Benchmark separating reasoning from knowledge retrieval in LLM evaluation using isomorphic cross-domain problem pairs.
Study of pruned Mixture-of-Experts models in biomedical domain, evaluating utility and factual reliability under resource constraints.
Research on gradient geometry of embedding tables in language models; introduces Ember optimizer for efficient finetuning and pretraining with minimal memory.
Framework for reducing LLM hallucinations in resume optimization through temporal validation, contamination detection, and structural checks.
Framework decomposing advantage function variants in RL-based LLM post-training to address training instability and diversity collapse in policy gradient methods.
Multi-Head Recurrent Memory Agents: Addresses reliability degradation in LLM agents with long contexts by decomposing memory capture and retention.
IntentTune: System for resolving ambiguous e-commerce queries by inferring latent user intent attributes using demand and personalization signals.
EFE: Framework using LLM-based evolutionary search to discover preprocessing transformations for structured data as composable Python programs.
X-LogSMask: Explainable multi-scale transformer variant for sparse, structured graph data with improved interpretability over standard transformers.
DiPS: Q-learning framework for dialogue policy selection in high-stakes persuasion agents, dynamically adapting strategies based on personality.
ADVENT: LLM-driven predicate invention mechanism for Inductive Logic Programming combining LLM abduction with Prolog deductive verification.
VLAFlow: Unified flow-matching framework for training vision-language-action models enabling controlled comparison of VLA objectives with robot manipulation datasets.
Comprehensive benchmark for LLM-based data agents automating data science workflows. Evaluates agent capabilities on heterogeneous datasets.
Probabilistic framework for merging task-specific models into multitask solutions. Scores statistical utility of parameter updates across tasks.
Diagnostic framework for predicting closed-loop performance of world models in model-based RL without explicit validation metrics.
Bayesian reinforcement learning method addressing data scarcity through prior knowledge and belief updates in sequential decision-making.
Open-vocabulary object detection calibration using frozen VLMs. Vision-language model application with limited novelty.
Adaptive expert pruning and growing for efficient MoE fine-tuning using LoRA. Parameter-efficient training for large models.
Expander sparse autoencoders for mechanistic interpretability with reduced parameters. ML research on interpretability and efficient dictionaries.
Causal study of AI coding agent adoption on open-source projects, analyzing impact on newcomer participation. Developer tools and OSS ecosystem.
Continuously evolving multimodal benchmark using multi-agent pipeline for VLM evaluation. Developer tools and evaluation frameworks.
Memory-efficient training stack combining parallelism techniques for MoE models. ML research on scalable training infrastructure.
Evaluation of chunking strategies for RAG systems on academic texts using RAGAS framework. Directly relevant to LLM applications and retrieval techniques.
Empirical study of LLM-generated code and comments in real repositories, analyzing prevalence and quality concerns. Developer tools and LLM applications.
Binarization technique for vision-language models reducing memory/latency for deployment. Model optimization for efficient inference.
Reward-free reinforcement learning from video using VLM as progress scorer and GRPO objective. AI agents and novel training approach.
Brain disease diagnosis framework integrating LLM semantics with brain connectivity analysis via hypergraphs. LLM application with healthcare focus.
Pipeline for adapting Qwen 27B model to perform reasoning in Turkish rather than English, addressing multilingual LLM reasoning.
Dataset of 1,639 K-12 science explanations with human and LLM-generated alternatives for training risk assessment auditors.
CausalSTeward agentic divide-conquer-combine system for causal discovery integrating prior knowledge to identify causal models from high-dimensional data.
PhysMani framework combines physics-principled 3D Gaussian world model with action policy for dynamic object manipulation in embodied AI.
Conditional co-ablation technique reveals transformer self-repair mechanisms where dormant backups activate after primary component ablation.
Analyzes representational geometry showing LLMs become robust to science skepticism through problematic mechanisms rather than genuine understanding.
Object Aligner provides configurable JSON schema similarity scoring for measuring LLM output structure alignment in tool calling and agentic systems.
Evaluates Vision-Language Model reliability for medical image quality assessment under corruption and bias conditions.
MolSight vision-language model combines molecular LLMs with graph-aware visual understanding for molecular structure and drug discovery tasks.
Controlled study comparing nine lightweight CNN architectures across multiple datasets and hardware to assess efficiency claims.
Proposes load-aware prefill deflection technique to improve disaggregated LLM serving efficiency by balancing prefill and decode GPU pools.
OpenSafeIntent benchmark evaluates whether LLMs calibrate assistance appropriately across benign, dual-use, and malicious intent variants.
SPLIT benchmark evaluates LLM cross-lingual empathy and cultural grounding in emotional-support contexts across English and Ukrainian.
Demonstrates performance evaluation failures in spatiotemporally correlated domains due to data leakage from non-i.i.d. splits.
Proposes prompt coverage adequacy testing framework to guide LLM and autonomous agent testing when prompts replace traditional code.
kNNGuard presents training-free guardrail for LLMs using activation space of off-the-shelf models to detect unsafe/adversarial prompts with minimal labeled data.
Introduces emotional self-correction mechanism for vision-language models to improve reasoning reliability without post-training or engineered feedback.
Proposes test-time guidance framework for vision-language-action policies using learned critic to guide flow-matching inference without retraining base models.
Presents vLLM-based inference pipeline for unified audio understanding and generation in speech language models with multi-token prediction support.
Develops behavioral monitoring techniques to detect and analyze guardrail activations in LLMs, enabling black-box security testing of production AI systems.
Reviews AI risk assessment and management methodologies under EU AI Act and other regulatory frameworks, covering identification, analysis, and mitigation approaches.
Systematically categorizes 53 human-AI team studies into five clusters using psychological teaming taxonomies to understand collaboration patterns and diversity.
Presents open infrastructure for operationalizing AI audits by moving beyond taxonomies to actionable tests. Reviews 74 existing risk taxonomies and proposes executable evaluation framework.