Procedural Memory Distillation method for self-improving language models that retains cross-episode procedural information from rollouts beyond episode-local rewards.
Study showing AI agents capture analytical variation among human researchers; different personas enable agents to report divergent conclusions from identical datasets.
Janus is a playground system for designing and evaluating user-involved permission management in autonomous AI agents that execute tool calls.
Semi-CoT framework reuses LLM-generated reasoning traces as semi-supervised learning signals to improve chain-of-thought reasoning with limited labeled data.
OPINE-World enables data-efficient agent adaptation by synthesizing programmatic world models through LLM-generated source code refined via counterexample-guided inductive synthesis.
Scaling analysis of SOLiD, a lie detector-based oversight method for identifying deceptive LLM responses, showing improved scaling with larger models in preference learning.
EO-Agents uses a three-agent LLM pipeline grounded in NASA Earth Observation Knowledge Graph with graph neural networks to generate scientific hypotheses from structured data.
Hawk uses LLMs with hardware-aware knowledge and semantic understanding to generate high-performance kernels for Neural Processing Units, addressing implicit constraints.
Neural-symbolic world model approach for cloud system fault detection and recovery using LLMs for semantic understanding and DRL for policy optimization.
SemHash-LLM framework combines semantic hashing, MinHash, contrastive learning, and LLM-based filtering for efficient large-scale document deduplication while preserving semantic equivalence.
Research on calibrating LLM confidence via RL to accurately express uncertainty during test-time scaling for reasoning and QA tasks.
Graph fusion method for rainfall reconstruction reconciling point gauges, path microwave links, and gridded radar/satellite with geometry-aware approach.
Autonomous AI system discovers universal traffic laws from observational data to identify recurrent congestion patterns across cities.
Multi-agent LLM deliberation improves forecasting when agents receive diverse evidence; identical evidence causes herding rather than genuine belief revision.
Research on continual ECG model deployment separating expert retention from source inference using frozen backbone and isolated classifiers per source.
Goggles is a learned module using gradient editing to fix Negation Neglect, where finetuned LLMs fail to recognize fictional content (9% accuracy baseline).
COMFYCLAW introduces self-evolving skill harnesses for agentic workflow-based image generation, enabling agents to recall patterns and preferences.
Generic TB-Coverage proposes coverage-aware expert pruning for sparsely-activated Mixture-of-Experts language models without downstream calibration data.
Distributionally robust listwise preference optimization for LLM alignment under ranking-label uncertainty from annotator inconsistency and noise.
Meta-benchmarking framework organizing 452 public benchmarks into 41 O*NET work activities for evaluating financial-services LLM performance.
Phi-Nav addresses semantic drift in vision-language navigation by generating hindsight instructions for on-policy exploration training.
Mastermind trains agents to reproduce vulnerabilities at repository scale by learning strategies for codebase inspection, input grammar inference, and PoC construction.
SimWorlds enables LLM agents to generate dynamic 4D scenes with physics simulation from text, supporting liquids, particles, and articulated motion.
Addresses world-model correction in long-horizon agent planning by fixing underlying issues rather than replanning entire graphs after mistakes.
Proposes retrieval-augmented SLM framework using formal concept analysis as symbolic verification loop for ontology and knowledge structure expansion.
Studies latent time representations in Diffusion Language Models, showing they internally encode denoising progress information without explicit timestep conditioning.
Vera is an automated safety testing framework for LLM agents performing autonomous actions, scaling risk discovery beyond expert-designed violations.
Pre-Flight is an open-source benchmark of 300 multiple-choice questions evaluating LLMs on aviation-specific operational knowledge and safety reasoning.
CLAP presents closed-loop methodology for domain agent post-training, converting business data into structured samples with validation and release-gate controls.
STEER identifies safety gaps in LLM multilingual and code-switching inputs, revealing models generate harmful responses outside English safety training distribution.
SkillCoach proposes self-evolving rubrics for evaluating and training LLM agents to reliably use skills from repositories with overlapping capabilities.
Spec-AUF training method for masked block drafters in speculative decoding addressing train-inference misalignment.
HECATE tool for measuring complexity in LLM-integrated applications across both prompt and code layers.
ContextSniper: token-efficient code memory layer for LLM agents performing repository-level program repair with reduced context overhead.
ElephantAgent: framework for maintaining verifiable contextual state continuity in agentic systems under tool and memory poisoning attacks.
A-TMA: method to address ghost memory failures in LLM agent long-term memory systems tracking state changes and fact transitions.
Atomic Task Graph framework for LLM-based agent planning and execution handling multi-step tasks without scaling or fine-tuning.
OntoLearner: modular Python library for automated ontology learning from text using LLMs with systematic evaluation framework.
Methods for continuous knowledge editing in multimodal LLMs with controlled scope to minimize disruption to unrelated behaviors.
Research on memory consolidation in long-running AI agents while preserving identity and audit compliance without changing agent behavior.
Traceable fault diagnosis system for battery storage using retrieval-augmented multi-agent assistant to combine alarms, measurements, topology and maintenance data.
InduceKV enables fixed-footprint continual adaptation of MLLMs by storing task-specific updates in induced KV memories without modifying backbone model.
Studies hidden forgetting in continual multimodal learning where standard metrics miss degradation of visual and textual grounding despite maintained answer accuracy.
PACE shows that cheap non-agentic LLM benchmark performance can accurately predict expensive agentic benchmark results, enabling cost-effective agent capability evaluation.
ADTC framework for exhaustive analysis of optimal decision trees using algebraic model counting for global explainability and reliability assessment.
Maven is an RL framework with editable evidence memory for long-context reasoning, defining evidence-state value and reward functions for intermediate reasoning steps.
SUNTA uses surprise-based chunking in hierarchical state-space models for long-horizon video prediction, aligning chunk boundaries with intrinsic temporal structure.
ContextNest formalizes context governance for autonomous AI agents, providing open spec for knowledge vaults with provenance, versioning, integrity, and traceability.
Domain-specific post-training of LLMs for Scientific Fitness Coaching, addressing knowledge gaps in general-purpose models for specialized fitness scenarios.
Paper-replication workflow enables coding agents to replicate computational claims from scientific ML papers and verify if generated evidence supports claims.