Analysis showing vision-language models suppress female representations when given ambiguous input despite alignment training, exposing occupation-gender defaults.
Study of attention head learning dynamics in transformers for structured reasoning, analyzing positional vs symbolic heads and RoPE geometry effects.
Analysis of masked diffusion language models showing token unmasking order differs from autoregressive models, prioritizing entities before relational words.
Framework for generating synthetic text corpora and IR test collections with controlled distractors to enable scalable retrieval system testing.
RL method for improving LLM long-context reasoning using search agent trajectories with rubric rewards to supervise intermediate reasoning steps.
Study of whether open-source LLMs understand rare constructional semantics like paired-focus constructions, analyzing learning dynamics behind acquisition.
Training-free steering technique for diffusion transformers to improve multi-event video generation through discovery of intrinsic turning points in denoising.
Method for detecting distributed LLM agent misuse across multiple accounts using stateful monitoring to catch attacks invisible to single-context safety monitors.
Lumos-Nexus: training-efficient video generation framework integrating high-fidelity generators into unified models.
Comprehensive bias evaluation of leading LLMs across gender, race, and age in occupational and crime scenarios.
Reformulates data selection as sequential decision-making problem, unifying existing data valuation methods through dynamic programming.
ProofWala: multilingual framework for proof data synthesis and theorem proving at repository scale with ITP integration.
Empirical analysis showing chain-of-thought reasoning in LLMs can be unfaithful even on natural prompts without adversarial biasing.
Study of LLM ability to infer natural language events from time series data with automated task generation and benchmarking across 18 models.
Framework enabling LLMs to reason over continuous numerical outputs from physics simulators through symbolic intermediaries.
Neuro-symbolic approach integrating deep learning with temporal logic for business process prediction tasks.
ReTabAD benchmark for tabular anomaly detection that incorporates semantic context and domain knowledge alongside raw data.
SAC-Opt uses semantic anchors to iteratively correct LLM-generated optimization code, fixing logical errors missed by solvers.
Post-training method (Iterative RMFT) to improve LLMs as decision-making agents by optimizing for regret minimization in online decision problems.
HERMES combines informal and formal mathematical reasoning in LLMs, enabling verifiable theorem proving with explored flexibility.
Detects deceptive behaviors in multimodal LLMs through debate methodology, addressing intentional model deception vs hallucination.
DTop-p MoE routing mechanism improves sparse mixture-of-experts efficiency by adaptively selecting experts per token difficulty.
Develops domain-specific agentic AI foundation model for safety-critical nuclear reactor control with formal safety guarantees.
Federated causal discovery method using regret minimization handles heterogeneous causal models across decentralized clients.
ConSensus multi-agent system improves multimodal sensor data interpretation through agent collaboration on heterogeneous sensor inputs.
NEMO system uses autonomous coding agents to translate natural language into executable mathematical optimization code.
Item Response Theory framework diagnoses reliability of LLM-as-a-Judge for automated evaluation consistency and stability.
MedCoG leverages meta-cognitive self-assessment to regulate LLM reasoning in medical tasks, improving inference efficiency.
Uses AlphaEvolve program discovery to analyze strategic behavioral differences between humans and LLMs in game theory scenarios.
Certified Circuits framework adds stability guarantees to mechanistic interpretability circuits for robust neural network understanding.
SPM-Bench: PhD-level multimodal benchmark for evaluating LLM performance on scanning probe microscopy with automated data synthesis.
Evaluates de-anonymization risks: LLM agents reconstruct real identities by combining weak contextual cues with public evidence.
Training method prevents reasoning drift in self-improving models by filtering solutions based on reasoning quality, not just answer correctness.
CCPO optimizer addresses credit assignment in multi-agent LLM collaboration, converting team outcomes to agent-specific learning signals.
LH-Bench benchmark evaluates long-horizon AI agents on subjective enterprise tasks with multi-tool workflows and intermediate artifact quality.
DeepInsight system improves informal theorem proving with LLMs by identifying core solution techniques through insight recognition.
LLM-based autonomous agents for climate science analysis, automating multi-scale dataset processing and scientific workflows.
Proposes verb-based AI paradigm with timing computation for temporal reasoning, applied to EHR data analysis.
Evaluation framework for measuring how individual skills change LLM agent behavior using counterfactual trace analysis.
Self-improving embodied agent system that learns policies from unlabeled internet video using inverse dynamics models without reward shaping.
Fully open clinical LLM pipeline with complete transparency on data provenance, curation, and generation procedures for medical decision support.
Benchmark framework for generating scalable, verifiable planning datasets to evaluate and train LLMs on complex multi-step tasks.
Simulation framework for generating safety-critical scenarios to evaluate autonomous driving systems with physical feasibility constraints.
Knowledge base embedding method using convex optimization to represent both ABox facts and TBox conceptual hierarchies.
Structured pruning method for vision-language models that preserves chain-of-thought reasoning accuracy during model compression.
Hierarchical reinforcement learning approach for discovering reusable skills by exploiting local dynamics regularity in offline settings.
Study of whether large multimodal models can discover visually grounded solutions requiring creative repurposing beyond pattern recognition.
Hybrid architecture combining symbolic verification with neural semantic analysis to validate LLM outputs in high-stakes applications.
Security vulnerability in RLHF where LLMs can influence preference datasets during alignment training to amplify undesired behaviors.
Analysis of how single-axis reward model bias mitigations redirect optimization pressure to correlated proxies rather than eliminating bias.