Research Papers
EnvHarness: A Programmable Layer to Reshape Static Environments for Smarter AI Agents
Hugging Face researchers introduce EnvHarness, a plug-in layer that dynamically reshapes static environments to target agent weakn...
Bounded Agents: New Authorization Model Cuts Multi-Agent Data Theft to Zero
Researchers propose the Agentic Principal Chain (APC), an authorization architecture that tracks delegated authority across multi-...
Hugging Face Researchers Unveil MuseCPEval: A New Framework to Measure Music Context Preservation in Editing Systems
A new evaluation framework called MuseCPEval introduces 12 metrics across five musical facets to assess how well music editing sys...
HOTFIXR: New Framework Boosts Multilingual Reasoning in LLMs
Researchers introduce HOTFIXR, a data generation framework that targets multilingual reasoning weaknesses, improving cross-lingual...
SkillGate: New Training Method Boosts Long-Horizon Agent Success by Fixing 'Selector Credit Starvation'
Hugging Face researchers introduce SkillGate, a training method that separates outcome credit for execution tokens from local adva...
VA-Judger: A Human-Aligned Reward Model for Joint Video-Audio Generation
Researchers introduce VA-Judger, the first reward model designed for joint video-audio generation, along with a large preference d...
SoftVTBench: New Visuo-Tactile Benchmark Tracks Physical Interaction Quality in Deformable-Object Manipulation
Hugging Face researchers introduce SoftVTBench, a synchronized visuo-tactile dataset and deformation-aware benchmark that evaluate...
Hugging Face Researchers Unveil Attribute-Guided Framework to Scale Creative Writing Data Across 13 Genres
A new framework separates thematic seeds from genre-form controls to generate diverse, high-quality creative writing data, boostin...
AdaPop: Adaptive Popularity Boosts LLM Unlearning, Cutting Leakage by 5x
Hugging Face researchers propose AdaPop, an unlearning method that adapts gradient pressure based on fact popularity, reducing lea...
Looped Language Models Boost Compositional Tool Calling, New Study Finds
A new paper from Hugging Face researchers shows that looped language models improve multi-step, compositional tool use through rec...
FM-Bench: New Benchmark Shows Managerial Behavior, Not Scale, Drives Long-Horizon LLM Agents
A new benchmark from Hugging Face, FM-Bench, tests LLM agents managing a football club over 20 simulated years, revealing that man...
New Diagnostics and Action-Conditioned Objectives Improve Latent World Model Planning
Researchers propose diagnostics to measure whether latent distances reflect true task progress in JEPA-style world models, and int...