forecasters
Jiawei Lin (University of Leeds ); Yu Zhang (UC Santa Cruz)*
Abstract: Load forecasting models that leverage future weather are almost universally evaluated using ground-truth weather, despite real-world deployments relying exclusively on imperfect numerical weather forecasts. This mismatch can substantially distort comparisons between task-specific forecasting models and time-series foundation models (TSFMs). We introduce a benchmark for evaluating load forecasting under realistic operational weather forecasts by replaying archived HRRR forecasts at two operational lead times across five major U.S. power grid operators. We compare Chronos-2, TimesFM 2.5, and a state-of-the-art attention-based specialist load forecaster using 534 paired-bootstrap evaluation windows. Under perfect weather conditions, the specialist model achieves the lowest errors. Under day-ahead operational forecasts, however, the fine-tuned foundation model exhibits greater robustness, degrading less than the specialist, reaching statistical parity on ISO-NE and CAISO, and surpassing the specialist on the most weather-sensitive grid ERCOT. Our findings highlight the importance of benchmarking load forecasting models under operational covariate conditions rather than idealized weather observations.
#34 | Pose-Guided Jersey Numbers Recognition for Player Identity Anchoring in Sports Video
Ryan Renslow (Eluvio)*
Abstract: Multi-object trackers label players with arbitrary indices that persist as long as the track itself. Broadcast sports demand more: identities that survive occlusion, camera cuts, and re-entry. We argue that high-precision identity anchors, like jersey numbers, can initialize and verify tracks even if there are few frames of legibility. However, current OCR systems are not optimized for many sports scenarios to accurately recognize jersey numbers, and are not designed to handle the many frames in which the number is illegible. Thus, precision is the priority because a single incorrect anchor could propagate a false identity through every frame that follows. We tested several systems on 4,918 consensus-labeled jersey crops spanning soccer, basketball, and American football, using both public and in-house annotated datasets. To define a ground truth for testing, each crop was reviewed by three or more independent labellers. After comparing the results of multiple current OCR methods, we developed a new pipeline that uses a series of models: person detection and segmentation using YOLO26, legibility gating using ResNet-34, pose-guided torso cropping using YOLO26, and finally, OCR recognition using PP-OCRv6. We show that multiple models working on individual problems in series for OCR outperform single isolated models.
#35 | Reasoning Is Not Safety: Small Open Reasoning Models Are the Least Safe Systems in a Cross-Benchmark, Cross-Evaluator Audit
Ehsan Latif (Cognizant)*
Abstract: It is often assumed that stronger reasoning makes a language model safer: the premise of ``reasoning-improves-safety'' arguments such as deliberative alignment and the o1 system card. We test this assumption by auditing nine models that factor apart scale/access, alignment, and reasoning on two current safety suites, AILuminate~v1.0 and SORRY-Bench, with multiple exemplars per class (1.5--3B) so findings are not single-model artifacts. We evaluate three large closed frontier models (Claude Opus~4.8, Claude Sonnet~5, GPT-5.5); three safety-tuned small open \emph{non-reasoning} models (Qwen2.5-1.5B/3B, Llama-3.2-3B); and three small open \emph{reasoning} models (DeepSeek-R1-1.5B, VibeThinker-1.5B/3B). Under a unified LLM judge, corroborated by each benchmark's \emph{own} evaluator (SORRY-Bench's fine-tuned Mistral-7B judge; AILuminate's ModelBench grade via the open LlamaGuard-2 annotator), frontier models are safest and the two \emph{unaligned} 1.5B reasoning models are the least safe systems tested ($\sim$54--58\% unsafe vs.\ 2--5\% frontier and 6--22\% for the non-reasoning open models), with harmful content in both the chain-of-thought and the final answer. Crucially, a \emph{better-aligned} reasoning model (VibeThinker-3B) is far safer ($\sim$15--18\%): the safety \emph{level} is set by alignment, not by whether the model reasons. To remove the size/lineage confound in that cross-section, we add a \emph{same-weights} control, Qwen3 (0.6--8B) with its reasoning mode toggled on vs.\ off, so reasoning is the only variable, and find that enabling reasoning raises the unsafe rate at \emph{every} scale under all three evaluators (answer-level $+1.7$ to $+9.1$ points; a full AILuminate grade tier at $\geq$1.7B). Safety tracks alignment and scale, not reasoning capability; at matched size, safety tuning, not reasoning, is decisive, and reasoning itself can \emph{erode} safety rather than add it.
#36 | Noise In, Harm Out: A Harm-Weighted Evaluation of Acoustic-Noise Propagation in ASR-LLM Clinical Scribe Pipelines}
Ehsan Latif (Cognizant)*
Abstract: Ambient AI scribes couple automatic speech recognition (ASR) with a large language model (LLM) to draft clinical notes from doctor--patient conversations. They are typically validated on clean audio yet deployed in noisy clinics, and the standard metric, word error rate (WER), cannot distinguish a lost filler word from a lost drug allergy. We ask how acoustic noise propagates through the \emph{full} ASR-then-LLM pipeline into clinically meaningful documentation error. Corrupting real primary-care consultations with four clinical noise environments at three signal-to-noise ratios, transcribing with nine ASR systems, drafting notes with three LLM scribes, and grading every documentation error on a four-tier clinical-harm taxonomy with a scalable LLM-as-judge (over 200{,}000 fact-level judgements), we find that severe noise roughly doubles harm-weighted documentation error (7.9 to 15.8), that about 78\% of the added errors are \emph{omissions} that leave no trace for a reviewing clinician to catch, and that the ranking of both ASR and LLM models changes with the acoustic environment. We argue that cascaded clinical AI must be evaluated with harm-weighted, noise-aware metrics rather than clean-audio WER.
#38 | Training-Free Prefill Safety Blocking via Cross-Layer Interpretability
Yiwen Zhang (Amazon Web Services)*; Madhavi Evana (Amazon Web Services)
Abstract: Information carried in LLM hidden dimensions can be used for safety guardrails. We find that an LLMâs prediction for a concept (e.g. animal, country) is largely determined early in the network, but only becomes readable in the final 15â30% of depth. We turn this finding into a training-free safety control. Because a harmful topic becomes readable during the prefill pass, a per-dimension sign monitor can flag it before any answer token is generated, catching paraphrased requests that a keyword blocklist misses. We ran analysis across four model families and four harmful topics, and compared against hard negatives (benign prompts from the adjacent domain, e.g. chemistry vs. weapons). Our method blocks 63â84% of harmful requests.
#40 | Tunable Tool-Call Rates in LLM Agents via Representation Steering
Yuqi Chen (UCSC)*; Vincent Siu (UCSC); Yang Liu (UCSC); dawn song (UCSC); chenguang wang (UCSC)
Abstract: Deciding whether to call a tool is a core competence of an LLM agent, and a costly one to get wrong: needless calls add latency and cost, while missing calls leave the model confidently wrong on questions it could only answer through tools. We show that whether an instruction-tuned model calls a tool is controlled by a single linear direction in its residual stream, extracted without training from the model's own tool-use preference signal and applied at inference with no prompt change. Tool steering moves the call rate monotonically from near $0\%$ to over $90\%$ while keeping calls well-formed; the induced calls land precisely on the questions the model cannot answer from its own knowledge, and the direction generalizes to unseen tools without favoring any specific tool choice. With live tool execution, sweeping steering strength traces a cost/accuracy Pareto frontier and nearly doubles open-domain QA accuracy ($0.29 \rightarrow 0.56$); the same recipe transfers across five models spanning dense, MoE, and multimodal architectures.
#41 | Human-centered Evaluation of Caste-based Representational Bias in Multi-modal LLMs
Smirity Kaushik (The George Washington University)*
Abstract: Multimodal large language models (MLLMs) are increasingly deployed in high-impact domains. However, little is known about how these models implicitly represent caste, a historical social hierarchy system in India, that continues to discriminate against the Scheduled Caste (marginalized) groups. Existing AI audits primarily focus on race and gender in Western regions, overlooking the persistent social biases in the Majority World. We investigate implicit caste representations in MLLMs through multimodal occupational personas. We develop a systematic evaluation framework comprising a curated occupation dataset with documented historical caste associations and a standardized prompting protocol. In a between-subjects survey study, 200 participants of Indian ethnicity will interact with ChatGPT or Gemini to generate occupational personas and reflect on the outputs. By analyzing both the generated personas and participantsâ responses, we examine how MLLMs represent caste and how these outputs influence perceptions of model fairness, trustworthiness, and accountability. This work contributes an evaluation framework, a public benchmark dataset, and culturally grounded insights for responsible AI.
#42 | Adaptive Hybrid Random Forest--XGBoost Framework for Accurate Obesity Level Prediction Using Lifestyle and Health Factors
karthi talapala (NEC)*; Tirumala Rao S (Narasaraopeta Engineering College); Avula Manikanta (Narasaraopeta Engineering College); Srinivas Reddy S (Vardhaman College of Engineering); Karuna G (GRIET); Sireesha P (Narasaraopeta Engineering College); Jyothi B (Narasaraopeta Engineering College)
Abstract: Obesity has become a major public health concern worldwide due to its association with chronic diseases and reduced quality of life. Accurate prediction of obesity levels can support early intervention and healthcare planning. This study proposes an Adaptive Hybrid Random Forest--XGBoost Framework for Accurate Obesity Level Prediction Using Lifestyle and Health Factors. The Obesity Levels Dataset containing 2,111 records and 17 attributes was utilized for experimentation. Data preprocessing techniques including label encoding, feature selection, and train-test splitting were applied to prepare the dataset. Random Forest and XGBoost classifiers were trained individually and later combined using a hybrid ensemble strategy based on probability averaging. Experimental results demonstrated that the proposed hybrid model achieved an accuracy of 96\%, outperforming individual classifiers. The model also achieved high precision, recall, and F1-score, indicating its effectiveness in multi-class obesity classification. The proposed framework can assist healthcare professionals in identifying obesity risks and supporting preventive healthcare strategies.
#43 | Personalized Federated Reinforcement Learning for Client-Centric and Application-Aware Adaptive Traffic Allocation in Wi-Fi Multi-Link Operation
Brian Trinh (Santa Clara University)*; Duy Ngo (Santa Clara University); Krishna Ramamoorthy (Santa Clara University)
Abstract: Multi-link wireless systems provide multiple transmission paths, but the optimal link varies with channel quality, contention, mobility, and application requirements. We present personalized federated reinforcement learning (P-FRL) for packet-level uplink allocation in a heterogeneous multi-client network. Each client observes the current channel conditions , then selects one of three links. P-FRL aggregates a shared feature-extraction trunk while retaining client-specific decision heads. In a five-client ns-3 environment, P-FRL reduces average latency by 94.9%. In contrast, full-model federated reinforcement learning increases latency by 159.7%, indicating that indiscriminate parameter sharing homogenizes client policies and amplifies wireless contention. Partial personalization therefore combines collaborative representation learning with the local adaptation required for real-time wireless scheduling.
#44 | Time Series Foundation Models versus Supervised Machine Learning for Network Telemetry Anomaly Detection
Nurcan Durak (Cisco)*
Abstract: Network assurance platforms continuously collect telemetry, model normal network behavior, detect anomalies, correlate evidence across network entities, and identify potential outages. Because nearly every network metric is observed over time, time series anomaly detection is fundamental to automated network troubleshooting. Traditional production systems commonly rely on statistical thresholding methods such as moving z-scores or interquartile range (IQR). Although simple and scalable, these approaches often fail to capture seasonality, recurring traffic patterns, and evolving network behavior, resulting in false positives and alert fatigue. This paper compares supervised machine learning and Time Series Foundation Models (TSFMs) for anomaly detection on network telemetry. For supervised learning, we engineered statistical features that capture temporal and distributional changes, including rolling z-scores, normalized interquartile differences, lag features, percentile-based slopes, and measures of dispersion and asymmetry. We evaluated Logistic Regression, Random Forest, XGBoost, LSTM, and CNN models. We also evaluated the Chronos-v2 TSFM in a zero-shot forecasting setting using 48 hours of history to predict one hour ahead.Tree-based ensemble models consistently achieved the best performance, with XGBoost and Random Forest reaching 0.98 recall and 0.49 precision. In comparison, Chronos-v2 achieved 0.29 recall and 0.26 precision. While TSFMs eliminate the need for labeled training data, we found they struggle with sustained network outages, periodic traffic patterns, and highly variable signals. Persistent anomalies tend to be absorbed into the forecast baseline, and increasing forecast uncertainty over longer horizons further reduces anomaly sensitivity. Our results show that supervised machine learning currently provides substantially more reliable anomaly detection for network telemetry.
#45 | Whisper or MMS? Multilingual ASR Routing and Error Analysis
Zhidong Lin (Santa Clara University); Tokunbo Ogunfunmi (Santa Clara University)*
Abstract: Multilingual ASR models are often selected using a single overall score, even though their performance can vary substantially across languages. This work examines whether Whisper large-v3 or MMS-1B-all should be used as a universal system, or whether model choice should depend on the language. Within the evaluated set, the two models show clear and complementary strengths: Whisper is stronger on the widely represented languages, whereas MMS is more reliable on the underrepresented languages. Exploiting this complementarity reduces languagemacro WER by 16.89% relative to the best fixed model without retraining. The results show that aggregate rankings can hide practically useful language-level differences and motivate more adaptive multilingual ASR deployment.
#46 | đ€ Information-Theoretic Analysis of Vision Transformers: Preliminary Results
Cameron Refaee (Santa Clara University); Tokunbo Ogunfunmi (Santa Clara University)*
Abstract: Information-theoretic diagnostics---mutual information (MI) between layers, divergence to a reference prior, channel capacity---increasingly guide Vision Transformer decisions such as hardware quantization. We test them against analytic ground truth using a controlled, invertible perturbation with known geometric effect. Two findings are cautionary: layerwise MI between continuous representations fails three ways, and KL-to-prior cannot distinguish \emph{scale} from \emph{structure}. One finding is positive: label information per token view---a bounded, self-checking target---shows that our perturbation \emph{diffuses} class information into patch tokens the classifier never reads. This shrinks the label-information margin at the read-out token. Accurate representation analysis therefore requires bounded targets, which make quantities estimable and estimator failures detectable. Prior ViT analyses using unbounded targets warrant reconsideration.
#47 | Verification of LLM-Narrated SHAP Explanations
Sydney Lin (Google); Veena Mendiratta (Northwestern University)*; Veena Mendiratta (Northwestern University)
Abstract: LLMs as the explanation interface for production ML predictions is used in many domains: the underlying modelâs feature attributions (SHAP values) go in as context, and a plain-English rationale is output. In our demo, given attributions (+0.45,+0.25,â0.35) for (Credit Score, Annual Income, Debt Ratio), an LLM writes: âPrimary driver for your approval is your excellent Credit Score. Annual Income also played a positive role, but your high Debt Ratio slightly decreased your score.â This is factually correct, but by magnitude Debt Ratio outranks Annual Incomeâthe model ordered features by valence (good news first), not importance. A reader perceives the wrong causal picture, and no factuality or hallucination check fires, since nothing was fabricated. Unlike open-ended generation, this setting carries its own ground truth: the attribution vector is available at generation time, so verification is nearly free. We make the following contributions: (1) We name and demonstrate valenceâmagnitude inversion, a faithfulness failure that factuality checks cannot catch; (2) We perform runtime validation (rank agreement with the true |SHAP| order) with a no-human fallback ladder and a tunable alert rate; and (3) Preliminary evidence across seven workhorse-tier LLMs on synthetic loan-decision SHAP vectors that inversion is systematic and that instruction-compliance splits robust/partial/none by model, so no static property predicts whether a given output is faithful. We also ran a case study on a network-intrusion detectorâs real per-alert TreeSHAP; the results will be reported in the presentation/poster.
#49 | A Progressive Image Codec for Human-Machine Vision
Tianma Shen (Santa clara university)*; Ying Liu (Santa clara university)
Abstract: Image coding for machines (ICM) compresses visual data for efficient processing by computer vision models, rather than for human viewing. However, in many real-world scenarios, the same visual content also requires human interpretation, introducing the need for image compression for humans and machines (ICHM). ICHM has the challenge of balancing compression efficiency with the different task requirements. In this paper, we propose a progressive conditional image coding framework that supports both human and machine vision tasks. Our method shares bitstreams among a base layer and two enhancement layers, where each enhancement layer compresses image content conditioned on multi-scale features extracted from previous layers. This design helps minimize redundant information across tasks. We also proposed a multi-masked conditional residual coding (MMCRC) scheme in the second enhancement layer, which integrates multiple sources of conditional information to boost reconstruction quality for human perception. Experiments on the COCO2017 dataset show that our framework achieves best rate-accuracy performance compared to existing ICHM methods across object detection, instance segmentation, and image reconstruction tasks. It also offers a competitive balance among bitrate, accuracy, and model complexity.
#50 | RankEvolve: Automating the Discovery of Retrieval Algorithms via LLM-Driven Evolution
Jinming Nian (Santa Clara University)*; Fangchen Li (Independent); Dae Hoon Park (melten.ai); Yi Fang (Santa Clara University)
Abstract: Retrieval algorithms like BM25 and query likelihood with Dirichlet smoothing remain strong and efficient first-stage retrievers, yet improvements have mostly relied on parameter tuning and human intuition. In light of the recent paradigm of using large language models (LLMs) to drive evolutionary program search, we investigate whether a LLM, when guided by retrieval evaluator and evolutionary search algorithm, can automatically discover improved lexical retrieval algorithms. We introduce RankEvolve, a program evolution setup based on AlphaEvolve, in which candidate retrieval algorithms are represented as executable code and iteratively mutated and selected based on the performance across 12 information retrieval datasets from BEIR and BRIGHT. RankEvolve uses BM25 and query likelihood with Dirichlet smoothing as independent seed programs for evolution. The resulting algorithms are novel, effective, and show promising transfer to the full BEIR and BRIGHT benchmarks as well as TREC DL 19 and 20. Our results suggest that evaluator-guided LLM program evolution is a practical path towards automatic discovery of novel retrieval algorithms.
#53 | PRISM: Transporter-Conditioned Molecular Representations for Interpretable Chemical Screening Under Sparse Labels
Reyansh Malla (The Harker School)*
Abstract: Environmental chemicals can resemble endogenous nutrients and potentially inter-act with membrane transporters that regulate their uptake. Computational discovery of such mimics is difficult because confirmed chemicalâtransporter relationships are scarce, verified negatives are rarer, and generic molecular fingerprints omit transporter-specific biological context. We introduce PRISM, a reference-guided ranking framework that evaluates chemicals relative to endogenous substrates and family-specific recognition requirements. PRISM maps 1,086,930 standardized chemicals against 72 endogenous references spanning 25 transporter classes. On a literature-informed benchmark of 57 family-specific assignments (19 documented transported substrates or transportable inhibitors and 38 presumed negatives), four families contained both labels and were evaluable by average precision. PRISM achieved macro-average precision (macro-AP) of 0.830, compared with 0.807 for Morgan fingerprints, 0.781 for MACCS fingerprints, 0.676 for descriptor cosine similarity, 0.624 for descriptor Euclidean similarity, and 0.443 for global nearest-endogenous similarity. The leading methods were not statistically distinguishable on this small benchmark. Thus, PRISM does not establish predictive superiority; it provides competitive retrieval with auditable, transporter-specific explanations in a sparse-label setting.
#54 | Confidence-Aware Shared Autonomy via Early Human Intent Prediction in Human-Robot Collaboration
Kajal Awasthi (Santa Clara University)*; Maria Kyrarini (Santa Clara University); George Okoroafor (Santa Clara University); Basile Ouedraogo ( Santa Clara University)
Abstract: Human-robot collaboration requires robots to infer user intent early enough to provide timely and effective assistance. However, intent prediction from partial human motion remains challenging due to incomplete trajectory information and uncertainty in early-stage observations. In this work, we present a confidence-aware shared autonomy framework for early robotic intent prediction from partial human reaching trajectories. A human initiates a reaching task while the robot continuously observes partial motion trajectories and predicts the intended target using a multiclass Logistic Regression model trained on end-effector position and orientation features. A confidence-based takeover mechanism triggers autonomous assistance once prediction confidence exceeds a predefined threshold, after which the robot completes the motion using Dynamic Movement Primitives learned from demonstrations. Experiments conducted on a real robotic dataset containing five target objects demonstrate that intent can be predicted with 70% accuracy using only 10% of the trajectory and 87% accuracy at 30% observation. Most autonomous takeovers occur early in the motion trajectory, enabling proactive and efficient human-robot collaboration. These results demonstrate the feasibility of real-time probabilistic intent prediction for confidence-aware shared autonomy.
#56 | đ€ PRISM: Predictive KV Cache Discovery for Agentic LLM Serving at Internet Scale
Adhvay Iyer (Del Norte High School); Naman Khurana (Skyline High School)*; Kathan Thai (Cosumnes Oaks High School); Aarush Rachakonda (Gunn High School); Claire Shao (Lynbrook High School); Shiyi Ling (University of California, Santa Cruz); Chen Qian (University of California, Santa Cruz)
Abstract: Agentic Large Language Models (LLMs) issue many related requests that share the same prompt prefix, so reusing the KV cache of that prefix saves serving time. But, as LLM serving becomes geo-distributed, a workflow revisiting its prefix often finds the KV cache already evicted, or kept in another region that is expensive to reach. We present PRISM, a predictive KV cache discovery plane for agentic LLM serving at Internet scale. PRISM tracks reusable KV caches across regions, forwards each request to the fastest GPU worker, and prepares the KV caches of predicted future requests in their destination regions in advance.
#58 | TRRAIL and Error: Evaluating Multi-Agent NeSy LLMs for Logic, Law, and Medical QA
Jonathan Wellington Morris (UC Santa Cruz)*; Khin Saw Yone (UC Santa Cruz); Leilani Gilpin (UC Santa Cruz)
Abstract: Logical reasoning remains a fundamental challenge for LLMs, particularly in high-stakes domains where users must inspect, verify, and trust intermediate reasoning steps. Neuro-Symbolic (NeSy) approaches such as SymbCoT, LogicLM, and ProSLM have made progress toward more interpretable reasoning, yet they are mostly evaluated on logic datasets as opposed to real world applications. In this work we investigate whether a multi-agent self-correction loop in a NeSy LLM QA system yields more reliable and transparent answers than systems without refinement. We evaluate across law, medicine, and logical reasoning datasets against CoT across two backbone LLMs (gpt-5.4-nano and gpt-5.3-chat-latest). To assess the explainability of our system, we introduce the TRRAIL score, short for Transparent Reasoning-relevant Reasoning-grounded Answer-supported Interpretable Logical score, an LLM-as-judge rubric evaluating whether reasoning is logically sound, complete, interpretable, relevant, answer-supported, and transparent rather than merely persuasive. We find that explicit self-correction improves not only accuracy, but also the ability to expose, diagnose, and revise reasoning failures in the intermediate representation, which results in better explanations, and greater trust in models deployed for critical applications like law and medicine.
#59 | One Model, Five Roles: A Sub-Gigabyte Offline Visual Memory Assistant, and Why Dense Retrieval Underperforms at Personal Scale
Leena Rajan Katkar (University of Southern California)*; Abinas Roy (Apple)
Abstract: Personal assistants that see what you see are built almost exclusively as cloud services because the component stack, a vision-language model, a speech recognizer, a sentence encoder, a reranker, and a speech synthesizer, does not obviously fit on a phone. We describe a fully offline visual memory assistant that does the following: points a camera, asks by voice, and receives a spoken answer grounded in a persistent store of things you have previously seen, with no network call at any point. Two decisions carry the system. First, a single 1B-parameter VLM is reused as five components: visual QA, text QA, summarization, reranking, and the embedding model itself, taken from mean pooled hidden states of its own language backbone, eliminating the dedicated sentence encoder entirely. Second, retrieval is lexical-first with dense search as a narrow fallback, a design we converged on empirically and which we think reflects a real property of personal-scale corpora. We report the resulting memory budget, give the ablation that would settle the retrieval question, and argue that sub-gigabyte agents are a distinct regime rather than a scaled-down server one.
#60 | Integrating AI-Assisted Image Enhancement with Physics-Based Synthesis of Low-Field MRI from High-Field MRI
Dang Bich Thuy Le (Santa Clara University)*; Muller Gomes (Promaxo Inc); Anh Truong (Promaxo Inc); Hoang Le (Promaxo Inc); Dnyanada Kadiyal (Promaxo Inc); Aleksandar Nacev (Promaxo Inc); Ram Narayanan (Promaxo Inc); Yuling Yan (Santa Clara University)
Abstract: Low-field (LF) MRI offers distinct benefits in affordability, portability, and accessibility, but its broader adoption is constrained by an inherently low SNR and reduced spatial resolution. This study introduces an AI-assisted framework to enhance LF MRI image quality and address these constraints. Specifically, we propose a two-stage framework for producing high-quality LF MRI images. In the first stage, realistic LF images are generated from high-field (HF) acquisitions via a physics-informed forward model that incorporates spiral k-space trajectories and accounts for nonlinear magnetic field gradients, B0 inhomogeneity, and the reconstruction characteristics of LF systems. In the second stage, a 3D U-Net augmented with multi-head attention in a Vision Transformer module is trained on paired synthetic LF and HF images, functioning as post-processing to enhance reconstructed LF images. On the synthetic test set, our framework achieved a peak signal-to-noise ratio of 19.08 ± 2.85 for the baseline U-Net and 21.00 ± 2.50 with the ViT block, indicating high reconstruction fidelity. The structural similarity index measure reached 0.65 ± 0.08 without ViT and 0.66 ± 0.08 with ViT, alongside low normalized RMSE values of 0.39 ± 0.1 and 0.31 ± 0.07, respectively, demonstrating improved image quality and reconstruction robustness. Applied as post-processing following standard reconstruction, the trained network consistently improved the contrast-to-noise ratio of output images, confirming the observed gains in contrast and clarity. The proposed framework addresses key barriers to wider adoption of LF MRI, namely the noise, artifacts, and resolution loss characteristic of LF acquisitions. By combining deep learning with physics-based simulation, the framework delivers notable qualitative and quantitative improvements in denoising, artifact removal, and overall image quality. These findings underscore its potential to substantially improve the practical utility of LF MRI.
#61 | Why Token-F1 misranks long-horizon memory systems
Haowei Liu (Santa Clara University)*; Yi Fang (Santa Clara University)
Abstract: Long-horizon dialogue benchmarks often treat token-level F1 as a direct measure of memory quality. Under matched retrieval and generation conditions, we find that answer form can move this score more than the memory architecture and can reverse the ranking given by semantic correctness. On LongMemEval, a prompt-only change adds 0.246 F1 while reducing judged accuracy by 7--8 points; on LoCoMo, one exact abstention string adds 0.20 aggregate F1. Reproduced architecture effects are much smaller. We argue that memory benchmarks should separate retrieval, answer form, and semantic correctness.
#62 | đ€ Let Them Sleep: Recursive Self-Pruning Memory for Long-Horizon LLM Agents
Kargi Chauhan (University of California, Santa Cruz)*; Sadiba Nusrat Nur (University of California, Santa Cruz); Pratibha Revankar (University of California, Santa Cruz); Kyle Montgomery (University of California, Santa Cruz); Chenguang Wang (University of California, Santa Cruz)
Abstract: Agents that run for a long time have a memory problem, and it is not the one most systems optimise for. Storing every trace and retrieving the nearest neighbours works early and quietly stops working later: the store fills up with failures the agent then repeats, and with facts that were true fifty turns ago. RSPM (Recursive Self-Pruning Memory) borrows the fix biology already found. At intervals the agent sleeps. It clusters its recent failures, distils each cluster into a short rule, and deletes the raw episodes it just learned from. Rules decay unless something reinforces them, and rules that recur across many contexts get promoted into general ones. We tested this on three settings with little in common: 19,008 SWE-bench issues, temporal consistency across 12 conversational benchmarks, and long-horizon QA on LoCoMo. Performance rose by 7.6 to 11.9 points in each. More interestingly, three separately tuned experiments landed on nearly the same equilibrium, roughly 93% fewer retry loops and roughly 30% less stored memory, which suggests that knowing what to forget is a general property of agent memory rather than a benchmark-specific trick.
#63 | When a Change Point Is Not a Point: Stress-Testing Localization Sets on Forecast Residuals
Diyun LU (UNIVERSITE DE LORRAINE)*; Marianne Clausel (UNIVERSITE DE LORRAINE); Alessandro Leite (Institut National des Sciences Appliquées de Rouen); Charles Findling (Pernod Ricard); Wei Gong (Pernod Ricard); Pierric Kersaudy (Pernod Ricard)
Abstract: Change-point procedures usually return a single boundary, although operational use depends on how precisely that boundary is localized. We study set-valued localization on forecast-residual windows and explore a candidate-wise block-randomization adaptation of likelihood-ratio test inversion for serially dependent data, comparing it against Bayesian highest-posterior-density sets, raw conformal MCP sets, and a circular block bootstrap over a shared candidate domain. Recent conformal methods guarantee coverage under within-segment exchangeability; forecast residuals violate this, so we treat the block-based procedures as dependence-aware diagnostics rather than exact confidence sets. Across nine controlled and semi-synthetic settings one trade-off recurs: the most compact sets are also the most severely miscalibrated under dependence and heteroscedasticity, whereas the procedures that restore coverage often do so by retaining much of the window, and on one real series no method approaches the nominal level. Localization uncertainty therefore deserves to be reported alongside the point estimate in forecasting and model-monitoring systems.
#66 | Asymmetric Transfer in LoRA-Adapted Dual Encoders for E-Commerce Retrieval
Dharmil Chandarana (Shopsense AI)*
Abstract: Adapting a pre-trained dual-encoder retrieval model (CLIP-style, separate image and text towers) to an operatorâs own catalog is routine in e-commerce and judged by one retrieval score on a held-out slice of that catalog. This hides an asymmetry between the two towers. LoRA-adapting both towers on a proprietary apparel catalog improves cross-modal retrieval by up to +26 Recall@1 points in-domain but degrades it on unseen public benchmarksâthe text tower over-specializes to the in-domain caption distributionâwhile imageâimage retrieval moves the opposite way, improving on every distribution tested. The pattern reproduces in adapter placement and across four catalog domains. Since a catalog serves both textâimage search and imageâimage âsimilar productsâ off one embedding, a single âfine-tune the retrieverâ decision silently trades one featureâs out-of-domain quality for anotherâs in-domain gain. We quantify this against CLIP and SigLIP2 and argue retrieval adaptation should be reported per-tower and per-direction.
#68 | Bringing Up a Bilingual BabyLM: Investigating Multilingual Language Acquisition Using Small-Scale Models
Linda Zeng (Stanford University)*; Steven Feng (Stanford University); Michael Frank (Stanford University)
Abstract: Multilingualism is incredibly common around the world, leading to many important theoretical and practical questions about how children learn multiple languages at once. For example, does multilingual acquisition lead to delays in learning? Are there better and worse ways to structure multilingual input? Many correlational studies address these questions, but it is surprisingly difficult to get definitive answers because children cannot be randomly assigned to be multilingual and data are typically not matched between languages. We use language model training as a method for simulating a variety of highly controlled exposure conditions, and create matched 100M-word mono- and bilingual datasets using synthetic data and machine translation. We train GPT-2 models on these data organized to reflect a range of exposure regimes, and evaluate their performance on perplexity, grammaticality, and semantic knowledge. Across model scales, measures, and languages, bilingual models perform similarly to monolingual models in one language, but show strong performance in the second language as well. These results suggest there are no strong differences between different bilingual exposure regimes, and bilingual input poses no in-principle challenges for agnostic statistical learners.
#69 | When Screenshot History Does Not Help: A Matched-Budget Audit of GUI-Agent Retention
Rishit Sai Shiramshetti (Santa Clara University)*
Abstract: Long-horizon GUI agents often retain screenshots from previous steps, but it is unclear whether this visual history improves decisions enough to justify its compute cost. We audit four screenshot-retention rulesâinstruction similarity, spatio-temporal re-weighting, fixed recency, and random retentionâat a matched 20% visual-token budget, together with full-history and zero-history controls. On 23 GUI-Odyssey episodes, the six arms formed a near- tie: a 0.9-point spread over 344 shared steps (p > 0.99). A diagnostic audit then found that prompt placement was a much larger source of variation than the retention rule: with the template and parser held fixed, moving the prompt from the user turn to the system role increased official-evaluator accuracy from 17.4% to 39.1% on 46 identical steps. In the corrected configuration, removing all historical screenshots changed accuracy by â1.3 points (378 paired steps, 95% CI [â12.4, +9.7], McNemar p = 0.72), while reducing peak memory by 2.6Ă and median latency by 3.0Ă. The result is not evidence that visual history never helps: the retention-rule comparison was collected before the harness correction, the study uses one model and one seed, and the contrast has only 0.50 power against a previously reported 6.1-point effect. Instead, it shows why evaluation of context-compression methods should include zero-information controls, matched budgets, and explicit audits of prompt and scoring sensitivity.
#71 | Beyond the Stitching Assumption: A Unified Framework for Multimodal Synthetic Data Evaluation via Semantic Quantization
Yefeng Yuan (Santa Clara University)*
Abstract: Multimodal synthetic records combine structured attributes with free-form text, yet are often audited with separate tabular and language metrics. This creates a stitch-ing assumption: realistic components are treated as evidence of coherent same-row pairings. We present a projection-based extension of SynEval that quantizes text semantics and tabular variables, then evaluates selected cross-modal contingency tables for fidelity, dependence, utility, diversity, and calibrated proximity risk. The Tilted Data Protocol supplies a negative control by shuffling text while preserving unimodal marginals. Across Amazon Reviews, Kiva Loans, and Fake Jobs, conventional checks validate deliberately misaligned data, whereas our diagnostics reveal dependence loss, over-conditioning, utility degradation, and near-collision risk. Trustworthy multimodal synthesis therefore requires dependence calibrated to real paired data, not merely realistic components or stronger association.
#72 | It Ain't Over Till It's Over: Modeling the Human Cognitive Layer After AI Inference
Zahra Sajedinia (Saint Marys College of California)*; Nahid Aalam (Saint Marys College of California); Harsh Singh (Saint Mary's College of California); Sayan Patra (Saint Mary's College of California); Simran Bains (Saint Mary's College of California)
Abstract: AI inference is not always the endpoint of an AI-assisted system. When a human makes the final decision, the AI output and its presentation must still be interpreted. To make this post-inference transformation explicit, we introduce the Human Cognitive Layer (HCL), a model-agnostic computational framework for describing how presented information becomes a human judgment or action. At Marr's computational level, HCL specifies this input-output problem; at the algorithmic level, it raises a central question: can cognitive models of social information integration and judgment updating transfer unchanged to human-AI interaction? We examine one directional implication in a randomized experiment in which 70 participants evaluated the same 15 AI-generated responses with or without color-coded bias-severity labels. Under a simple directional transfer of established cognitive models' information-integration principles, favorable information should increase reliance and confidence, whereas unfavorable information should reduce them. In our experiment, however, presenting bias information increased reliance overall, even though the labels could convey either favorable information (e.g., low bias) or unfavorable information (e.g., high bias). At the same time, confidence in participants' judgments did not change. This pattern is not straightforwardly explained by a direct, unchanged transfer of the existing cognitive models we tested to human-AI interaction. These results do not establish that fundamentally different cognitive mechanisms are involved when humans interact with AI, nor do they imply that entirely new cognitive models are required. Instead, they motivate testing whether existing cognitive models require recalibration, extension, or structural modification in human-AI settings. HCL makes this question explicit and provides a framework for evaluating and ultimately optimizing the complete AI-plus-human pipeline rather than the AI output in isolation.
#73 | The Retrofitting Bottleneck in Post-Hoc Depth Folding
Da-Yi Wu (UCLA)*
Abstract: Depth folding compresses a pretrained transformer by replacing k consecutive blocks with one block applied k times, matching depth pruning in stored parameter count. We ask whether recursive weight sharing can be imposed post-hoc on a pretrained model. We compare the two on GPT-2 small, holding the checkpoint, span, parameter count, and recovery budget fixed. Post-hoc, folding underperforms pruning, and the gap widens from +0.29 PPL at k=2 to +1.04 PPL at k=3; it persists under different initialization, objective, and span position, and keeps its direction across evaluation domains. Trained from scratch, the two architectures show no significant gap at either k; at k=2 with 20 tokens/parameter and 6 seeds, the gap's 90% CI half-width is 1.8% of pruned model perplexity. A full recovery epoch narrows the post-hoc gap to +0.10 PPL without closing it. This is consistent with a retrofitting effect rather than an intrinsic cost of recursive sharing.
#74 | Reliability of LLM-based Evaluation in Multilingual Deception Detection
Jeromy Chang (Santa Clara University)*; Parth Bhalerao (Santa Clara University); Oana Ignat (Santa Clara University)
Abstract: Large language models (LLMs) have became useful for generating text in recent years. With more AI generated text, there has also been an increase of using LLMs to judge whether text was produced by another LLM. This creates a fundamental question for multilingual deception detection: can LLMs reliably recognize AI-generated text across languages and generators? We study this problem using using 9,985 human-written hotel reviews spanning ten languages and 79,880 matched generation requests across eight instruction-tuned LLM backends. We compare a trained character-level detector with LLM judges under single-review classification and matched pairwise discrimination, while also measuring sensitivity to ordering and repeated evaluation. In a preliminary study across English, Korean, and Turkish, the detector achieved F1 scores of 0.944--0.966, compared with 0.661--0.671 for single-review judging and 0.290--0.408 for pairwise judging. Providing both reviews therefore reduced performance in every language. These results show that current LLMs remain weak multilingual deception detectors and that comparative context can worsen, rather than improve, their judgments.
#75 | MemRerank: Preference Memory for Personalized Product Reranking
Zhiyuan Peng (Santa Clara University)*; Xuyang Wu (Independent Researcher); Huaixiao Tou (Independent Researcher); Yi Fang (Santa Clara University); Yu Gong (Independent Researcher)
Abstract: LLM-based shopping agents increasingly rely on long purchase histories for personalization, yet naively appending raw history to the prompt is often ineffective due to noise, length, and relevance mismatch. We propose MemRerank, a preference-memory framework that distills a user's purchase history into a concise, query-independent representation of shopping preferences that can be precomputed once and reused across future searches. We build a benchmark that pairs Amazon-C4 product-search queries with the user's temporally preceding purchase history from Amazon-Review-2023, and evaluate LLM reranking over fixed top-100 candidate pools retrieved from the full product corpus. MemRerank trains a Qwen2.5-7B-Instruct memory extractor with GRPO, using feedback from the downstream setwise reranker together with a lightweight deterministic memory-quality regularizer. On Electronics and Beauty \& Personal Care queries across four first-stage retrievers, MemRerank attains the best macro-average MRR@10 and MRR@5, as well as the best result on a held-out retriever, while using substantially fewer memory words than raw history and MR.Rec. These results show that compact, learned preference memory can match or exceed far longer memories at a fraction of their length, providing a practical building block for personalized agentic e-commerce search.
#76 | Evolving Semantic ID Generation for Generative Retrieval via Proxy-Guided Program Evolution
Jinming Nian (Santa Clara University)*; Jyun-Yu Jiang (Coupang); Henry Tsai (Coupang); Yi Fang (Santa Clara University); Zhirong Liang (Coupang)
Abstract: Generative retrieval (GR) is an emerging paradigm in search and recommendation. The core idea is to train a generative model as a differentiable corpus index that retrieves relevant documents generatively for a given query. Because generating raw documents directly is challenging, a common workaround is to represent documents using semantic identifiers (SIDs) and train the generator to decode SIDs instead. Since autoregressive beam search makes early cuts, a good SID structure prevents premature candidate pruning and helps guiding the generator to relevant documents. We present a framework to continuously evolve SID generation to enhance downstream retrieval performance by iteratively refining the SID solution using an island-based evolutionary loop with a 31B open-source mutator LLM and a 1M-parameter proxy estimator. We show that this loop can iteratively modify SID solutions to improve the proxy estimator's Recall@100. We also demonstrate that this proxy estimator can be used to describe the quality of SID solutions, as it correlates strongly with the full 1B-parameter target generator. Finally, we present preliminary results on NQ320K.
#77 | How Radiologists Respond to AI After Deployment: Evidence from Five Years of Clinical Use
Alexander Zentefis (Santa Clara University)*
Abstract: Most evidence on clinical AI comes from model benchmarks, controlled reader studies, or short follow-up after rollout. These designs reveal little about how experts respond during years of routine use. We study an FDA-cleared system that flags pulmonary embolism (PE), a blood clot that blocks arteries in the lungs and causes an estimated 60,000â100,000 U.S. deaths each year. The system was introduced across eight care sites and followed for up to five years. The data include more than 117,000 scans; for 53,880 post-rollout scans, we link the model output to a signed report from one of 169 radiologists. Response differs sharply by model output. Across radiologists, disagreement averages 15.9% after a positive flag but only 2.8% after no flag. Agreement on no-flag cases is tightly concentrated near 100%, while responses to positive flags vary widely. Among emergency-department scans, disagreement with positive flags falls from about 30% near rollout to the low teens within two years; disagreement after no flag remains near 2â3%. In a 153-radiologist subsample, preview use is not monotone in agreement: positive-flag disagreement is 19.1%, 12.3%, 9.1%, and 19.2% across increasing use quartiles (p=0.029). The model and report disagree on 3.3% of scans. Among first scans, adjusted 30-day mortality is 8.7% when only the model flags PE and 10.2% when only the radiologist diagnoses itâpoint estimates at or above the 8.1% rate when both identify PE and roughly twice the 4.7% rate when both do not. These findings show what one-time evaluations miss: responses differ by model output, vary across radiologists, and change after rollout. Even 96.7% overall agreement leaves a small set of clinically important cases on which the model and radiologist give different answers. That distinction matters whenever a model flags rare cases and a person makes the final decision.
#78 | You Wonât Believe This Click: Content Rewriting for Agentic Choice
Tianyi Jin (UC Berkeley)*; Zirui Wang (UC Berkeley); David Chan (UC Berkeley)
Abstract: Language models increasingly mediate what information users see, creating incentives to optimize content for agents as well as humans. We study selection shifts induced by rewriting in agentic decision-making. Given a fixed set of competing content snippets, we rewrite only one target snippet while leaving the rest unchanged and measure how this affects the agent's choice. We operationalize this setup with Agentbait, an advisor-rewriter framework in which the advisor learns to propose rewriting strategies and the rewriter revises the snippet. While a fixed-prompt rewriter improves target snippet selection from 17.1% to 34.8%, our Agentbait raises its selection to 98.5%. The advisor effectively transfers to setups with different agents, languages, and snippets in other domains. Yet higher selection can result from unsupported rewrites. Adding a reward for support from the original snippet redirects the advisor toward more supported rewriting strategies, revealing a trade-off between factuality and target selection. Together, our results show that agent decisions can become a learnable pressure on content presentation, even when selection and source support diverge.
#79 | Route, Donât Scale: A 450M VisionâLanguage Model that Beats a 30B Model at On-Device Spatial Grounding
YAHYA SOWTI KHIABANI (Mercedes Benz)*
Abstract: Multimodal assistants are usually built as monoliths: every request traverses the full vision--language stack, and capability is bought with parameter count. Both choices fail in embodied, latency-bound settings. We test this where the constraint is unavoidable --- resolving a spoken parking command (``park next to the red car'') against surround-view cameras on automotive-grade compute, with no cloud fallback. Spatial language, we observe, is \emph{modality-heterogeneous}: only about a third of commands need vision; the rest are exact functions of geometry the vehicle already knows. We combine LoRA fine-tuning of a 450M VLM, a learned modality gate deciding per query whether images are needed at all, and a deterministic solver for the non-visual majority. This reaches 87.1\% on a held-out 8-way benchmark, beating a 30B mixture-of-experts VLM by 17.8 points at $\sim$1.5\% of its parameters while cutting visual tokens $\sim$75\%. A negative result also follows: chain-of-thought supervision, helpful \emph{before} fine-tuning, hurts after.
#80 | Reliable and Efficient Evaluation of AI Tutors
Parth Bhalerao (Santa Clara University)*; Jeromy Chang (Santa Clara University); Oana Ignat (Santa Clara University)
Abstract: As AI tutors become more common, their responses must be evaluated on more than factual correctness. A useful evaluation should determine whether the tutor recognizes a student's mistake, locates the error, provides meaningful guidance, and suggests an actionable next step. Existing work often studies these abilities separately, which can produce strong results for one dimension without ensuring balanced performance across the others. We define reliable evaluation as producing accurate and balanced judgments across all four pedagogical dimensions while consistently returning the required labels. Our work aims to answer the following main research question. Which modeling and training choices enable reliable and efficient evaluation of AI tutor responses across these dimensions? We study single task and multitask learning, LoRA fine-tuning, synthetic augmentation, and CoT+Reasoning while also measuring carbon emissions. Our findings show that multitask learning produces more balanced judgments, while the benefit of augmentation depends on the underlying model. CoT+Reasoning is more useful for generating training data than for direct classification, and self verification increases carbon cost without consistent gains. These findings highlight the importance of evaluating accuracy, balance, output stability, and environmental cost together when designing AI evaluation systems.
#83 | Tinted Frames: Question Framing Blinds Vision-Language Models
Wan-Cyuan Fan (University of British Columbia)*; Jiayun Luo (University of British Columbia); Declan Kutscher (University of California, Berkeley); Leonid Sigal (University of British Columbia); Ritwik Gupta (University of California, Berkeley)
Abstract: Vision-Language Models (VLMs) have been shown to be blind, often underutilizing visual inputs even on tasks that require visual reasoning. In this work, we demonstrate that VLMs are selectively blind: they modulate attention to visual inputs based on linguistic framing even when alternative framings demand same visual reasoning. Using visual attention as a probe, we quantify how framing alters both the amount and distribution of attention over the image. Constrained framings, such as multiple choice and yes/no, induce substantially lower attention to image context than open-ended and reduce focus on task-relevant regions. We further show that this attention misallocation is a key mechanism contributing to degraded accuracy and cross-framing inconsistency. Evaluating common prompting strategies, we find that simple image-focused instructions provide limited benefit, while structured reasoning and grounding prompts partially restore target-region attention and reduce but do not eliminate cross-framing inconsistency. Our findings suggest that question framing and prompting should be considered part of VLM evaluation to better reveal models' visual capability.
#84 | A Parameterized Dynamic Movement Primitive Library for Reusable Robot Cooking Motions
Tran Khanh Phuong Cao (Santa Clara University)*; Maria Kyrarini (Santa Clara University)
Abstract: We present a parameterized Dynamic Movement Primitive (DMP) library for three robot cooking motions: spreading, stirring, and pouring. Kinesthetic demonstrations were collected with a Franka Emika Panda robot, segmented into task-relevant phases, and encoded as DMPs. For spreading and stirring, the model learns end-effector position trajectories while assuming a fixed and stable spatula orientation. For pouring, both position and orientation are learned because the arm must tilt the container to transfer its contents. Preliminary results show accurate trajectory reproduction and initial parameterized adaptation. The current study focuses on kinematic trajectory reproduction and open-loop temporal scaling; velocity and acceleration evaluation, physical contact, and material-flow behavior remain for future work.
#85 | The iPhone Architecture of Human Cognition: Leveraging Supertasker Mechanics for Edge-AI and Asymmetric Token In-Filling
Long Le (Santa Clara University)*
Abstract: This work addresses a critical bottleneck in edge-AI: routing concurrent, high-bandwidth asynchronous data streams under severe resource constraints. To explore solutions, this paper introduces a novel, bio-inspired architectural framework modeled on an extreme cognitive exception. Operating as both the principal investigator and the subject, the presenter frames this study strictly as an empirical proof-of-concept prototype (N=1) to bridge the gap between cognitive neuroscience and hardware-aware machine learning. The presentation features a dual-methodology demonstration strategy designed for an in-person symposium. First, a recorded YouTube case study will analyze the independent generation of advanced 7:4 jazz drumming polyrhythms during continuous speech. Second, the presenter will conduct a live, interactive whiteboard demonstration executing simultaneous, bilateral text in-filling while holding an unscripted conversation with the audience. Contribution A: Neuromorphic Isolated Subsystems (The "FaceID" Approach). Currently, the iPhone offloads the task to an isolated Neural Engine, completely bypassing the main CPU/GPU clusters. We propose an ML routing framework as an automated, deterministic motor loop, the cognitive architecture isolates it into a "Neuromorphic Subsystem." This frees up the main computational graph (the Unified Memory). Contribution B: Asymmetric Dual-Head Token In-Filling (The "Predictive Text" Approach). Standard predictive text models predict the next token sequentially (Left â Right). We introduce a Dual-Head In-Filling Architecture inspired by the presenter's simultaneous dual-hand whiteboard writing. The model deploys twin, independent generation heads operating out of a shared latent workspace. Head A generates tokens forward (Left â Right) while Head B generates tokens backward (Right â Left). The model reduces sequence generation latency by half, mapping perfectly to multi-threaded edge processors.
#86 | Budget-Aware Test-Time Scaling via Discriminative Verification
Kyle Montgomery (UC Santa Cruz)*; Sijun Tan (UC Berkeley); Yuqi Chen ( UC Santa Cruz); Siyuan Zhuang (UC Berkeley); Tianjun Zhang (UC Berkeley); Raluca Ada Popa (UC Berkeley); Chenguang Wang (UC Santa Cruz)
Abstract: Test-time scaling has become a common recipe for boosting the performance of large language models on complex reasoning tasks. State-of-the-art approaches often employ generative verifiers to select the best solution from a pool of candidates, but their computational cost limits their practicality. We therefore pivot to a more budget-aware paradigm: discriminative verification. Although discriminative verifiers may underperform in isolation, combining them with self-consistency yields an effective and efficient test-time scaling mechanism. Under fixed compute budgets, this hybrid approach achieves accuracy up to 15.3 points higher than state-of-the-art generative verification on AIME2025. Our findings establish budget-aware scaling with discriminative verifiers as a practical alternative to costly generative techniques.
#87 | Memory-based Self-Explainable Early Exit Networks
Evan Kaiden (University of California, Santa Cruz)*; Leilani Gilpin (University of California, Santa Cruz); Biagio La Rosa (University of California, Santa Cruz)
Abstract: Early Exit Neural Networks accelerate inference of deep neural networks by allowing samples to exit early through internal classifiers. Despite recent progress in performance for these methods, it remains unexplored whether these early exits can provide interpretability alongside the prediction. In this work, we propose Memory-based Self-Explainable Early Exit Networks (M-SEEEN), a design to make the internal classifiers of exit networks more interpretable. M-SEEEN uses a sparse content-based attention mechanism to select relevant examples from a memory set and uses them to aid and explain the decision process. We show that, compared to alternative self-exlainable designs, M-SEEEN best preserves the performance of black-box baselines while providing means to inspect its inner mechanisms.
#88 | Signoff: Bringing Human Accountability Back to Agent-Powered Research
Jerry Lin (Boston University)*
Abstract: As coding agents author a growing share of research and production code, review degrades toward rubber-stamping: the human approves a diff they cannot explain. This \emph{cognitive surrender} is most dangerous where code fails silently rather than loudly---unit slips, data leakage, numerical drift. We present \textbf{Signoff}, a harness-agnostic protocol that inverts the review relationship: before merge the agent interviews the human across four fixed axes and, only on a passing interview, emits a \emph{Git Signoff Attestation} (GSA)---a signed empty commit mirrored into \texttt{refs/notes/signoff} binding the reviewed tree, a transcript digest, the interviewing model and profile, and the trade-offs and risks the human explicitly accepted. We report a self-hosted deployment of 14 attestations across two independent agent harnesses, and state plainly what the artifact does and does not prove.
#89 | Designing Neural Value Models under Global Resource Constraints
Jinwen Xu (Pinterest)*; Yiling He (Pinterest); Sara Asgari (Pinterest); Lin Zhu (Pinterest); Sen Yang (Pinterest); Tingting Zhu (Pinterest); Sudarshan Lamkhede (Pinterest)
Abstract: Modern resource allocation systems increasingly rely on neural value models to estimate the marginal value of candidate interventions under global resource constraints, including notification delivery, promotional campaigns, personalized incentives, and healthcare interventions. Existing research has substantially advanced value modeling, decision-focused learning, and constrained optimization, yet comparatively little attention has been paid to how neural value models should be designed for globally constrained optimization. We argue that globally constrained optimization induces a common decision scale, under which downstream allocation depends not only on the accuracy of learned marginal values but also on the reliability of the comparisons they enable. This observation reveals design requirements for neural value models beyond predictive accuracy and motivates Optimization-Consistent Value Learning (OCVL), a unified framework combining optimization-consistent representations, decision-scale calibration, and decision-aware learning. Experiments on a large-scale production notification budgeting system demonstrate a 0.51% improvement in Weekly Active Users (WAU) in an online A/B test while maintaining comparable predictive performance.
#90 | OSGuard: A Benchmark for Safety in Computer-Use Agents
Mina Mohammadmirzaei (University of California, Santa Cruz)*; Jeffrey Flanigan (University of California, Santa Cruz)
Abstract: Computer-use agents are usually evaluated by task completion, yet they may succeed while taking unsafe actions such as altering unrelated content, exceeding the intended scope, or unnecessarily accessing sensitive information. We introduce OSGuard, a diagnostic benchmark at action and execution levels. Its action benchmark contains 324 contextualized candidate actions human-labeled as allowed when appropriate for completing the task, unrelated when irrelevant to task completion but not unsafe, or unsafe when they violate task constraints. Its execution benchmark contains 45 risk-augmented tasks derived from OSWorld that retain the original benign instructions while introducing latent environmental hazards; state-based invariants distinguish safe success from unsafe completion. On the action benchmark, the strongest guardrail attains 80% accuracy and 0.80 macro-F1 overall, while the best macro-F1 for actions collected from the risk-augmented tasks is only 0.48. On the risk-augmented tasks, an unguarded agent completes 38% of these tasks unsafely, while the strongest guardrail reduces this rate only to 33% without improving the overall task-completion rate. The results expose a gap between local risk recognition and reliable constraint preservation.
#91 | Emotion Is a Distribution, Not a Single Answer
Ethan Lin (Santa Clara University)*; Hsin-Tai Wu (DOCOMO Innovations); Mei Takeda (NTT DOCOMO, Inc.); Yi Fang (Santa Clara University)
Abstract: People increasingly turn to LLMs as friends and confidants, but LLMs are not people and do not respond like them. A growing number of benchmarks evaluate their emotional intelligence, nearly all against a single gold answer labeled by a handful of experts or an LLM judge. But there is no single right response. The same situation can call for comfort, a joke, or a hard truth, and different people want different ones. Thus, emotion is better modeled as a distribution. Even models show this in their own way. When sampled, their answers rarely land on the one answer a benchmark labels. Worse, interpretability work shows post-training reshapes a model's internal emotional profile and steers its preferences, so an LLM judge is a poor proxy for human ones. We take a different approach. Instead of relying on experts or judges, we observe implicit preferences that people already express online, community votes on millions of replies to real emotional disclosures. Votes also reward being early, so we keep only pairs where a later reply outscored an earlier one and thus the win must come from the reply itself. Each pair looks binary, but it is decided by a population of voters, not an answer key. Across millions of comparisons the pairs trace out the community's preference distribution. To our knowledge, no existing evaluation treats emotion as a distribution over many people's implicit preferences rather than a single labeled answer. Preliminary results show that LLMs judge these preferences barely better than random, while a far smaller reward model trained on our pairs predicts them substantially better, a first step toward models that respond the way people actually want.
#92 | Signal-Driven Observation: Decoupling Observation from Action in Long-Horizon Web Agents
Shubham Gaur (UC Santa Cruz)*; Ian Lane (UC Santa Cruz)
Abstract: Long-horizon web agents re-read the full page stateâaccessibility trees routinely spanning 20,000 to 80,000 tokensâbefore every single action, whether or not anything has changed. We name this coupling observation over-ingestion and argue it is an architectural defect rather than a context-length shortfall: it appears in frontier models with 200K-token windows as readily as in small ones, and acts as a triggering precondition for context rot, loop-trapping, and goal drift. We sketch Signal-Driven Observation (SDO), in which a dedicated sub-call returns only task-relevant elements and their selectors, re-invoked only when a zero-LLM-cost detector registers a genuine state change. A traced e-commerce checkout shows how the standard architecture cascades into failure and how decoupling averts it. We argue observation compression deserves treatment as a first-class architectural decision, and lay out the open problems that follow.
#93 | Autonomous Inspection of Wind Turbine Interiors with Multimodal Digital Twins and Vision-Language Agents
Jack Landers (Santa Clara University)*
Abstract: Abstract not authorized to be published
#94 | Impact Prediction Is Not Technology Foresight
Ethan Lin (Santa Clara University)*; Hsin-Tai Wu (DOCOMO Innovations); Yi Fang (Santa Clara University)
Abstract: Technology foresight, judging whether an idea will matter before it does, is the decision behind every research agenda, funding call and corporate strategy. Humans are not reliably good at it, including experts at the top of their field. It is now being asked of machines too, since language models propose research faster than anyone can read it. The obstacle is measurement. Impact is recorded, because citations count how much later work built on an idea. Novelty is not, because novelty is a property of an idea at the moment it appears while citations only accumulate afterwards. The two run against each other in time, because an idea that has been built upon is no longer new, so scoring by citation counts recommends the past. We propose learning foresight from the structure of the citation graph rather than its counts. Two relations can be derived from the graph alone, one from the direction of a citation and one from what two papers share, and together they record both how new an idea was and how much followed it. 39.5M abstracts yield 416.7M dated comparisons, and a 22M-parameter model trained on them traces a field's lineage from text alone, placing each idea below the work that grew out of it even on papers after its knowledge cutoff.
#95 | Generating the Modal Worker: A Cross-Model Audit of Race and Gender in LLM-Generated Personas Across 41 Occupations
Ilona van der Linden (Santa Clara University); Sahana Kumar (Santa Clara University); Arnav Dixit (Santa Clara University); Aadi Sudan (Santa Clara University); Smruthi Danda (Santa Clara University); Julianna Dietrich (Santa Clara University); David C. Anastasiu (Santa Clara University); Kai Lukoff (Santa Clara University)*
Abstract: Large language models (LLMs) now routinely generate portrayals of people in professional roles, feeding user personas, fictional characters, and synthetic datasets. Across 1.5 million personas generated by GPT-4, Gemini 2.5, DeepSeek V3.1, and Mistral Medium for 41 U.S. occupations, we benchmark race and gender against Bureau of Labor Statistics (BLS) data. Models collapse each occupation onto a single "modal worker": for three of four models, the median occupation receives one gender in 99-100% of personas. A shift/exaggeration decomposition shows White (-31pp) and Black (-9pp) workers underrepresented and Hispanic (+17pp) and Asian (+12pp) workers overrepresented, with existing occupational segregation amplified. These patterns recur across models from three countries, pointing to shared structural sources of bias and motivating multi-model audits as standard evaluation practice.
#96 | The Gaussian Is Enough: Flow-Matching Priors Do Not Help When Fine-Tuning LBMs
Chen Xu (Toyota Research Institute)*; Rishi Shah (Woven by Toyota); Hadas Kress-Gazit (Cornell University); Haruki Nishimura (Toyota Research Institute); Masha Itkina (Toyota Research Institute)
Abstract: Modern robot imitation learning increasingly relies on generative policies based on diffusion or flow-matching models, which generate actions by transforming samples from a prior distribution. Replacing the standard Gaussian with a closer-to-target, non-Gaussian prior has been shown to substantially improve performance when training from scratch. A natural next step is to ask whether these gains transfer to fine-tuning pretrained Large Behavior Models (LBMs) such as LBM 1.0, Ï0.5, and GR00T N1.5, where one might expect even larger gains. Surprisingly, they do not, except possibly at very low fine-tuning data fractions. Across over 100K simulation rollouts spanning all three aforementioned LBMs on 40+ tasks in two simulation platforms, and 1250 rollouts on bimanual robot hardware, non-Gaussian priors that are demonstrably closer to the target yield statistically indistinguishable or worse fine-tuning performance than a standard Gaussian prior. Diagnostic analyses suggest why: fine-tuned policies result in similar action predictions across priors, despite their fine-tuned encoder embeddings diverging substantially from the pretrained embeddings and each other. A learning-rate ablation further confirms that encoder training is the dominant factor, outweighing the effect of prior choice.
#97 | Towards a Physiological Scaling Law: Model Quality vs. Cohort Size for Stochastic Sequence Data
Gayathri Sunil (University of Massachusetts Amherst)*; Bhuvana Ravi Kumar (University of California Davis); Bharath Ramsundar (Deep Forest Sciences, Inc.); Sandya Subramanian (University of California Berkeley)
Abstract: Scaling laws help determine the optimal data size for training large models but are established in domains where the target is deterministic. Physiological signals are different: heartbeat sequences are stochastic, so part of the error is irreducible even with large amounts of data. Metrics such as MAE do not account for non-deterministic behavior, and therefore assessing scaling requires evaluating distributional calibration (measuring how well predicted probability densities capture true conditional characteristics). We formulate a scaling law metric(n) = E + A*n^-α and evaluate it with five metrics: accuracy (MAE, RMSE), distributional calibration (KS distance, goodness-of-fit), and training objective (negative log loss) using a neural temporal point process trained on a cohort of four-ECG datasets. The law fits all five metrics. While point accuracy is near saturation at n=183, KS distance and goodness-of-fit improve by 6% and 12% respectively when extrapolated to 10,000 subjects, showing that scaling decisions in stochastic domains must be guided by distributional calibration rather than point accuracy.
#100 | SPADE-KD: Adaptive Peer-to-Peer Collaboration for Heterogeneous Learning Ecosystems
Harikrishna Kuttivelil (UC Santa Cruz)*; Katia Obraczka (UC Santa Cruz)
Abstract: Distributed learning is a structural paradigm for deep learning that enables collaborative training between clients distributed over a network. However, heterogeneity within learning ecosystems presents challenges to distributed learning. Data heterogeneity challenges consensus-based training towards globally optimal solutions, as they may diverge significantly from locally optimal solutions. Heterogeneity in model training or quality necessitates discernment in collaboration choices. Model architecture heterogeneity makes impossible the parameter averaging core to many federated learning approaches. While distillation-based collaboration overcomes this limitation, existing heterogeneity-aware approaches are largely dependent on centralized orchestration, limiting scalability and robustness. To enable collaborative learning in heterogeneous learning ecosystems, we present SPADE-KD, a fully decentralized model distillation framework that optimizes collaboration towards personalized rather than global performance. SPADE-KD leverages two mechanisms adaptive distillation and utility-driven peer selection built on a shared utility signal. Adaptive multi-distillation calculates this utility signal as the performance advantage of incoming models compared to the native model measured by their task losses, and uses it to modulate the strength of each incoming model transfer such that a client will distill more aggressively from incoming models that can outperform it. Utility-driven peer selection reuses this loss-informed utility signal to rank prospective partners, focusing on exchanges that measurably improve the local model while suppressing low-utility model transfers. Using federated classification tasks, we demonstrate that SPADE-KD and its individual components outperform naive decentralized distillation with lower loss and faster convergence.
#101 | Confidence Is Not a Design Criterion: Four Ways a Structure-Predictor-Driven Design Loop Fools Itself
Longfu Xu (UC Berkeley)*
Abstract: Structure predictors are now the acceptance test inside generative design loops. But what they expose is confidence about a fold, not evidence about a function, and usually there is no ground truth to grade it against â the only evidence the predictor is right is the predictor. We break that circularity on a ring-ATPase task with two quantities fixed by an experimental structure and optimised by no predictor. Across 27 constructs and three predictors, interface closure passed 45/45 at every cut site, conditioning alone flipped one sequence from 5/5 to 0/5, and the substrate channel (20.8, 14.2, 9.3 Ă
vs 20.6 Ă
native) separated live from dead. Two transferable8 rules follow.
#103 | @skills: ATTENTION IS ALL YOU HAVE
Li Yin (SylphAI); Haebin Seong (SylphAI); Haoran Zhang (SylphAI); Aria Shi (Santa clara university)*; Aknur Abdikarim (SylphAI); Aruzhan Sapayeva (SylphAI ); Shanglin Yang (Google)
Abstract: Agent skills package procedural knowledge as portable \texttt{SKILL.md} directories. Our July 2026 crawl identifies 56,804 public skill directories across 1,133 GitHub repositories, indexed by community registries such as skills.sh, while teams maintain many more privately. Yet delivery remains fragmented across agent-specific directories and commands: a skill may be installed for implicit activation, invoked manually after local setup, or fetched through a client-specific one-shot command. These options do not provide one portable lifecycle across three distinct decisions: whether to use a skill now, whether to keep a project-owned copy, and whether to let it activate without being named. We introduce \texttt{@skills}, an open delivery protocol that separates these decisions behind one portable, path-shaped address . \texttt{@skills:\textless path\textgreater} resolves a skill, subtree, or collection at the point of use and injects it adjacent to the task, without persistent prompt residency; \texttt{:save} vendors a reviewed copy into the project's git-tracked tree; and \texttt{:install} records a project-scoped auto-trigger policy, for which only the selected skills' frontmatter remains resident. Resolution is local-first, directories act as menus rather than indivisible bundles, and the underlying \texttt{SKILL.md} format is unchanged. The result is a shared lifecycle for public expertise, provider integrations, and private team workflows across agents. Install less, use more.
#104 | Ablating a Multi-Agent Pipeline for Spatial Transcriptomics Deconvolution: Critics Catch Errors but Miss the One That Matters Most
Tianbao Yang (Santa Clara University)*; Hamed Akbari (Santa Clara University)
Abstract: Multi-agent scientific pipelines commonly evolve by adding specialized agents, critics, retrieval modules, and increasingly detailed instructions. Despite their widespread adoption, the contribution of individual architectural components remains poorly understood. We present a systematic ablation study of a four-role multi-agent pipeline for spatial transcriptomics deconvolution using a benchmark with controlled data corruptions. Increasing tool granularity reduces opaque code execution from approximately 20% to 4.5% (n=45, p=0.009). Replacing detailed procedural instructions with high-level objectives improves report quality by 3.8 points on a 45-point rubric. In contrast, critic-based iterative revision provides no measurable improvement, and enabling tool access for the critic decreases performance by 4.7 rubric points. Across all ablations, architectural components that expand agent capabilities consistently improve performance, whereas additional scaffolding and procedural constraints provide limited or negative benefit. These findings identify which design choices materially affect multi-agent scientific pipelines and offer practical guidance for building simpler, more effective scientific AI systems.
#105 | Auditable Objectives, Brittle Behavior: Fitness-Function Tuning in Autonomous Driving
Leilani Gilpin (UCSC)*; Sumedha Kumar (Monta Vista High School)
Abstract: Autonomous driving systems balance progress, speed, and safety, but it is hard to see how these priorities affect their decisions. In simulation, we expose meaningful fitness weights and use a genetic algorithm (GA) to evolve a neural controller. This makes the reward easier to inspect and change, although it does not explain the network itself. Across three trials, a distance-dominant objective reduced the logged search/evaluation count by 35.7\% and runtime by 70.4\% relative to an all-zero control. Yet increasing the speed weight degraded both metrics, while positive sensor weighting frequently produced stalls or loops. Understandable fitness weights can therefore aid examination without making a neural policy explainable, reliable, or safe.
#106 | đ€ Drop the Act: Probe-Filtered RL for Faithful Chain-of-Thought Reasoning
Swapnil Parekh (NYU)*
Abstract: Reasoning models post-hoc rationalize answers they have already committed to internally, producing chains of reasoning theater: deliberative-looking steps that contribute nothing to correctness. This wastes inference tokens, pollutes interpretability, and obscures what the model actually computed. We introduce ProFIL (Probe-Filtered Reinforcement Learning) to reduce theater, increase chain-of-thought faithfulness, and shrink chain length in a single, drop-in extension to Group Relative Policy Optimization (GRPO). A multi-head attention probe is trained once on the frozen base model to detect post-commitment steps from internal activations alone; during GRPO, rollouts whose probe score exceeds a threshold have their advantage zeroed. Our central finding is that a probe trained on a frozen base, with verifier-derived labels and no human annotation, provides a stable signal that suppresses theater while resisting the RL-obfuscation failure mode predicted by prior work. Across four reasoning domains (GSM8K, LiveCodeBench, ToolUse, MMLU-Redux) and two model architectures (Llama-8B, Qwen-7B), ProFIL reduces post-commitment theater by 11â100%, raises faithful-fraction (e.g., +24pp on LiveCodeBench under an independent LLM judge), and shortens chains by 4â19%, all while preserving or improving task accuracy. ProFIL also beats a matched length-penalty GRPO baseline, isolating the gain as semantic commitment-detection rather than chain compression.
#107 | đ€ Blind Deference: A Diagnostic Benchmark and Vulnerability Index for Authority-Driven Sycophancy in LLMs
Sujit Noronha (Independent Researcher)*; Steven Au (Independent)
Abstract: Large language models frequently abandon correct answers when users invoke authority credentials and express disagreement, a failure mode we term authority-driven sycophancy. In production systems such as retrieval-augmented or memory-augmented applications, authority signals embedded in the user context can override accurate outputs across multi-turn interactions. We introduce the Authority Vulnerability Index (AVI), a composite metric for systematically measuring this vulnerability during model selection, fine-tuning, and pre-deployment auditing. Our evaluation spans 10 domains and 200 questions in a factorial design, assessing eight models ranging from 7B to 70B parameters. AVI integrates attack surface, overall flip rate, and worst-case flip rate into a 0--100 score. Across more than 21{,}000 conversations, we find that: (1) AVI scores range from 1.9 to 57.3, with reasoning-trained models exhibiting 77\% lower vulnerability than standard RLHF models; (2) all models exhibit blind deference, with statistically indistinguishable flip rates regardless of domain-authority alignment (BDI = 93--100\%, $p = 1.000$); (3) training paradigm, rather than model scale, predicts authority resistance, as Gemma-3-27B matches Gemma-3-12B in vulnerability while Qwen3-30B exceeds all 7--12B models; and (4) linear probes show that authority and sycophancy are encoded in nearly orthogonal activation subspaces ($|\cos \theta| < 0.07$). We release the benchmark, scoring pipeline, and AVI toolkit to support standardized authority-vulnerability assessment.
#111 | Grounding LLM Decisions with Lexicographically Ordered Constraints for Autonomous Driving
Madeleine Finn (Unaffiliated); Aditya Singla (Unaffiliated); Stephen Sheng (Unaffiliated); Esteban Ramos (University of California, Santa Cruz)*
Abstract: In autonomous driving LLM based safety controllers demonstrate high level reasoning but lack physical safety grounding, while hard constraint systems guarantee safety but are not able to perform complex reasoning and adapt to the situation. We propose an adaptable hierarchical constraint system, where an LLMâs semantic description of the scene activates a scenario specific constraint set; which is then solved by a propagator network with a Truth Maintenance System that hypothesizes the consequences for every viable action at a given step. Actions are then scored lexicographically, managing sensor uncertainty without defaulting to conservative fallback. Our method achieves zero crashes and full waypoint completion in an unoccluded sensor environment, while demonstrating graceful performance degradation in scenarios with increasing information masking.
#112 | Adaptable and Transparent Assistance in Appendicitis Diagnosis
Aditya Singla (Unaffiliated); Stephen Sheng (Unaffiliated); Madeleine Finn (Unaffiliated); Esteban Ramos (University of California, Santa Cruz)*
Abstract: Machine learning approaches have shown strong performance for appendicitis diagnosis but are often limited by a lack of adaptability in clinical settings, uncertainty quantification, and interpretability. Adaptability to new information from diverse sources is an operational priority to support throughout the diagnostic process, while transparency about predictions are essential to provide clinicians with contextual information required to incorporate AI outputs into decision making. We propose a framework that integrates five probabilistic machine learning models leveraging propagator networks to continuously update predictions as information becomes available, quantify uncertainty though intervals, and identify feature set contributions to predictions. Using the Regensburg Pediatric Appendicitis Dataset, we provide a transparent and adaptive foundation for clinically interpretable AI assisted diagnosis.
#114 | How Pruning Reshapes Knowledge: Insights from Compositional Explanations
Rishika Srinivas (UC Santa Cruz)*; Arnav Kartikeya (UC Santa Cruz); Leilani Gilpin (UC Santa Cruz); Biagio La Rosa (UC Santa Cruz)
Abstract: Pruning aims to reduce the size of deep neural networks while preserving their performance. Although recent techniques achieve strong compression rates, it remains unclear how pruning affects the knowledge encoded by these models. In this paper, we contribute to this direction by studying how pruning reshapes conceptual knowledge using compositional explanations, a neuron-level method based on logical formulas that can be applied across multiple activation ranges. We analyze unstructured pruning, with and without retraining, across BiLSTM, BERT, and LLaMA on a natural language inference task. We also introduce a set of metrics to quantify knowledge changes during pruning. Using this framework, we find that (i) activation ranges remain stable under pruning, while neuron-level representations change, and (ii) retraining-based methods preserve more concepts but reorganize them across neurons, whereas one-shot pruning preserves structure but loses more knowledge. We also observe a strong link between knowledge stability and model performance, and show that redundancy in transformer models supports robustness under pruning. The combination of these findings provides a characterization of how pruning reshapes internal knowledge, which can be used by future research directions to improve pruning methods.
#115 | EcoChef: Intelligent Food Inventory for Everyday Sustainability
Yu Ji (Santa Clara University)*; I-Han Hsiao (Santa Clara University)
Abstract: Smart refrigerators have demonstrated the potential of technology-assisted household food management by helping users track stored food. However, these systems can be costly and rely on dedicated appliances, limiting accessibility and adoption. Therefore, we present EcoChef, a mobile application that brings similar household food-management support to a widely available smartphone-based workflow without requiring specialized hardware. Users take one photo of a mixed set of groceries, and EcoChef recognizes multiple items at once and adds them to a persistent household inventory. This inventory supports expiration awareness, meal planning, dietary preferences, and waste-aware recipe recommendations, while confirmed cooking activities update monetary and sustainability estimates derived from public data sources. The supporting OWLv2 vision module achieves 90.82\% matched household-category accuracy on the public D2S validation set. By combining one-photo grocery recording, persistent inventory management, meal planning, dietary support, and interpretable food-waste feedback, EcoChef provides an accessible one-stop workflow for everyday household food management.
#116 | SAFE-Chat: A Modular Framework for Graph-Based Self-Harm Risk Detection in Conversational AI
Smita Ghosh (Santa Clara University)*; Xiaochen Luo (Santa Clara University); Thai Nguyen (Santa Clara University); Jonah Engelmann (Santa Clara University); Karl Heinemeier (Santa Clara University); Atharva Sonar (Santa Clara University); Tharun Venkatesh (University of Waterloo); Shivvrat Arya (New Jersey Institute of Technology)
Abstract: Conversational AI systems are increasingly used as trusted interfaces for emotional support. Recent incidents have raised concerns about whether these systems can reliably recognize and respond to implicit self-harm signals. Such risk may emerge gradually across multiple turns and remain difficult for general-purpose conversational models to recognize reliably. To address this, we propose SAFE-Chat, a Self-harm Aware Framework for Escalation in Chat-based AI. SAFE-Chat complements the underlying model's safety mechanisms with a model-independent monitoring module trained on both overt and implicit self-harm trajectories. We construct a synthetic benchmark of multi-turn interactions and formulate this monitoring task as a conversation-level graph classification problem. Compared with baselines, our fine-tuned Confidence-Oriented GCN achieves an AUROC of 0.9667 and a Macro-F1 score of 0.8804 on a three-class classification task. These results support model-independent monitoring as a complement to safeguards embedded within conversational models. Ultimately, this work invites broader discussion on whether future AI safety should rely on capabilities embedded within foundation models or be reinforced through independent oversight.
#117 | Machine Learning Guided Design of Thermostable PET-Degrading Enzymes to Reduce `Persistent Polymer Exposure Risks
Ayush Iyer (Dougherty Valley High School)*; Abhinav Iyer ( Dougherty Valley High School)
Abstract: Recent surges in plastic pollution from polyethylene terephthalate (PET) has created an urgent need for faster and sustainable recycling technologies. PET-degrading enzymes offer a promising solution, but many lose activity under the heat, controlled pH, metal ions, and interfaces found in industrial reactors. Because laboratory screening of stabilizing mutations is costly and slow, computational methods are needed to identify strong candidates before laboratory testing. In this study, we developed a machine learning platform that predicts protein mutations to optimize thermal stability. We compiled 1,001,888 protein stability records from seven public databases and developed a novel dual-stage framework approach for plastic-based protein engineering, primarily focused on PET hydrolases. It consists of a protein screener predicting melting temperature (Tm) and a mutation optimizer that predicts folding free energy change (ÎÎG). We represented proteins using ESM-2 protein language model embeddings, which were then modeled with gradient-boosting ensembles. The ensemble was evaluated on BRENDA and S669 benchmarks, representing optimum temperatures and ÎÎG values respectively. On these benchmarks, the ensemble received an ROC of 0.732 (95% CI 0.708â0.755) and 0.669 (95% CI 0.62â0.72). Because the intended use of our pipeline is candidate development, we prioritized precision over recall. At strict thresholds, the framework reached high precision (0.93 and 0.81) with lower recall values (0.39 and 0.10). Thus, our pipeline yields results strongly enriched for accuracy. This framework aims to develop high performing candidates for experimental testing while enabling condition-aware protein optimization, an area that remains largely unexplored. This approach could be particularly valuable in recycling plants and waste management systems, where optimized enzymes could be integrated into treatment processes to help break down microplastics and other plastic waste.
#118 | From Search to Action: Multi-Stage RAG for Sequential Planning
Izat Temiraliev (University of California Santa Cruz)*
Abstract: When a robot encounters an unseen long-horizon task, the knowledge it needs usually exists, but inside a reference document it must first find. Retrieval-augmented robotics showed that grounding a planner in a retrieved visual manual materially improves assembly planning, but assumed the matching manual could be identified directly; in the field, the input is a single camera view of the unassembled parts. We present RAG-Planner, which closes this perception-to-plan loop. A two-stage retriever, a frozen DINOv3-7B encoder proposing candidates and a 2.1M-parameter learned cross-encoder re-ranking them, recovers the manual from a real customer photograph at 88.5% recall@1 and 96.9% recall@5 on held-out items; a four-stage visionâlanguage planner then converts the retrieved manual into an executable assembly tree. The two halves respond to different interventions: retrieval is closed by a re-ranker that trains in about a minute, while planning is bounded by backbone capacity; fifteen inference-time techniques move an open 32B baseline by at most two points, one frontier-backbone swap lifts strict-tree accuracy up to 2.8Ă.
#119 | AuditLM: A Framework to Trace AI Reasonings and Verify its Generated Outputs
Reza Fayyazi (Rochester Institute of Technology)*; Michael Zuzak (Rochester Institute of Technology); Shanchieh Jay Yang (Gonzaga University)
Abstract: When a language model produces a grounded-looking answer with citations attached, we currently have no principled way to ask the two questions that actually matter: where did this text come from, and do the references it points to match to the generated text? Fluency and formatting are often mis-treated as proxies for provenance. This work argues that provenance is measurable, not as a single trust score, but as an auditable trail. We present AuditLM, an open framework that measures provenance through three complementary lenses operating at different granularities to verify AI reasonings
#120 | đ€ Supervision-Efficient Robot Learning in Deployment
Lars Ankile (Stanford)*
Abstract: Learning from human demonstrations is remarkably data-efficient, but often yields diminishing returns, especially for high-precision tasks with wide initial-state distributions. Continued improvement can instead come from deployment, where a supervised policy performing the task produces autonomous rollouts and receives human corrections. Common methods use only one signal each: interactive imitation learning learns from the corrections, real-world reinforcement learning from the rollouts. We combine the two in a batch-online loop (run a frozen policy, collect a batch under supervision, update offline, repeat), which also lets us choose where to collect: failures concentrate in parts of the initial-state space, so observed performance steers each round's initial states toward where supervision buys the most improvement. The learning side strengthens interactive imitation with a reinforcement learning objective that learns from the full stream, including autonomous failures and successes. Across three real-world high-precision tasks and simulation, the two axes deliver complementary gains over a uniform interactive imitation baseline, round after round. Collection itself succeeds at near-perfect rates, so the robot does productive work while it learns.
#121 | ARGUS: Efficient Passive Person Identification from Wi-Fi CSI
Nayan Bhatia (UCSC)*
Abstract: Passive person identification is attractive for access control and personalized environments because it can operate without cameras, wearable devices, or deliberate user actions. Yet it is difficult to identify people from commodity Wi-Fi channel state information (CSI): the signal is high dimensional, noisy, and coupled to the room and radio placement. We introduce ARGUS, a compact pipeline that converts short spans of CSI into statistical maps (âstatgramsâ) and classifies them with a lightweight decoder-only Transformer. The representation summarizes calibrated amplitude and phase views across time and subcarriers, while attention-guided occlusion identifies a small subset of informative regions. On a 154-person dataset using a strict physical-segment split, ARGUS obtains ~78.88% Top-1 accuracy from 6-second windows and ~84.85% after combining a minute of evidence. It improves 60-second identification by 7.75 percentage points over a raw-CSI Transformer while requiring 4.4Ă fewer FLOPs per window. Tests on a multi-user benchmark further show that compression can preserve accuracy across rooms and Wi-Fi bands, while cross-room transfer and open-set rejection remain important deployment challenges. These findings suggest that compact statistical representations can make passive wireless identification more scalable and more interpretable.
#124 | Local Hebbian Learning for Fine-Tuning Spiking Neurons
Russ Webb (Apple)*; Sophia Webb (Valley Christian)
Abstract: Adaptability is an essential attribute in biological and artificial neural networks. We present experiments highlighting the success of local reward-modulated Hebbian [2] learning in tuning a spiking neural network for accuracy and stability. This improvement is illustrated through a line-following task, where the model uses input from a simulated 1D retina to follow a line. The model is updated via a per-neuron learning rule and compared to its initial simple pre-programmed behavior. Experiments were conducted to tune the synapse learning rate, memory retention of previous reward spikes, probability of the neural networks synapses spiking, and the initial model weights of the integrate-and-fire neurons. Improved task performance and greater stability under out-of-distribution hyper-parameter perturbations are achieved by the tuned model.
#126 | Closing the Reasoning Gap in Graph-RAG with Structured Decomposition and Graph-Aware Compression
Yasaman Zarrinkia (University of Victoria); Venkatesh Srinivasan (Santa Clara University)*; Alex Thomo (University of Victoria)
Abstract: Graph-based retrieval-augmented generation (Graph-RAG) retrieves evidence from documents as knowledge graphs to support multi-hop question answering, but access to the correct facts does not guarantee that a language model will combine them accurately. Across three benchmarks, the gold answer appears in 77--91\% of the retrieved contexts, yet accuracy is only 35\%--78\% and 73--84\% of all errors occur even when retrieval succeeds. We propose two inference-time interventions that require neither retraining nor re-indexing: a SPARQL-inspired decomposition prompt and an LLM-free graph-walk procedure that reduces context size by approximately 60\%. When paired with question-type routing, an 8B open-weight model matches or outperforms an unaugmented 70B model on all three datasets while operating at roughly one-twelfth of the per-token cost. Our experiments with LightRAG show that our augmentations transfer across Graph-RAG systems.
#127 | Who Endorsed It? Measuring Authority Bias Across Expertise Levels in Language Models
Priyanka Mary Mammen (UMass Amherst)*; Emil Joswin (Independent Research); Shankar Venkitachalam (Independent Research)
Abstract: Prior research demonstrates that the perfor- mance of language models on reasoning tasks can be influenced by suggestions, hints, and en- dorsements. However, the influence of endorse- ment source credibility remains underexplored. We investigate whether language models ex- hibit systematic bias based on the perceived expertise of the provider of the endorsement. Across 4 datasets spanning mathematical, legal, and medical reasoning, we evaluate 11 mod- els using personas representing four expertise levels per domain. Our results reveal that mod- els are increasingly susceptible to incorrect or misleading endorsements as source expertise increases, with higher-authority sources induc- ing not only accuracy degradation but also in- creased confidence in wrong answers. We also show that this authority bias is mechanistically encoded within the model and a model can be steered away from the bias using, thereby im- proving its performance even when an expert gives a misleading endorsement.
#128 | MathAtlas: Benchmarking Autoformalization in the Wild
Nilay Patel (University of California, Santa Cruz)*; Jeffrey Flanigan (University of California, Santa Cruz)
Abstract: Current autoformalization benchmarks are largely focused on olympiad or undergraduate mathematics, while graduate and research-level mathematics remains underexplored. In this work, we introduce MathAtlas, the first large-scale autoformalization benchmark of in the wild graduate-level mathematics, containing ~52k theorems, definitions, exercises, examples, and proofs extracted from 103 graduate mathematics textbooks. MathAtlas is enriched with a mathematical dependency graph containing ~178k relations, and is the first autoformalization benchmark to include such relations, facilitating evaluation and development of dependency-aware autoformalization systems. Our extensive experiments show that MathAtlas is high quality but challenging: strong single-pass baselines achieve at most 9.8% correctness on theorem statements and 16.7% on definitions. Furthermore, we find performance degrades substantially with dependency depth: on MA-Hard, a subset of 700 entities with the deepest dependency trees, the best model achieves only 2.6% correctness for autoformalization on this challenging dataset. We release MathAtlas to the community as a benchmark set for large-scale autoformalization of graduate-level mathematics in the wild. In ongoing and future work, we continue exploring more complex, dependency- and feedback-aware autoformalization systems on MathAtlas, expanding MathAtlas to research-level mathematics, discuss pitfalls in current semantic faithfulness metrics, and investigate methods to improve them.
#129 | The Clinical Utility of a Promising Biomarker for Two Overlapping Lung Diseases: What 26 Cases Support
Jordan Smith (Santa Clara University)*
Abstract: Pulmonary veno-occlusive disease (PVOD) and idiopathic pulmonary arterial hypertension (IPAH) present alike and are easily confused, and the standard PAH vasodilators can cause life-threatening pulmonary edema in a patient who has PVOD instead. A recent research letter proposes a blood biomarker, GDF15, and a single cutoff to separate them. We reanalyzed the published patient-level data to ask not whether GDF15 separates the two groups overall, but what evidence one patient's value provides. GDF15 far outperformed routine clinical variables (AUC 0.956 versus 0.673). But replacing the cutoff with a continuous likelihood ratio exposed an inconclusive band around the threshold, and a high-value region where the data cannot support a precise estimate. Standard internal validation moved the aggregate score by 0.0001; leave-one-case-out refitting moved the estimate at the edge of the data by 15%. Strong aggregate validation does not establish that every individual prediction is equally supported.
#131 | Mem-Engram: A Self-Evolving Typed Memory System for LLM Agents with Iterative Retrieval
Shanglin Yang (Google); Aria Shi (Santa clara university)*; Li Yin (SylphAI); Haebin Seong (SylphAI); Haoran Zhang (SylphAI)
Abstract: LLM agents require persistent cross-session context to preserve architectural decisions, known bugs, and user preferences. Existing memory systems typically impose structure at write time, through mechanisms such as pre-indexed RAG or hierarchical memory paging. However, multi-hop retrieval remains challenging when relevant evidence becomes identifiable only after partial information has been recovered, while prior solutions rely on costly query decomposition, planning, and reflection. We present Mem-Engram, a persistent memory framework based on the principle that raw event storage combined with LLM-time interpretation can outperform pre-structured indexing. Mem-Engram introduces a five-type memory taxonomy that serves only as a confidence-weighted index over untouched raw events, rather than as the source of truth. It also employs an iterative retrieval engine that resolves multi-hop queries using a single coverage check and targeted follow-up retrieval. These components improve multi-hop and multi-entity retrieval while reducing the multi-call overhead of prior memory-augmented pipelines. Mem-Engram further supports a self-evolving loop in which retrieval configurations are refined across sessions based on diagnosed task outcomes.
#133 | Evidence-Grounded Narrative Understanding for Long-Form Professional Video
Haoyu Zhang (Eluvio Inc)*
Abstract: Long-form video understanding requires reasoning over minutes to hours of multimodal content while maintaining factual consistency and narrative coherence. Existing Video-LLMs and retrieval pipelines rely on temporal chunking or flat aggregation, forcing the model to reconstruct semantic structure implicitly under a constrained context budget, which leads to redundancy, fragmented narratives, hallucinated reasoning. We argue that the core bottleneck is not perception quality, but the absence of explicit semantic memory organization. We propose Hierarchical Evidence Memory (\hem), which transforms raw multimodal observations into progressively abstract representations through three layers: perceptual evidence aggregation, semantic evidence binding, and narrative evidence memory. Given these memories, \hem{} answers in one generation pass--no iterative agentic reasoning, no large-context decoding. On a diversified medium-to-long summarization benchmark, the structured levels beat a flat baseline granted the model's full 262k context (event coverage up $13\%$, unsupported claims down $16\%$, paired bootstrap $p<0.05$) at 7.5--16$\times$ fewer prompt tokens and up to 47$\times$ lower latency, and position-swapped pairwise judges prefer the most interpreted level on quality, completeness, and coherence.
#135 | Untethered Continuous Thought: Outcome-Only Supervision for Latent Reasoning
Aria Shi (Santa clara university)*; Zhan Shi (Google.com)
Abstract: Latent-reasoning methods let language models compute in continuous hidden-state space, with a key motivation, articulated by COCONUT, that continuous thoughts can represent a superposition over alternative reasoning paths, or an ``implicit BFS.'' Yet existing methods often tether latent steps to tokens from a single reference trace, potentially collapsing this superposition toward one path. We study this tension with a controlled frontier-tracking task whose ground-truth belief state is exactly the BFS frontier, making reasoning-path superposition directly measurable. On an identical continuous-thought backbone, we compare token-tethered training with \textbf{Untethered Continuous Thought (UCT)}, which uses final-answer supervision together with latent consistency and verifier-based halting, without pulling latent states toward a reference path. UCT improves reasoning accuracy and training stability, while frontier-recovery analysis shows that tethered supervision preferentially encodes the supervised path whereas UCT better preserves alternatives. Surprisingly, UCT is also the least linearly decodable into the symbolic frontier despite reasoning best. These results establish supervision target as a first-class design choice for latent reasoning and caution against equating latent legibility with reasoning quality.
#136 | TRACE: Agentic Catalog Enrichment with Multi-Source Evidence Grounding
Rohan Kumar (DoorDash)*; Steven Xu (DoorDash); Kyle MacDonald (DoorDash); Sudeep Das (DoorDash)
Abstract: Structured attributes drive search, filtering, and recommendation in e-commerce, yet seller-provided catalogs are chronically attribute-sparse. TRACE, our agentic LLM enrichment framework deployed on an industry-scale marketplace catalog, separates proposing values from deciding what may be published: a ScoutAgent triangulates evidence across catalog text, syndicated feeds, product imagery, and identity-matched web search, and a JudgeAgent sits in the write path, re-verifying every candidate before it reaches the catalog. On a human-labeled audit, extracted values were 98.2% accurate at 74.7% attribute coverage. In production, enrichment now spans 31 million product records with 90.4% impression-weighted coverage across four retail verticals, and a randomized experiment surfacing the enriched attributes lifted checkout conversion by 0.48%. We share design lessons for systems that write LLM output into durable, customer-facing state.
#137 | When Benchmarks Break: Stress-Testing Agent Benchmarks with Behaviorally Grounded User Personas
Jeshwanth Bheemanpally (UCSC)*; Yash Malegaonkar (UCSC)
Abstract: Benchmarks for agentic systems summarize capability as a single task success rate [4, 8], computed against fixed tasks and one generic, goal-directed user simulator [2, 3]. Real users vary on factors, such as tone, patience, tolerance for policy friction, and willingness to escalate or abandon. An agent that tops a leaderboard against a compliant simulator may behave inconsistently across user groups, which is a liability that aggregate scores are structurally unable to report. Existing user-centric evaluations rely on hand-crafted profiles, or treat the persona as the evaluation target [6]. We treat persona-conditioned simulators as an evaluation instrument: holding tasks, tools, and rewards fixed, we perturb only the user across a population of behaviors grounded in real behavioral artifacts. We do not claim to estimate performance on the "true user population," rather we ask whether benchmark conclusions are stable under such user variation derived from real world behavior. In addition, we study whether personas conditioned on real behaviorial artifacts can surface plausible errors that standard synthetic user simulators may not find, either on the fault of the agentic harness or underlying model itself.
#138 | Mitigating Stylometric Attacks in LDP Text Generation via Latent Separation
Aria Shi (Santa clara university)*; Yefeng Yuan (Santa clara university); Liang Cheng (ebay.com); Yuhong Liu (Santa clara university)
Abstract: Privatized text rewriting under local differential privacy (LDP) enables release of sensitive documents with formal per-input privacy guarantees. However, existing content-level DP rewriting methods perturb a monolithic representation in which semantic content and writing style are entangled, leaving a style privacy gap: at utility-preserving noise levels, generated text retains stylometric signals exploitable by authorship attribution models, whereas stricter budgets reduce attribution mainly by degrading coherence. We formalize \emph{author-attribute DP}, which protects author identity at fixed content, and identify an \emph{author-conditional sensitivity} $\du$ that prices disentanglement quality in $\eps$ units. We propose a per-document DP rewriting framework separating semantic content from writing style with asymmetric protection: the semantic channel is privatized via Public-Anchored PCA-DP, projecting a clipped semantic residual onto a public PCA subspace with calibrated Gaussian noise, while source style is replaced by a sampled public donor style, on a BART encoder--decoder backbone. Since the decoder accesses only privatized semantic information, fixed public artifacts, and donor style, each document satisfies a formal DP guarantee by post-processing. Experiments show the method substantially reduces attribution accuracy while preserving fluent generation and utility.
#141 | Generalization in Prompt Optimization for LLM Judges in E-Commerce Relevance Labeling
Jinming Nian (Santa Clara University)*; Zongjun Tan (Walmart Global Tech); Rakesh Malviya (Walmart Global Tech); Prateek Verma (Walmart Global Tech); Eva Song (Walmart Global Tech); Yi Fang (Santa Clara University); Matthew Mu (Walmart Global Tech)
Abstract: E-commerce search systems rely heavily on query-item relevance judgments as supervision and ranking features, as well as for debugging and sanity checking. Traditionally, these judgments are produced by expensive human labelers following a large, continuously updated policy manual, which leaves room for inconsistent guidance and interpretation. LLM judges offer a scalable alternative, but their performance still depends on the rubric prompt derived from these manuals. We study whether automatic prompt optimization can improve an LLM query-item relevance judge by comparing full-prompt rewriting, diff-based GEPA editing, island-based evolutionary search, and divide-and-conquer evolution. On a development set stratified by label, all methods improve over using the original manual as prompt. However, these gains generally fail to transfer to a much larger held-out set. The best held-out result comes from a short GEPA run guided by the smallest development subset, which outperforms the baseline, while more aggressively optimized prompts regress. We find that validation data is the main bottleneck rather than the optimization method. E-commerce relevance data is highly skewed toward relevant query-item pairs, and preserving the label distribution alone does not provide sufficient coverage for reliable prompt selection. These preliminary findings suggest that prompt optimizers can reliably optimize on the evaluator metric, but better evaluator signal design is needed to turn optimization gains into robust improvements in LLM relevance judging.
#142 | Refusal Steering Creates Capable and Dangerous Agents
Vincent Siu (UCSC)*; Vlad Pavlovich (UCSC); Glenn Grant-Richards (UCSC); Navodit maheshwari (UCSC); Yujin Potter (UC Berkeley); Dawn Song (UC Berkeley); Chenguang Wang (UCSC)
Abstract: Refusal direction steering is an established weight level jailbreak for single turn language models. We test whether it transfers to the agentic setting on four open weight families spanning dense, mixture of experts, and hybrid attention architectures. Refusal collapses across all four, with AgentHarm refusal falling from a released range of 29\% to 85\% down to 0\% to 3\% and harmful task completion reaching 73\% on steered GPT-OSS-120B. General knowledge, function calling accuracy, and hazardous procedural knowledge are largely preserved. Indirect prompt injection resistance survives on three of four families. Any user with download access can therefore produce a capable harmful agent by running a known recipe.
#143 | FaultLoc: Evaluating Coding Agents for Vulnerability Localization From Crash to Cause
Jianhong Tu Tu (UCSC)*; Shubham Gaur (UCSC); Rathik Murtinty (UCSC); Chenguang Wang (UCSC); Zhun Wang (UCB); Tianneng Shi (UCB); Dawn Song (UCB)
Abstract: Coding agents powered by large language models have demonstrated strong performance on security tasks such as vulnerability discovery and automated patching. However, existing benchmarks primarily evaluate patch generation through functional tests, allowing agents to succeed by mitigating symptoms at the crash site without identifying the underlying root cause. As a result, it remains unclear whether these systems locate the defect or merely suppress the observed failure. We introduce FaultLoc, a benchmark for crash-to-cause fault localization in large codebases. Given the vulnerable codebase, a sanitizer report, and a proof-of-concept input, agents must identify the code region responsible for the vulnerability, evaluated against developer patches. FaultLoc comprises 900 real-world memory-safety vulnerabilities with verified PoCs and fixes across 124 C/C++ open-source projects. Across 10 models and 4 agent harnesses, Opus 5 with Claude Code leads at 58.1 Hunk Recall. Tracing the root cause from the crash site remains challenging even when agents identify the file where the fault resides. Agents anchor on the stack trace, and removing it degrades performance sharply, while the PoC alone is not sufficient to localize the fault. These results highlight that strong code reasoning skills are required to trace the fault to its origin.
#144 | Knowledge Distillation for Modern YOLO
Timothy Tong (Santa Clara University)*
Abstract: When deploying robots on the edge, researchers face constraints ranging from network bandwidth to edge compute limitations. Under these conditions, it is essential to have small models that are both easily transmittable and fit on limited hardware, while maintaining adequate performance. Knowledge distillation is a common model efficiency technique that seeks to boost the performance of a small model to match that of a larger model. Our contributions are as follows: (1) a multi-seeded evaluation of a YOLO distillation method on a small object benchmark, demonstrating positive results, (2) an ablation revealing augmentation as the component that flips distillation gains in mAP from positive to negative, and (3) preliminary findings as to where in the network the teacher fails to transfer knowledge to the student.
#148 | Data Scout: Targeted Web Crawling for Domain-Specific Pretraining Corpora
Chirag Garg (IBM)*
Abstract: Domain-specific pretraining corpora are usually built by filtering a general web archive. When the target domain is rare, the filter rejects nearly everything it reads and cannot recover pages the archive never crawled. We introduce Data Scout, which uses a domain classifier to decide where to crawl. An LLM generates search queries; the returned seed URLs are grouped and surveyed by hostname before a deep crawl begins. Against a random RedPajama-v2 sample scored by the same classifier, Data Scout raises the pass rate from 0.31% to 21.9% in mathematics (70x) and from 0.63% to 51.9% in astronomy (82x). In mathematics, pages on the seed hostname pass at 28.5%, compared with 1.33% on sibling hostnames, and 63.2% of sampled URLs are absent from CommonCrawl. Continued pretraining of Llama-3.2-3B on 1.9B Data Scout tokens matches training on an equally sized FineMath corpus on GSM8k.
#150 | How Does Retrieval Topology Shape Memory Failure? A Controlled Ablation for Agentic Memory Systems
Alexander Anokhin (Santa Clara University)*; Rynne Whitnah (IBM)
Abstract: LLM-based agents increasingly require external memory to operate across different sessions and users, but existing agentic memory systems bundle many design choices, making it difficult to tell which components help and which hurt. We present a controlled, single-variable ablation of five retrieval topologies over a shared extraction and embedding pipeline, evaluated on the LongMemEval and BEAM benchmarks. Our preliminary findings include: (1) topology rankings change across benchmarks, with co-mention expansion winning on sparse conversation histories but typed-edge routing winning on dense ones, (2) a systematic failure mode where graph-expanded candidates are buried by the cross-encoder reranker due to missing path traversal context, addressable by a simple path-injection technique, and (3) edge-only memory storage performing more weakly on most question types but uniquely scoring on temporal, summarization, and contradiction categories. These observations suggest that the right topology depends on the workload, which might be hidden by single-benchmark evaluations.
#151 | LLM Chess: Benchmarking Reasoning and Agentic Reliability through Full-Game Play
Sai Kolasani (UC Berkeley); Maxim Saplin (Independent Researcher); Nicholas Crispino (UCSC)*; Kyle Montgomery (UCSC); Jared Quincy Davis (Stanford); Matei Zaharia (UC Berkeley); Chi Wang (AG2); Chenguang Wang (UCSC)
Abstract: We introduce LLM Chess, an evaluation framework that probes the generalization of reasoning in large language models (LLMs) by having them play complete games of chess against a weak random move player and a chess engine opponent of variable strength. We play 200 configurations of 159 models and rate them on a single Elo scale, scoring separately how much of each game a model is able to finish. Because every position is generated during play and difficulty is controllable, the benchmark can resist saturation, as shown by two years of continuous evaluation. The best performing model still only reaches 1548 Elo, about half way up the chess engine opponentâs calibrated ladder.
#152 | Cracking Reasoning in Diffusion Models: Evolution Strategies Sidestep the limitations of Reinforcement Learning in Diffusion Large Language Models
Roberto Dailey (Cognizant AI Lab)*
Abstract: Diffusion Large Language models represent a new frontier for language modeling research, allowing for bidirectional reasoning and faster inference than standard autoregressive models. Yet their post-training pipeline remains bottlenecked by something they cannot provide: unbiased token sequence log-probabilities. Numerous attempts have been made at working around this limitation within a reinforcement learning paradigm, yet methods either introduce biases that hurt performance, or introduce structures that cannot be extended to long-horizon workflows. This preliminary work shows that these limitations can be side-stepped by using Evolution Strategies as a replacement for existing RL methods, achieving high performance while being naturally extendable to multi-hop, long-horizon, and multi-agent learning environments.
#154 | A Powerful Probe of the Cosmos: Agentic Strong Gravitational Lens Modeling
Sean Xu (UC Berkeley)*; Xiaosheng Huang (Santa Clara University); Aniket Jain (UC Berkeley)
Abstract: Strong gravitational lensing is on the cusp of a new, data-rich era. New wide-field survey telescopes are expected to discover and confirm hundreds of thousands of new gravitational lenses, $\sim1000\times$ more than the number of confirmed lenses today. These gravitational lenses provide insight into the expansion history of the universe, and the nature of dark matter and dark energy. Unlocking these insights requires large-scale, accurate modeling of these systems. Each lens model must be fast and statistically robust, enabled by the advent of highly parallelized, differentiable lens modeling codes running on GPUs. Even with such codes, however, human inspection is required to ensure convergence and physicality of the model. We demonstrate the ability of frontier large language models (LLMs) to produce successful lens models with minimal human intervention. This will lead to an automated pipeline to model $O(10^5)$ strong lens to measure the properties of dark matter and energy.
#155 | Harness as Model: Training Executable Agent Infrastructure from Verifier-Grounded Traces
Aria Shi (Santa clara university)*; shanglin yang (Google); li yin (sylph.ai); Haebin Seong (sylph.ai); Haoran Zhang (sylph.ai)
Abstract: A language-model agent is a fixed base model wrapped in software: prompts, tools, execution policies, context management, retries, and the agent loop itself. That wrapper is where most engineering effort goes and where most failures live, yet it is treated as infrastructure rather than as something learnable. The obstacle is not that harness state is unlearnable but that improving it safely is hard: the object is executable code that must survive a real environment, the evidence for what to change is buried in traces rather than exposed as a loss, the reward comes from a verifier that must stay unreachable by the thing being optimized, and infrastructure failures mimic capability failures. We propose \emph{Harness as Model}, giving agent infrastructure a model-like lifecycle and report two implementations. The first treats declared prompt and code parameters as trainable state, updated transactionally against a benchmark verifier. Its outcome motivates the second: we argue the binding limit is the \emph{proposal channel}, since a structured call filling a declared slot cannot add a tool, restructure a loop, or read the trace that would say which to do. The second therefore makes the candidate a whole harness module and the proposer a coding agent with trajectory access, selects by a frontier over per-task results, and adds two safeguards: rejecting candidates that encode benchmark knowledge, and classifying each trial by whether the harness was exercised before its reward counts. We contribute the lifecycle, a safety policy keeping measurement outside the search space, the diagnosis relating proposal expressiveness to what can be learned, and evidence that a trajectory-reading proposer surfaces defects a parameter-leveloptimizer structurally cannot.
Call for Abstracts
BayLearn 2026 will be an in-person event, held on Thursday, October 15th, 2026.
BayLearn 2026 will be hosted at Santa Clara University (Santa Clara, CA) and co-organized with San José State University, UC Berkeley, and UC Santa Cruz.
Note: BayLearn 2026 will not be a hybrid event, and it will not be live-streamed.
The BayLearn 2026 abstract submission site is now open for submissions:
Submit via Microsoft CMT: https://cmt3.research.microsoft.com/BAYLEARN2026
Please submit abstracts as a 2-page pdf in NeurIPS 2023 format. An extra page for acknowledgements and references is allowed.
About BayLearn
The BayLearn Symposium is an annual gathering of machine learning researchers and scientists from the San Francisco Bay Area. While BayLearn promotes community building and technical discussions between local researchers from academic and industrial institutions, it also welcomes visitors. This one-day event combines invited talks, contributed talks, and posters, to foster exchange of ideas.
Meet with fellow Bay Area machine learning researchers and scientists during the symposium that will be held on October 16th, at Santa Clara University.
Feel free to circulate this invitation to your colleagues and relevant contacts.
Key Dates
- Updated Abstract submission deadline: Thu, Aug 6th, 2026 at 11:59pm PDT
- Author notifications: Mon, Sep 14th, 2026
- Registration Lottery opens: Tue, Sep 15th, 2026
- Symposium: Thu, Oct 15th, 2026
Submissions
We encourage submission of abstracts. Acceptable material includes work which has already been submitted or published, preliminary results, and controversial findings. We do not intend to publish paper proceedings; only abstracts will be shared through an online repository. Our primary goal is to foster discussion! For examples of previously accepted talks, please watch the paper presentations from previous BayLearn Symposiums: https://baylearn.org/previous
Submit your abstracts via CMT: https://cmt3.research.microsoft.com/BAYLEARN2026
Relevant Topics
We welcome submissions on the theoretical, experimental, applied and positional arguments on Machine Learning. We invite abstracts on all topics related Machine Learning, including but not limited to the topics highlighted in NeurIPS:
- Computer vision
- Language and multimodal language models
- Robotics, embodied systems, and engineering
- AI/ML for physical sciences
- AI/ML for health and biotechnology
- AI/ML for sustainability
- AI/ML for social sciences
- AI/ML for creatives
- Neuroscience and cognitive science
- Socio-technical aspects of AI
- Human interaction in AI systems
- Decision-making, reinforcement learning, and control
- Generalization and multi-task learning
- Optimization
- Probabilistic methods
- AI and network science
- Data-centric aspects of AI
- SysML Infrastructure
- Theory
- Deep learning
- General machine learning: core contributions in supervised and unsupervised methods
Submission guidelines
All submissions must be made before the deadline to Microsoft CMT.
Please use the Latex template for NeurIPS, available in this link: NeurIPS format
While abstracts can be based on work that has been published elsewhere, the abstract should be original text of a self-contained idea and not copy-paste from a previously accepted paper.
All submissions must be anonymized and must not include any information that could intentionally or unintentionally compromise the double-blind review process. Authors may share versions of their work on preprint platforms such as arXiv. The conference will be held in-person. At least one author for each accepted abstract is requested to physically attend the conference to present their work.
The organizers will desk reject submissions based on the following criteria: * Have formatting issues, e.g., not following the formatting guidelines or template. * Exceed the 2 page limit. This includes using hyperlinks for extended content. * Have non-anonymized submissions.
High quality abstracts satisfactorily address the following:
- Motivation & Problem Framing
- Is the problem clearly stated?
- Is there a compelling reason this work matters (scientifically or practically)?
- Novelty / Idea Contribution
- Does the work introduce a new idea, perspective, or research question?
- Clarity and Organization
- Is the abstract understandable?
- Is the abstract well-structured and relatively self-contained?
- Results / Proofs / Findings
- Does the abstract include evidence?
- Evidence includes but is not limited to experiments, qualitative results, case studies, examples, analysis, etc. Preliminary results are acceptable (and encouraged)
- Conclusions and Discussion
- Does the work provide any conclusions, insights or lessons?
- Does this work invite discussion or debate? Would this spark conversation at the event?
- Broad Applicability
- Why should the broader ML community care about this? What is the broader impact of this work?
- Is this exciting to a larger audience?
Note: The Microsoft CMT service was used for managing the peer-reviewing process for this conference. This service was provided for free by Microsoft and they bore all expenses, including costs for Azure cloud services as well as for software development and support.