Latest stories
GeoPair Introduces Training-Free Cross-Layer Compression for Transformers
A new Hugging Face paper proposes GeoPair, a training-free framework that jointly optimizes cross-layer weight pairings and shared...
Researchers Extract Hidden Chain-of-Thought from Frontier Models Like GPT-6 Astra
A new preprint shows that a simple tool-calling setup can expose the hidden reasoning traces of closed-source frontier models, rev...
Schrödinger's Repository: New Benchmark Tests Whether Coding Agents Truly Understand Code or Just Memorize It
A new evaluation framework called SchrodingerRepo dynamically transforms test repositories at evaluation time to strip away famili...
JitMem Defers Agent Memory Curation to Read Time, Boosting Task Success
A new paper proposes Just-in-Time Memory (JitMem), which stores raw agent trajectories losslessly and curates them at read time on...
PACT: New RL Method Aligns Critics with Policy Updates, Boosting Math and Coding Performance
Hugging Face researchers propose three regularity conditions that uniquely determine token-level credit in LLM reinforcement learn...
Hugging Face Paper Trains VLMs to Reason About Space Through Physical Interaction
A new research paper introduces Spatial-Interactor, a framework that teaches vision-language models to model physical-world state...
RewardVerse Uses Dynamic Rubrics to Stabilize Video Reward Models
Hugging Face researchers propose RewardVerse, a rubric-based video reward framework that inserts explicit evaluation criteria betw...
SpeakerMem-R1 Brings Speaker-Centered Dual-Track Memory to Multi-Party Dialogue
A new Hugging Face paper proposes SpeakerMem-R1, a dual-track memory framework that separates verbatim speaker-labeled messages fr...
HappyWorld-Bench: A New Yardstick for Testing Whether AI World Models Stay Reliable Under Interaction
Hugging Face researchers introduce HappyWorld-Bench, a benchmark spanning video, spatial, and embodied world models that measures...
Hugging Face Survey Maps Memory Mechanisms in Autoregressive Video Generation
A new survey frames temporal persistence in autoregressive video generation as a memory problem, organizing existing work across f...
RoboFollow Benchmark Exposes the 'Instruction Following Mirage' in Embodied AI
A new diagnostic benchmark argues that high success rates for embodied agents mask weak language grounding, tracing the problem to...
LatentPort Shows a 4B Model Can Hand Its Memory to a 9B Model Without Replaying Context
A new Hugging Face paper demonstrates the first cross-model handoff of persistent recurrent inference state between differently si...