ReGen proposes hierarchical multi-prompt representation generation for efficient waveform diffusion models to address representation alignment issues in diffusion Transformers.
Empirical evaluation of vLLM inference engine configurations measuring trade-offs between energy consumption, performance, and output quality.
Analysis of task interference as mechanism for modeling forgetting in continual learning using path-averaged curvature.
Git-Assistant combines LLMs with automated planning to help developers with repository management tasks and version control workflows.
SAMPAT neural architecture using polynomials and analytic transformations for interpretable learning of continuous functions.
Online experiment examining user perceptions of benefits/risks and disclosure willingness when using AI chatbots for health topics.
QLoRA-tuned LLM system for telecom fraud detection with blockchain-based decision management and auditability.
Study testing whether four LLMs show geopolitical bias when evaluating international policies presented with different country endorsements.
Research on risk-aware Markov decision processes where agents optimize risk measures of objective value distributions based on state visitation frequency.
Studies emergence of creativity, honesty, and selective memory in small hyperbolic language models as companions.
ML approach to automatic thematic indexing of literary corpora, demonstrated on Voltaire's Complete Works.
Evaluates NLP approaches for automated keyword extraction from crowdsourced digital archives using the Their Finest Hour collection.
WILDTRACE benchmark evaluating LLM long-context reasoning requiring evidence integration across distant document passages.
Shortcut Trajectory Planning: consistency-based approach reducing inference cost in diffusion-based offline RL planners.
Identifies deceptive grounding failure in clinical RAG systems where correct citations mask incorrect entity attribution.
Diversify2Verify: LLM pipeline generating diverse program implementations to improve automated verifiability of Why3 code.
STEEL: sparsity-aware fused attention optimization for energy-efficient long-sequence LLM inference on AMD NPUs.
Self-guided test-time training method for improving long-context LLM reasoning and evidence utilization during inference.
Soofi S 30B-A3B: open-source Mixture-of-Experts hybrid Mamba-Transformer foundation model for German and English with efficient inference.
Examines test-time scaling techniques for small open-source vision-language models on multilingual visual QA benchmarks.
Parameter-efficient CLIP adaptation framework for animal re-identification with continuous metadata conditioning.
Pipeline combining reverse engineering, code retrieval, and LLM reasoning to recover source code from stripped binaries.
Explores using LLMs to generate multiple explanations for ML models, arguing explainability and prediction are complementary rather than trade-off objectives.
Probes VGGT geometric foundation model for emergent co-visibility encoding useful in 3D reconstruction and robotic localization.
Large-scale study analyzing failure trajectories of LLM coding agents in terminal environments as temporal process rather than outcome.
Energy profiling of on-device VLM inference reveals language generation, not visual processing, as primary energy bottleneck.
ALICE unified pathology foundation model using multi-stage distillation from eight vision and vision-language teacher models.
TCLA: Training-free class-wise logit adaptation for medical vision-language models handling domain shift and class bias.
PAC-ACT post-training method using actor-critic for action chunking transformers in precision industrial robot manipulation.
Task-specific multimodal QA agents with confidence calibration and incremental reasoning for QANTA 2026 challenge.
Semantic Pareto-DQN: Multi-objective reinforcement learning framework for financial anomaly detection addressing class imbalance.
VEXAIoT applies LLM agents to autonomous IoT vulnerability exploitation and penetration testing in constrained environments.
Introduces Complex Social Behavior dataset and evaluates vision-language model accuracy and error types over decade of progress.
Scalable visual pretraining approach for foundation models incorporating figures, equations, and layouts beyond text conversion.
IFAR framework using LLMs for multi-perspective causal discovery with DeepAbduction dataset for abductive reasoning evaluation.
Mechanistic analysis of white-box LLM monitor evasion strategies and proposed defenses through red-team experiments.
Study on how human beliefs about agent capabilities affect preference generation in RLHF, proposing theoretical framework beyond reward functions.
QAgent: Multi-agent LLM framework for autonomous OpenQASM quantum circuit code generation with domain-specific planning and iterative synthesis.
Multi-agent system with fine-tuned small LMs for telecom network troubleshooting. Combines multiple agents and SLMs for domain automation.
BREW framework distills agent interaction trajectories into retrievable knowledge base enabling agents to learn from experience across sessions.
LLM agents solve multi-constraint planning by programming candidate solutions rather than pure reasoning. Efficient constraint satisfaction via code generation.
Personalized adaptive curriculum engine for 9-1-1 call-taker training. Domain-specific education application with limited tech generalizability.
Self-evolving agentic framework for metasurface inverse design coupling coding agent with physics-based evaluator. Agent autonomously generates optimization code.
Framework for allocating human verification budget when using LLM predictions in surveys. Budget optimization with variable LLM accuracy.
DReST reward function trains agents to lack preferences over trajectory length, promoting shutdownable behavior. Safety-focused agent training method.
Hierarchical Preference Optimization improves DPO for complex reasoning by providing fine-grained feedback on solution subsections. Extends preference learning to multi-step reasoning.
Information-bottleneck method for learning coordination graph topology in multi-agent RL. Theoretical approach to agent communication capacity allocation.
Evaluates concept-based explainability of multimodal LLMs in few-shot in-context learning. Studies transparency of MLLM reasoning processes.
Latent Reward Steering adaptively promotes cognitive behaviors in reasoning LLMs at inference time. Addresses reasoning behavior control without explicit guidance.
SHARP learns long-range temporal patterns in streaming settings via hierarchical replay. Temporal sequence modeling for specialized domains.