21
From Prompts to Contracts: Harness Engineering for Auditable Enterprise LLM Agents
The paper introduces a harness-engineering architecture designed to make enterprise large language model agents more traceable, deterministic, and auditable.
Top stories in Research
21
ContextSniper: AntTrail's Token-Efficient Code Memory for Repository-Level Program Repair
The paper introduces ContextSniper, a token-efficient code memory module designed to select precise evidence for repository-level program repair using language model agents.
21
Collective Intelligence with Foundation Models
This paper outlines a multi-agent framework that coordinates solver, critic, and aggregator foundation models to build more reliable cooperative reasoning systems.
21
Beyond Attention Scores: SVD-Based Vision Token Pruning for Efficient Vision-Language Models
Researchers proposed an SVD-based vision token pruning method to reduce the computational and memory overhead of processing long sequences in Vision-Language Models.
21
Concept-as-Tree: A Controllable Synthetic Data Framework Makes Stronger Personalized VLMs
Researchers have introduced Concept-as-Tree, a synthetic data framework designed to improve the personalization capabilities of vision-language models by generating high-quality positive and negative samples.
21
LSRM: High-Fidelity Object-Centric Reconstruction via Scaled Context Windows
The Large Sparse Reconstruction Model demonstrates how expanding transformer context windows can improve the recovery of fine-grained textures in feed-forward 3D reconstruction.
21
Hierarchical Control in Multi-Agent Games: LLM-based Planning and RL Execution
Researchers propose a hierarchical framework that utilizes a centralized large language model for strategic planning to coordinate specialized reinforcement learning policies in multi-agent environments.
21
Out of Sight: Compression-Aware Content Protection against Agentic Crawlers
Researchers have developed a content protection method that exploits context compression to prevent LLM-based web crawlers from scraping online information.
21
CriterAlign: Criterion-Centric Rationale Alignment for Code Preference Judging
CriterAlign is a criterion-centric evaluation framework designed to improve pairwise code preference judging by aligning rationales across specific criteria.
21
VEGAS: Human-Aligned Video Caption Evaluation via Gaze
VEGAS is a training-free evaluation metric that uses viewer gaze data to measure how well generated video captions align with human visual attention.
21
Understanding Axes of Difficulty For Long Context Tasks Via PredicateLongBench
A new benchmark called PredicateLongBench has been introduced to systematically analyze the specific factors that make long-context tasks difficult for large language models.
21
TypeProbe: Recovering Type Representations from Hidden States of Pre-trained Code Models
This study investigates how pre-trained code models internally represent programming types by probing their hidden states across different programming languages.
21
Who Analyses the Analyser? Self-Validating LLM Hazard Analysis with Constitutional Meta-STPA
This study introduces a self-validating framework to analyze and mitigate safety hazards and hallucinations in LLM-assisted safety engineering tools.
21
CausalDS: Benchmarking Causal Reasoning in Data-Science Agents
CausalDS is a new benchmark designed to evaluate the causal reasoning and data analysis capabilities of language model agents.
21
Search Beyond What Can Be Taught: Evolving the Knowledge Boundary in Agentic Visual Generation
The paper introduces SearchGen-20K and SearchGen-Bench to address the world-knowledge limitations of visual generators by integrating search capabilities.
21
Narration-of-Thought: Inference-Time Scaffolding for Defeasible Ethical Reasoning in Large Language Models
The paper proposes Narration-of-Thought, a system prompting technique that structures chain-of-thought reasoning to improve ethical decision-making and reduce bias in large language models.
21
Persona Cartography: Charting Language Model Personality Traits in Weight Space
A new framework maps large language model personality traits within weight space using low-rank adapters to measure and control model personas.
21
Multi-agent Autoformalization of Tensor Network Theory
A new multi-agent workflow utilizing large language models has been developed to autonomously formalize complex theorems in theoretical physics.
21
ProjAgent: Procedural Similarity Retrieval for Repository-Level Code Generation
Researchers have introduced ProjAgent, a repository-level code generation system that retrieves functions based on procedural logic rather than simple lexical or semantic similarity.
21
LiteOdyssey: A Lightweight Reasoning AI Agent for Interpretable Rare-Disease Diagnosis
Researchers developed liteOdyssey, a lightweight framework that guides general-purpose language models through diagnostic reasoning for rare diseases without altering model weights.
21
Remember When It Matters: Proactive Memory Agent for Long-Horizon Agents
Researchers have introduced a proactive memory agent designed to mitigate behavioral state decay in long-horizon tasks by actively retrieving relevant context.
21
When RLHF Fails: A Mechanistic Taxonomy of Reward Hacking, Collapse, and Evaluator Gaming
The study presents a mechanistic taxonomy and diagnostic pipeline to track how failures like reward hacking and policy collapse emerge during reinforcement learning from human feedback.
21
Curriculum Learning for Efficient Chain-of-Thought Distillation via Structure-Aware Masking and GRPO
Researchers have introduced a three-stage curriculum learning framework designed to efficiently distill chain-of-thought reasoning from large language models into smaller student models.
20
Best-of-$N$ TTS Evaluation is Confounded by ASR Family Alignment
This paper demonstrates that evaluating zero-shot text-to-speech systems using automatic speech recognition verifiers is biased by the specific model family used for judgment.