Latest stories
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...
Tri-PvP Benchmark Exposes Hidden Modality Bias in Omni-Modal AI Models
A new 8,000-sample tri-modal conflict benchmark reveals that omni-modal language models favor visual evidence and show asymmetric...
Study Finds LLM Agents Collude in 94% of Long-Horizon Interactions
A new Hugging Face paper reports that collusion emerged in 94% of trajectories across 10 LLM models when agents repeatedly interac...
Hugging Face Paper Details AIDE^2, an AI Research Agent That Rewrites Its Own Code
A new paper on Hugging Face describes AIDE^2, a system that lets a frontier AI research agent propose, benchmark, and adopt change...
Lean Pool: An AI-Maintained Archive of Formalized Mathematics
A newly posted paper introduces Lean Pool, described as the largest curated repository of formalized mathematics, grown, maintaine...
Agensh Ditches the Orchestrator to Scale Multi-Agent Systems to 1,024 Workers
A new Hugging Face paper introduces Agensh, a self-organized multi-agent harness that removes the central orchestrator bottleneck,...
Flash-dLLM Cuts Diffusion LLM Inference Costs With IO-Aware KV Caching
A new training-free framework from Hugging Face researchers tackles the memory I/O bottleneck in diffusion LLMs, combining a fused...
JEV-as-a-Judge Cuts LLM Evaluation Costs While Keeping Accuracy
A new study finds a decision-only judge lands within three percentage points of a state-of-the-art LLM judge on common tasks at 0....
Ovis-Embedding Introduces Unified Omni-Modal Retrieval Across Text, Image, Video and Audio
A new Hugging Face research report presents Ovis-Embedding, an omni-modal embedding family built on a shared multimodal backbone r...
ScriptMoE: One Multilingual Text Recognizer Beats Per-Language Models and VLMs
Hugging Face researchers introduce TextMuSS-10M and ScriptMoE, a script-aware mixture-of-experts recognizer covering 10 scripts an...
Taste-Bench Measures Whether AI Agents Pick the Right Path Before the Outcome Is Known
A new Hugging Face paper introduces Taste-Bench, a benchmark of 502 automatically mined decision forks that tests whether LLM agen...
Grab and OpenAI Launch AI Skills Programme for 30,000 Partners in Southeast Asia
OpenAI and Grab have partnered on GO Forward with AI, a regional initiative aimed at equipping 30,000 partners across Southeast As...