Research Papers
LatentStream: A New Framework for Streaming Video Understanding via Progressive Latent Memory
Researchers introduce LatentStream, a progressive latent working memory framework that shifts streaming video understanding from e...
Principia Benchmark Exposes Major Physics Gaps in Video Generation Models
A new benchmark, Principia, tests video generators on Newtonian physics using calibration-independent relational consistency, find...
Environment Evolution: A New Method to Keep Terminal Agents Learning as They Improve
Researchers propose 'environment evolution,' an off-policy method that incrementally increases task difficulty to provide continuo...
AutoTraceGT: Grounded Theory Meets AI to Decode Agent Behavior at Scale
Hugging Face researchers introduce AutoTraceGT, a multi-agent pipeline that automates grounded theory coding on agent trajectories...
FlashRender: Few-Step Generative Rendering Achieves 25x Speedup with Camera Control
FlashRender introduces a few-step generative rendering framework that retakes a source video along a target camera trajectory in s...
Hugging Face Unveils Last Translation Benchmark to Stress-Test AI Translators
The Last Translation Benchmark introduces peer-reviewed, multimodal examples that break leading translation models, alongside hand...
RoboTok: Web-Scale Data Engine Boosts Dexterous Robot Learning from Human Videos
RoboTok is an internet-scale data engine that retrieves relevant human manipulation videos from the web using a latent motion spac...
Temporal Context Routing: New Method Boosts Script-Timed Audio-Video Generation
Hugging Face researchers introduce Temporal Context Routing (TCR), a module that aligns script timestamps with the shared audio-vi...
Hugging Face Researchers Unveil Editable Visual Design: A Coding Agent Paradigm for Layered, Aesthetic Graphics
A new paradigm from Hugging Face combines vision-language models with coding agents to generate editable, layered designs with rea...
Scal3R: Efficient Multi-Relative Pose Query for Scalable Online 3D Reconstruction
Hugging Face researchers introduce Scal3R, a method that improves long-video online 3D reconstruction by querying multi-reference...
Knowing When Not to Reuse: New Method Boosts Autonomous LLM Post-Training
Researchers introduce Boundary-Calibrated Intervention Transfer (BCIT), a method that selectively reuses past training evidence in...
Terminal-Universe: Turning Agent Trajectories into Scalable Training Environments
Hugging Face researchers introduce Terminal-Universe, a framework that reconstructs executable workspaces from agent trajectories,...