Research Papers
RIDE: New Distillation Method Treats RL Teacher as a Direction, Not a Destination
A new Hugging Face paper proposes RIDE, an on-policy distillation method that extrapolates the representation-space shift induced...
Google DeepMind Unveils SynthID Bio, Watermarking AI-Designed Proteins
Google DeepMind introduced SynthID Bio, a family of watermarking methods that embeds imperceptible, verifiable signatures into AI-...
LongCat-DeepResearch Debuts Multi-Agent Workflow for Evidence-Grounded Reports
A new technical report introduces LongCat-DeepResearch, a system pairing an enhanced LongCat model with a multi-agent workflow tha...
EngiWorld Benchmark Exposes Big Gap in AI Agents for Professional Engineering
Hugging Face paper introduces EngiWorld, the first benchmark covering the full engineering design loop with 1,301 expert tasks acr...
Hugging Face Paper Unveils VoxPolyMem: Interaction-Aware Memory for Multi-Party Spoken Conversations
A new Hugging Face paper introduces VoxPolyMem, an interaction-aware multimodal memory framework for multi-party spoken conversati...
Hugging Face Paper Proposes AsyncLLM: General Asynchronous Agents Without Task-Specific Training
A new Hugging Face paper introduces AsyncLLM, an open-source framework that moves concurrency inside LLM inference, letting pretra...
CorpusMap Builds an Entity-Based Map to Make Agentic Search Cheaper and Sharper
A new Hugging Face paper introduces CorpusMap, an offline navigation layer that organizes document collections around recurring en...
SAKI Reroutes Teacher Supervision for On-Policy Distillation
A new Hugging Face paper introduces SAKI, a method that uses KL-constrained teacher-guided rollouts and maximal coupling to route...
Hugging Face Paper Reveals 'Phase Sensitivity' Weak Spots in Chunked KV-Cache Compression
A new study finds that chunked KV-cache compression creates periodic retrieval weak spots tied to a token's phase, shifting long-c...
LEGO-Anything Turns Single Images into Editable 3D Scene Programs
Hugging Face researchers propose LEGO-Anything, an Image-to-Code framework where coding agents write and revise Blender programs t...
PanoVLN Boosts Panoramic Vision-and-Language Navigation
Hugging Face paper introduces PanoVLN, a panoramic VLN approach combining longer-horizon action prediction, targeted route supervi...
Omni-IO Skills Wraps Existing Agents to Make Them Omni-Native Without Retraining
A new Hugging Face paper introduces Omni-IO Skills, a plug-and-play harness that gives general-purpose agents multimodal abilities...