Research Papers
CoWindow Attention Cuts Long-Context Training Latency Up to 8.6x While Matching Full Attention
Hugging Face researchers introduce CoWA, a structured attention design in which heads collectively cover the full causal history i...
Hugging Face Paper: MALA Lets Attention Allocate Its Own Compute
A new fused attention primitive, MassAlloc Attention (MALA), preserves full causal QK scoring but uses normalized softmax contribu...
Duplex-MPE Benchmark Tests When Voice Assistants Should Stay Silent in Group Chats
Hugging Face researchers introduce Duplex-MPE, a 2,000-scenario benchmark that evaluates full-duplex speech assistants on selectiv...
Xiaomi's GAGAR Rewrites Code Agent RL by Grading Solution Quality, Not Just Test Passes
A new framework from Xiaomi and Hugging Face uses a groupwise agentic grader to redistribute advantage among test-passing trajecto...
Hugging Face Paper: Domain-Normalized Distillation Fixes Multi-Teacher Imbalance
A new paper introduces Domain-Normalized Multi-Teacher On-Policy Distillation (DN-MOPD), which rescales each specialist's feedback...
Hugging Face Paper: Post-Training Leaves 'Behavioral Shadows' That Transfer Skills
A new paper introduces Active Taskless Distillation (ATD), showing that a language model can learn coding, science, and reasoning...
CompoWorld Composes Reusable Services to Train More Capable General Agents
A new Hugging Face paper introduces Compositional Environment Scaling, which builds agent training tasks by composing a library of...
EmbodiedMemory-Bench Tests Whether Embodied Agents Can Actually Remember
A new benchmark of 2,554 interactive episodes probes four memory weaknesses in long-horizon embodied agents, alongside a structure...
Adaptive Looped Transformers Boost Test-Time Scaling by 53% on AIME
A new paper introduces TaH2, a post-training method that directs extra loop iterations only to tokens that benefit from them, impr...
Hugging Face Paper Proposes 'Physical Coding' to Make Robots Learn Like Software Agents
A new paper argues that representing robot task state and execution as code can overcome the brittleness of vision-language-action...
TraceDance Turns Real Agent Deployment Traces Into Targeted Behavior Benchmarks
Hugging Face paper TraceDance automatically builds benchmarks from real agent deployment traces to test specific undesirable behav...
YuE2 Unifies Symbolic and Audio Music Generation, Rivaling Suno
Hugging Face researchers introduce YuE2, a single Mixture-of-Transformers model that plans music as a readable score before render...