Proposes SketchVerify for structured inference-time scaling where LLMs enumerate algorithmic strategies and write program sketches verified by execution for cost-effective improvement.
Empirically analyzes fairness in LLM explanations across demographic groups, introducing Explanation Fairness Taxonomy framework for evaluating explanation quality disparities.
Survey of attention-based graph neural networks covering mechanisms for adaptive feature selection and noise filtering in graph representation learning.
Compares voting strategies for LLM code generation including Semantic Voting, analyzing how textual voting, ranking, and execution-based agreement contribute to code selection.
Introduces PrepBench benchmark evaluating natural language-driven data preparation using LLMs, assessing progress toward replacing GUI-based data transformation tools.
Introduces Gate-and-Merge, a zero-shot framework for compositional personalization of vision-language models using lightweight LoRA adapters paired with concept tokens.
Develops end-to-end autonomous parking using reinforcement learning with Gaussian Splatting simulator for real-to-sim-to-real transfer in extreme parking scenarios.
Proposes AgentForesight for early failure prediction in LLM-based multi-agent systems through online auditing, enabling intervention before trajectory-level failures cascade.
Introduces PROBE, a framework for structured recovery in software engineering agents that converts runtime failures into grounded recovery guidance for subsequent attempts.
Proposes AdaPreLoRA, an optimization method for Low-Rank Adaptation using Adafactor preconditioning to handle rank-deficiency issues in LoRA factor-space preconditioners.
Systematically studies compression techniques (pruning and knowledge distillation) for large-scale MoE model pretraining, examining initialization quality and expert compression strategies.
Research on measuring causal dimensionality in transformer representations using sparse autoencoders and attribution patching to characterize feature importance.
Framework quantifying LLM dispositions as stable model-specific regularities measuring consistency and diversity across narrative tasks.
Lexical acoustic coding framework where LLM agents transmit audio through natural language text and generated synthesis code.
Study on dishonesty in LLM unlearning methods, showing hallucination and inconsistency issues, with evaluation framework and improvements.
GPU-accelerated solver for entropic-regularized optimal transport using sparse-plus-low-rank decomposition.
PPU-Bench benchmark for personalized partial unlearning in multimodal LLMs with fine-grained factual control and real-world deletion requests.
SimReg regularization technique reducing intra-class variance in token embeddings to improve LLM pretraining efficiency and representation learning.
Vision-language-action model for autonomous driving using expert routing to balance semantic understanding with trajectory prediction.
Cognitive science study comparing how LLMs versus humans ground abstract concepts, replicating property-generation experiments across 21 models.
Framework for creating leakage-free RAG evaluation benchmarks by identifying and filtering questions answerable from LLM parametric memory.
BubbleSpec addresses RL training efficiency bottlenecks for LLM improvement by converting synchronization delays into speculative rollout drafts.
Mechanistic study identifying structural vulnerabilities and refusal-escape directions explaining why aligned LLMs remain vulnerable to jailbreak attacks.
Research on smoothness degradation in extremely quantized LLMs beyond numerical precision loss, proposing methods to improve deployment efficiency.
FragileFlow addresses margin-aware error flow in LLMs and VLMs where predictions remain correct but confidence shifts toward wrong classes near decision boundaries.
LLM-Agnostic semantic representation attack circumvents alignment safeguards by optimizing adversarial prompts at semantic level rather than token level.
DAPE improves efficient vision-language models by dynamically aligning non-uniform information density between text and image modalities.
Improves lexical difficulty prediction with contrastive learning and ridge ensembling for cross-lingual language learning and readability assessment.
MolWorld introduces world models for actionable molecular optimization in drug discovery ensuring reachability through valid local structural transformations.
Proposes spatio-temporal object tracking monitoring for video LLMs to reduce hallucinations in dynamic scenes by persistent identity and relation tracking.
GuardVLA introduces backdoor-based ownership verification framework for Vision-Language-Action models to protect model ownership in robotic applications.
Studies in-context learning of LLMs for sequential decision-making tasks including MDPs and POMDPs using supervised fine-tuning for few-shot reasoning.
Investigates predictive information geometry across LLM layers using representation lenses as diagnostic tools to track how next-token prediction capabilities emerge.
Evolutionary Ensemble organizes coding agents into decentralized co-evolving system for algorithmic discovery by evolving cumulative guidance and learned skills.
ShadowMerge demonstrates poisoning attacks on graph-based agent memory systems by injecting crafted relations that influence LLM agent behavior and reasoning.
Octopus Protocol enables AI agents to control new hardware with single command by using language models to generate drivers without pre-existing SDKs or primitives.
Position paper on evaluating jailbreak attacks comprehensively across parameter distributions rather than single configurations for fair LLM security benchmarking.
Research on defending multi-agent LLM systems against Byzantine faults in peer-to-peer networks to prevent adversarial manipulation and system degradation.
FactoryNet introduces a 51M datapoint dataset for industrial time-series foundation models with a unified S-E-F-C schema for zero-shot cross-embodiment transfer and anomaly detection.
Theoretical characterization of personalized LLM alignment, establishing conditions for optimal online regret and sample complexity under user diversity.
Framework grounding LLM agent reliability techniques (retry, voting, self-consistency) in Shannon coding theory for cost-aware adaptive reliability.
First controlled comparison of internal deliberation versus external evolution for discovering behavioral rules in multi-agent AI systems.
Empirical comparison of LLMs versus traditional taggers for part-of-speech tagging in Medieval Romance languages with limited annotated resources.
Hierarchical architecture for traffic simulation combining multi-agent interaction reasoning with continuous trajectory prediction beyond self-play RL.
Survey and analysis of optimization algorithms for LLM training, comparing Adam variants and memory-efficient optimizers for large-scale pretraining and fine-tuning.
DARE improves RL training of LLMs through difficulty-adaptive data selection with co-evolved estimation, addressing sample inefficiency and policy drift issues.
RigidFormer uses transformers to learn rigid-body dynamics from mesh-free representations like point clouds, improving computational efficiency over vertex-level methods.
Flame3D uses agentic language models for zero-shot 3D scene understanding, combining object grounding, spatial reasoning, and integration with external tools.
Krone uses hierarchical log abstraction and LLM augmentation to detect and explain system anomalies by transforming flat log sequences into semantically coherent units.
Large-scale compositional jailbreak dataset with 114K adversarial prompts and systematic framework for generating and evaluating LLM attacks.