Research Papers
UNREAL Unifies Retrieval and Long-Context Reasoning in a Single Frozen LLM
A new framework called UNREAL derives retrieval queries and chunk encodings from a frozen LLM's internal representations, adding u...
WING Framework Transfers Human Video Skills to Robots via Spectral Latent Guidance
A new research framework called WING learns interaction-centric latent actions from egocentric human videos and transfers them to...
MiniCorp Simulator Lets AI Agents Run a Company — and Generates Enterprise Training Data
Hugging Face researchers introduce MiniCorp, an office simulator where AI agents collectively run an e-commerce company, generatin...
TRIAGE: Hugging Face Paper Stabilizes Native NVFP4 Reinforcement Learning
A new paper proposes TRIAGE, a direction-aware stabilization method that keeps native NVFP4 4-bit execution on both sampler and le...
New Method Lets Robot Policies Compensate for Their Own Execution Errors
Researchers propose self-compensating VLA, a deployment-time adaptation method that updates a vision-language-action policy online...
EVISKILL Ties LLM Agent Skill Edits to Replayable Evidence
A new Hugging Face paper introduces EVISKILL, an evidence-driven framework that stores execution observations as Replayable Eviden...
HuatuoGPT-3 Brings RL-Only Medical Adaptation to Base LLMs
A new Hugging Face paper introduces OnePO, a one-stage policy optimization method that adapts base LLMs to the medical domain with...
Hugging Face Paper Exposes Where Tool-Using AI Agents Break Down
A new Hugging Face paper introduces SafeActBench, a 656-case benchmark showing that tool-using agents can produce correct outcomes...
TRACE Aligns Train and Rollout Quantization for FP4 RL of MoE Models
A new Hugging Face paper introduces TRACE, an FP4 quantization framework that uses rollout-side quantization outcomes to guide tra...
AutoSciBench Lets Scientific Agents Build the Tests They Take
Hugging Face paper introduces AutoSciBench, a framework that automatically generates and iteratively refines scientific-agent benc...
DuoMatching Adds Image-Teacher Supervision to Few-Step Video Generation
A new Hugging Face paper proposes DuoMatching, a joint-marginal distribution matching framework that pairs video-teacher joint mat...
Hugging Face Paper Argues Supervision Reliability Beats Alignment Coverage in Cross-Tokenizer Distillation
A new Hugging Face paper rethinks cross-tokenizer on-policy distillation, finding that compact supervision at strictly aligned pos...