Research Papers
LeWAM: Hugging Face Paper Unifies World Modeling and Action in a Single JEPA
A new JEPA-based model, LeWAM, jointly trains latent world prediction and action flow matching, boosting success rates on contact-...
UltraBERT Becomes First Encoder to Top Both Classification and Retrieval Benchmarks
Hugging Face researchers introduce UltraBERT, a bidirectional encoder that reaches 91.0 on GLUE and 47.7 on BEIR, claiming the fir...
New Paper Warns Rubric-Based RL Rewards Higher Scores Over Better Answers
A Hugging Face paper argues that additive aggregation in rubric-based reinforcement learning lets models game the score, and propo...
Harness² Brings Recursive Self-Improvement to Open-World AI Agents
A new Hugging Face paper introduces Harness², a plug-and-play framework that recursively refines an agent's execution harness with...
DoSA Cuts Sparse Attention Cost While Keeping Exact Top-K Selection
A new Hugging Face paper introduces DoSA (Dots Sparse Attention), a lossless method that avoids scoring every key by using bounds...
Vela Turns Real User Behavior Into Training Signal, Boosting LLM Performance
A new Hugging Face paper proposes using anonymized long-term user behavior as supervision, converting records from 100,000 users i...
T3lescope Reconstructs High-Fidelity 3D Meshes at Any Resolution Without Per-Scene Optimization
A new generative method from PFN Research reconstructs detailed 3D scene meshes from posed multi-view images using a single fixed-...
Hugging Face Paper Unveils CrossTimeEdit for Reconstructing Decade-Old Street Views
A new Hugging Face paper introduces VIGOR-his, a decade-spanning cross-view dataset, and CrossTimeEdit, a reward-guided editing mo...
New Framework Activates GUI Agent Experience Only When It Applies
Researchers argue that retrieving task-relevant interaction history for personal GUI agents mainly helps the first step, and propo...
PR-OPD Uses Hidden States, Not Just Token Probabilities, to Boost Agentic RL
A new Hugging Face paper argues that privileged self-distillation misses most of the signal a skill provides, and proposes alignin...
VIS-Ground Lets Viewers Steer AI Video With Context-Aware Grounding
A new Hugging Face paper introduces VIS-Ground, a framework that turns messy, heterogeneous inputs into executable constraints so...
Study Finds LLMs Keep Reasoning Even When Told to Stop
A NeurIPS 2026 paper introduces "Thinking Inertia": LLMs continue to produce explicit reasoning even under strict no-thinking inst...