X-FEMR proposes token-level explainability approach for Electronic Health Records foundation models using transformers for clinical prediction tasks.
Study investigates reward function design for reinforcement learning to improve LLM-generated BPMN process models beyond supervised fine-tuning limits.
UBEP optimizes communication library for Mixture-of-Experts models on high-bandwidth superpods, addressing serialization and latency bottlenecks.
Spider 2.0-AIFunc benchmark evaluates text-to-SQL models on AI-native SQL workflows with LLM functions for classification, filtering, and search tasks.
UI2App benchmark for evaluating LLMs on visual interaction inference in executable web application generation from screenshots.
Large-scale evaluation of 9 uncertainty estimation methods across 22 languages for LLM multiple-choice question answering.
Framework using LLM code agents for automatic software verification, enabling autonomous proof generation for interactive theorem provers like Coq.
RuBench 1.0 benchmark with 25 repository-level agentic coding tasks in Russian and other languages from live open-source projects.
Experimental design framework for systematically evaluating autonomous model discovery behavior of LLM coding agents with variability quantification.
TILDE method for concept unlearning in text-to-image diffusion models balancing unwanted concept removal with model quality retention.
Analysis of localization signals in vision-language models used as condition encoders for diffusion-based image editing.
Question-type-aware LLM pipeline framework for biomedical question answering in BioASQ 14b using agent collaboration and evidence grounding.
DataGovBench benchmark evaluating LLMs on real-world data analysis tasks including multi-tabular datasets and exploratory insight discovery.
RSF-GLLM framework for multi-hop knowledge graph QA using differentiable graph reasoning with Recurrent Soft-Flow module and decoupled LLM generation.
arXiv paper analyzing graph attention mechanisms for denoising from spectral perspective in graph diffusion models.
arXiv paper on ELSA3D unified 3D foundation model with elastic semantic anchoring for improved text-3D interaction and explicit structural representation.
arXiv paper proposing federated learning framework using multimodal LLMs to address data heterogeneity across distributed clients.
arXiv paper investigating socioeconomic indicators in satellite imagery and LLM-generated text for poverty mapping in African neighborhoods.
arXiv paper discovering what thinking models learn via Sparse Autoencoders on reasoning traces and constructive model diffing analysis.
arXiv paper introducing PROBE benchmark for measuring proactive problem-solving in LLM agents across longer time horizons and multiple information sources.
arXiv paper formalizing bystander disempowerment: how AI agents optimizing for one user unintentionally erode bystanders' agency in shared environments.
arXiv paper presenting VASP Agent, LLM-based coding agent for autonomous first-principles materials computation with workspace-state management.
arXiv paper on agentic AI system for insurance underwriting with adversarial self-critique mechanism ensuring reliability in regulated environments.
arXiv paper measuring faithfulness gap in LLM agents: whether agents act on stated reasoning vs conclusions in controlled Texas Poker simulator.
arXiv paper on EEG motor decoding using video-derived motor priors for cross-subject brain-computer interface generalization without calibration.
arXiv paper proving active inference can be rewritten as VFE minimization with entropy-correction terms, unifying goal-directed and information-seeking behavior.
Three-layer framework for AI in scientific discovery: search/retrieval, model formation/evolution, and execution via optimization/simulation.
TAC benchmark: first agentic benchmark assessing animal welfare in AI agent behavior, extending beyond question-answer evaluations.
Framework using reward functions as agents to enable broader exploration in embodied world models beyond conservative training distributions.
FADE: mechanism for reducing language-prior dominance in vision-language models to mitigate hallucinations through information flow analysis.
HARC: approach coupling harmfulness and refusal directions in residual streams to improve LLM safety alignment robustness.
Nemotron-Labs-3-Puzzle-75B-A9B: compressed hybrid MoE LLM variant optimized for interactive deployment with 2x throughput improvement.
Agent Step Value: replay framework for evaluating which agent transitions helped or harmed task traces using LLM evaluators.
LLM-as-a-Verifier: framework for verification as a scaling axis, providing fine-grained feedback for agentic tasks without ground truth.
Survey of memorization and replication phenomena in visual diffusion models, covering privacy, security, and copyright concerns.
Metamemory Agent approach for data-free code generation in LLMs without requiring manually curated reference examples.
Analysis of o3 reasoning token usage versus accuracy gains in LLMs, showing improved efficiency rather than longer reasoning chains.
Vision paper exploring LLMs solving computational problems via prompting rather than programming, examining challenges and opportunities.
Distance Explainer: novel post-hoc method for interpreting embeddings by explaining distances between data points using saliency-based attribution.
EU AI Act's research exemptions and their potential impact on publication norms at major AI conferences.
Framework learning minimum action distance from state trajectories for environment structure representation in MDPs.
Detoxify: LLM framework for transforming abusive text to non-abusive while preserving meaning.
Research on sparse autoencoders for LLM interpretability showing L0 hyperparameter effects on feature quality.
SmartMixed: two-phase training strategy enabling per-neuron adaptive activation functions in neural networks.
SWITCH: Benchmark for embodied agents handling tangible control interfaces in long-horizon scenarios.
SecureCode: Production-grade dataset of 2,185 multi-turn examples for training security-aware code generation models.
OpenGround: Planning-based perception for zero-shot 3D visual grounding in open-world scenarios without predefined object lookups.
KernelEvolve: Agentic framework for automated kernel coding targeting heterogeneous AI accelerators at scale for DLRM.
Multi-task instruction tuning for Arabic-English audio LLM covering ASR, speech/text summarization via data scheduling.
StepShield: First agent safety benchmark measuring detection timeliness for rogue agents via Early Intervention Rate metric.