Research Papers
StepAudio 3 Gen Unifies Speech, Sound Effects and Music in One Autoregressive Model
A new technical report presents StepAudio 3 Gen, a discrete autoregressive audio generator built on residual vector quantization t...
ReactHuman Benchmark Exposes Reactive Safety Gaps in Embodied Multimodal LLMs
Hugging Face researchers introduce ReactHuman, a physics-grounded benchmark that tests multimodal LLMs as decision cores for simul...
Feyospace-v1: Seven-Person Team Trains Open-Weight Cyber Agents to Top-Tier Results
A data-centric framework with five specialized systems lets a small independent team post-train open-weight cyber agents that impr...
DataFlex-RL Finds Uniform Sampling Hard to Beat in RLVR Data Policies
A new Hugging Face evaluation platform, DataFlex-RL, benchmarks rollout selection, reweighting, and domain mixing for reinforcemen...
SNAP3D Brings Physics to Single-Image 3D Part Assembly
A new physics-guided framework from Hugging Face researchers resolves inter-part penetration, recovers contact graphs, and refines...
SAS Ranks Context End-to-End to Make Sparse Attention Count
A new Hugging Face paper proposes Simple Attention Sparsification (SAS), which injects a selector's continuous scores into attenti...
COBRA-Skills Uses Contextual Bandits to Cut LLM Agent Skill Optimization Costs by Up to 58%
A new Hugging Face paper introduces COBRA-Skills, a framework that treats agent skill optimization as a budgeted sequential proble...
PLC-DPO Routes Noisy Preference Labels Instead of Filtering Them
A new Hugging Face paper proposes Posterior Label Correction DPO, which uses calibrated policy-reference margins to route each pre...
Hugging Face Paper Unveils Benchmark Radar, a Living Database of 1,283 AI Benchmarks
Benchmark Radar is a searchable living database and discovery engine that aggregates benchmark sources, score histories, and evalu...
Latent Interface Training Cuts Robot Models' Reliance on Visual Shortcuts
A new two-stage method called Latent Interface Training (LIT) teaches robot foundation models to act from spatial goals before exp...
Adaptive Bridge isolates ROS 2 subscribers from DDS backpressure
A new proxy layer splits ROS 2 topics into separate reliable and best-effort DDS writers to stop one degraded subscriber from stal...
ActReview Turns Author Rebuttals into Training Signal for Actionable Peer Review
A new Hugging Face paper introduces ActReview, a rebuttal-guided post-training framework that generates diagnostic claims and conc...