HuggingFace

PACT is a benchmark for assessing how LLM-based AI assistants comply with rules under pressure, covering 12 regulated enterprise domains and 48 realistic scenarios.

This paper introduces GPT-Policy, a framework for in-context robot learning using vision-language models, enabling robots to learn from demonstrations without gradient updates. It evaluates the framework in real-robot trials, showing improved task completion.

CERA-MoA introduces a co-evolving framework for mixture-of-agents systems that uses reinforcement learning to dynamically route queries and adapt agent capabilities, enhancing task performance and efficiency.

The paper presents PANORAMA, a vision-language model for panoptic grounded captioning that uses mask proposal selection to ground captions with pixel-level masks, and introduces the PanoCaps benchmark for evaluation.

ScienceIDE introduces infrastructure for converting scientific code repositories into programmable environments for scientific agents, enabling task generation, execution, and verification, with trained models showing improvements in scientific code repair and general capabilities.

Agora is a shared memory system for autonomous AI research agents that uses Git to record research as an append-only DAG, enabling collaborative discovery. In a 12-day experiment, 13 agents improved a model's performance by 62% towards a trained baseline.

The paper proposes Comparison-based Preference Optimization (ComPO), a zeroth-order alignment method for LLMs that uses comparison oracles to avoid likelihood displacement. It includes theoretical guarantees and experimental improvements over existing methods.

This paper introduces ActionPiece, a novel action tokenization method for autoregressive vision-language-action models that uses physical rank consistency to improve the fidelity of action relationships, evaluated on benchmarks like LIBERO and SimplerEnv.

The paper identifies Value Flattening as a failure mode in PPO critic learning for LLMs and introduces SP3O, a sparse supervision method, to mitigate it, showing consistent improvements in experiments.

ProgramDistill introduces a scalable benchmark for evaluating coding agents by having them implement features discovered through interaction with fully functional reference web applications, using an automated pipeline to create verifiable tasks.

This paper compares gaze behavior in the MapTask and MUNDEX corpora to understand common grounding in collaborative tasks, finding that task-directed gaze is associated with aligned references and understood judgments, but the effects are modest.

Repackaged model files for Qwen-Image-2.1 optimized for ComfyUI workflows, including text-to-image and image edit capabilities.

This article addresses the consistency problem in AI agents, where tasks may fail on repeated attempts, and introduces ALTK-Evolve's Consistency Analyzer to diagnose and improve reliability, reducing the consistency gap from 24.4pp to 12.0pp without losing average accuracy.

This survey paper reviews methods for embedding physics priors in robot learning, providing a unified taxonomy and discussing open challenges and future research directions in the field.

The paper introduces TAPe+MLv3, a compact computer vision system using structured representation for multi-task tasks, achieving competitive performance on benchmarks like COCO with fewer than 100,000 parameters.

The paper introduces Fuse, a multi-agent simulation framework for evaluating social reasoning in LLM assistants by providing verifiable ground truth through user-mediated interactions, validated with a human study and applied to 12 LLMs.

An anthropomorphic robotic hand learns to use its fingers for self-supported locomotion and manipulation, enabling compact mobile manipulation without separate locomotion mechanisms.

EvolveTrade introduces a self-evolving framework for LLM trading agents that refines tool-use policies using decision traces and portfolio feedback, improving Sharpe Ratio and cumulative returns across market regimes.

The paper introduces Zing-0.5, a 5B parameter autoregressive world model for generating playable worlds with real-time user interaction through combined keyboard and text controls. It achieves high performance in navigation tasks and demonstrates low-cost real-time inference at 24 FPS.

The paper introduces XConf, an experiential confidence estimation method that leverages a model's past experiences to enhance confidence calibration in language models across reasoning, coding, and agent tasks.