Latest stories
DeepSeek-V4.1-Flash Cuts KV Cache to 890 Bytes per Token
DeepSeek's new multimodal Mixture-of-Experts model pairs a Causal Encoder-Decoder design with CSA2 cross-layer reuse and FP4 KV ca...
JEPA-Anything: One Predictive Framework for Seven Scientific Worlds
Hugging Face researchers introduce JEPA-Anything, a domain-agnostic world-modeling framework built on orthogonal predictive factor...
Hugging Face Paper Unveils SoL-Pi, a Token-Efficient Agent Harness
A new Hugging Face paper, SoL-Pi, applies recursive self-improvement at the agent harness layer, matching Pi's EdgeBench performan...
Hugging Face Study Dissects What Actually Makes Coding Agents Work
A new empirical study from Hugging Face isolates planning, action space, and context management inside a fixed coding harness, eva...
PANORAMA Brings Pixel-Level Grounding to Detailed Image Captions
A new Hugging Face paper introduces PANORAMA, a vision-language model that grounds every referring phrase in a caption with pixel-...
New Memory-Bounding Techniques Let MoE Models Train at 1M Context
A new paper introduces four complementary methods that cap every memory peak in long-context Mixture-of-Experts training, enabling...
GPT-Policy Lets Robots Learn New Tasks In-Context via VLM Agents
A new Hugging Face paper introduces GPT-Policy, a general-agent framework that combines a context compiler, a commercial vision-la...
EventEgoHands++ Brings Instance-Aware 3D Hand Reconstruction to Event Cameras
A new Hugging Face paper introduces EventEgoHands++, an event-camera framework for egocentric 3D hand mesh reconstruction that add...
Gaze as a Cue to Mutual Understanding: Cross-Corpus Study of MapTask and MUNDEX
A new study maps two collaborative dialogue corpora into a shared gaze vocabulary and finds that aligned references and "understoo...
Hugging Face Paper Proposes Zeroth-Order Method to Fix LLM Preference Alignment
A new paper introduces Comparison-based Preference Optimization (ComPO), a zeroth-order alignment method that extracts directional...
SpectralShift: Extending Gated DeltaNet's Context Window via Spectral Reparameterization
A new research paper proposes SpectralShift, a spectral reparameterization method for long-context continual pretraining of Gated...
HypoEvolve Uses Genetic Algorithms to Coordinate Multi-Agent LLMs for Scientific Hypothesis Discovery
A new framework called HypoEvolve applies a generational genetic algorithm to coordinate specialized LLM agents that propose, crit...