SF

Accepted Papers

All accepted papers are presented as posters. Find your paper below to see its assigned poster session. Paper titles link to their OpenReview pages.

Paper AwardsPresented at the closing session, 17:15 - 18:00

Poster Session I12:00 - 13:30

CLIO-BENCH: Beyond Accuracy in AI Historical JudgmentPoster · Poster Session IControlling for Feature Frequency in Sparse Autoencoder Steering of Mathematical ReasoningPoster · Poster Session IDemystifying Classifier-Free Guidance for Auto-Regressive Image GenerationOral · Poster Session IDimension Is Free, Horizon Is Not: When Evolution Strategies Rival Policy Gradients for LLM Fine-TuningPoster · Poster Session IBest Paper Runner-upDiscretizing Reward ModelsPoster · Poster Session IDo as I Say, Not as I Do: Instruction-Induction Conflict in LLMsPoster · Poster Session IDon’t Blink: Evidence Collapse during Multimodal ReasoningPoster · Poster Session IDr. Post-Training: A Data Regularization Perspective on LLM Post-TrainingPoster · Poster Session IEconomy of Minds: Emerging Multi-Agent Intelligence with Economic InteractionsPoster · Poster Session IEncoded Early, Used Late: Where Transformers Begin to Act on an Inferred Partner's ExpertisePoster · Poster Session IFinding the Minimal Parameter Budget for Implicit Reasoning: A Data Complexity Driven Scaling Law for Language ModelsPoster · Poster Session IFrom Refusal to Fabrication: An Affective Cause of Hallucination That Bypasses the Honesty GatePoster · Poster Session IH-Probes: Extracting Hierarchical Structures From Latent Representations of Language ModelsPoster · Poster Session IHow Faithful Is Trajectory-Based Data Attribution? Error Sources, Remedies, and Practical GuidelinePoster · Poster Session IHow Much Do Circuits Tell Us? Measuring the Consistency and Specificity of Language Model CircuitsOral · Poster Session IINFUSER: Influence-Guided Self-Evolution Improves ReasoningPoster · Poster Session IMatching the Distribution, Not the Mean: Learning Human Rating Distributions with VLMsPoster · Poster Session IMeasuring Representation Reorganization from Language Models to Vision-Language ModelsPoster · Poster Session IMode-seeking in on-policy distillation of language modelsPoster · Poster Session INeuron Populations Exhibit Divergent Selectivity with ScalePoster · Poster Session IBest PaperPlaying Psychic: Using Thought Trees to Predict Reasoning Models Accuracy on Coding TasksPoster · Poster Session IRoundtable Policy: Scaling Multi-Agent Reasoning with Confidence-Weighted-ConsensusPoster · Poster Session IShorter Context, Fewer Negatives: How Information Reduction Biases LLM Scientific ReasoningPoster · Poster Session ISingle Canonical Prompts Underestimate LLM Safety's Surface-Form SensitivityPoster · Poster Session ISingle-cell foundation models obey data scaling laws on the optimized lossPoster · Poster Session IBest Paper Runner-upStructure Before Collapse: Transient semantic geometry in next-token predictionPoster · Poster Session IThe Future of Facts: Tracing the Factual Generation-Verification GapOral · Poster Session IBest PaperToken GeometryPoster · Poster Session ITowards Physical Intuitions for Alignment Dynamics: A Case Study With Randomness CrystallizationPoster · Poster Session ITwistBench: Benchmarking Transformational Creativity in LLMs via Literary Plot TwistsOral · Poster Session IWhen Do Looped Transformers Need More Loops? Calibrated Measurement of External Stop SignalsPoster · Poster Session IWhy GRPO Needs Normalization: A Local-Curvature Perspective on Adaptive GradientsPoster · Poster Session IYou Only Judge Once: Multi-response Reward Modeling in a Single Forward PassPoster · Poster Session I

Poster Session II17:15 - 18:00

Algorithmic Blindness in Large Language Models: A Calibration Study of Performance PredictionPoster · Poster Session IIBoNVoyage: Learning Better Rewards without RankingPoster · Poster Session IICharacterizing Learning-Resistant PreferencesPoster · Poster Session IIClayBuddy: A Framework, Evaluation, & Mitigation of Coding Agent FailuresPoster · Poster Session IIConfidence as a Compass: Regime-Dependent Dual Steering for Safety–Utility Tradeoffs in Language ModelsPoster · Poster Session IIDemonstrations, CoT, and Prompting: A Theoretical Analysis of In-Context LearningPoster · Poster Session IIEmergent Misalignment as a Distributional Phase Transition in Rank-1 LoRA Fine-TuningOral · Poster Session IIEMO: Pretraining Mixture of Experts for Emergent ModularityPoster · Poster Session IIEVALUATING THE ROBUSTNESS OF FACT-CHECKING LLMS ON MULTILINGUAL EVIDENCEPoster · Poster Session IIGOPO: Policy optimization using ranked rewardsPoster · Poster Session IIHow Language Models Choose Sides: Internal Representations of Instruction HierarchyPoster · Poster Session IIInterpreting Attention Sink Neurons in Large Language ModelsPoster · Poster Session IILanguage model suffer from a curse of ambiguityPoster · Poster Session IIMany Preferences, Few Policies: Towards Scalable Language Model PersonalizationPoster · Poster Session IIMMIB: A Mechanistic Interpretability Benchmark for Vision-Language ModelsPoster · Poster Session IIMultiView-Bench: A Diagnostic Benchmark for World-Centric Multi-View Integration in VLMsPoster · Poster Session IINatively Unlearnable Large Language ModelsPoster · Poster Session IIOn-Policy Self-Distillation with Sampled Demonstrations Reduces Output DiversityPoster · Poster Session IIPeer-Predictive Self-Training for Language Model ReasoningPoster · Poster Session IIPredicting Neural Scaling Laws without Training: A Data Manifold OraclePoster · Poster Session IIRepresentation Steering for Emotions in Diffusion Language ModelsPoster · Poster Session IIRethinking On-Policy Self-Distillation for Thinking ModelsPoster · Poster Session IISpend Less, Fit Better: Budget-Efficient Scaling Law Fitting via Active Experiment SelectionOral · Poster Session IIStochastic Collapse: Video World Models Fail to Reproduce Outcome FrequenciesPoster · Poster Session IITaming Polysemanticity in LLMs: Theory-Grounded Feature Recovery via Sparse AutoencodersPoster · Poster Session IITemporal Conflict Resolution in Video Large Language ModelsPoster · Poster Session IITruncation Is Not the Bottleneck: Reliable Calibration Conclusions from Truncated API Log-ProbabilitiesPoster · Poster Session IIWeight Decay Improves Language Model PlasticityPoster · Poster Session IIWhat AstroPT knows about galaxies, and what that can teach us about LLMsPoster · Poster Session IIWhat Black-Box Benchmarks Miss: Open-Ended Generation Under-Measures Memorization in Foundation ModelsPoster · Poster Session IIWhere did the ambiguity go? Examining how multimodal models interpret polysemous wordsOral · Poster Session IIWhere Does Social Reasoning Come From? Capability Provenance in Language ModelsPoster · Poster Session II