Research Papers
Pivot-SD: Training Masked Diffusion LMs on Just the Tokens That Matter
A new self-distillation framework, Pivot-SD, supervises only the high-impact denoising commitments in masked diffusion language mo...
Hugging Face Releases HyperBrowseComp, a 13-Language Stress Test for Web-Browsing AI Agents
A new benchmark of 423 human-validated questions across 13 languages challenges AI browsing agents to chase obscure, multi-step ev...
Latent-MOPD Brings Representation-Level Multi-Teacher Distillation to LLMs
A new Hugging Face paper introduces Latent-MOPD, described as the first representation-level multi-teacher on-policy distillation...
Hugging Face Paper: On-Policy Update Direction May Explain Why Post-Training Generalizes Better Than SFT
A new Hugging Face paper argues that the cumulative direction of parameter updates in on-policy post-training, not just the paradi...
NAVA-WAM: Pretraining Robot Action Policies Directly from Observation-Only Videos
A new Hugging Face paper introduces NAVA-WAM, a two-stage method that pretrains an action policy natively from observation-only vi...
ProWAM: Sparse Visual Sub-Goals Push World Action Models to New Robotic Control Records
A new Hugging Face paper introduces ProWAM, a progressive world action model that predicts an ordered sequence of sparse visual su...
RealCompanion Benchmark Tests Whether AI Companions Truly Understand Us Over Time
A new Hugging Face paper introduces RealCompanion, a benchmark built from ten real human-AI companion relationships spanning 27,21...
MotorMind Lets General VLMs Drive Robots Zero-Shot, Topping LIBERO-PRO
A new harness called MotorMind connects general-purpose vision-language models directly to deterministic robot control, reaching 6...
Hugging Face Paper: Recursive Self-Rewrite Turns Harness-Assisted Wins into Reusable Model Skills
A new paper proposes Recursive Self-Rewrite (RSR), a framework that uses a single base model to discover solutions under specializ...
Study Finds LLM Agents Favor Items by Source, Even When Worse
A new Hugging Face paper shows that LLM agents across 12 models and three domains systematically prefer items from certain sources...
Hugging Face Unveils World Embedding Benchmark to Test Physical Understanding in Video AI
A new benchmark of 8,000 simulated video-text cases probes whether video representations truly capture physics, revealing a trade-...
FrameMorrow Uses Future-Guided Frame Selection to Improve Long-Horizon Video Generation
A new research paper introduces FrameMorrow, a prospective frame selector that predicts future information needs to choose relevan...