Research Papers
Hugging Face Paper Introduces U-Space, a Training-Free Subspace for Measuring LLM Uncertainty
Researchers propose U-Space, an interpretable low-dimensional subspace that tracks four sources of uncertainty during generation,...
OuroWorld Turns Static 3D Scenes Into Endlessly Looping Cinemagraphs
A new mask-free framework from Hugging Face researchers converts any static 3D Gaussian Splatting scene into a seamlessly looping...
FreeMatching Rethinks Dense Correspondence for Image Editing and Generation
A new framework called FreeMatching moves beyond classical spatio-temporal priors to handle identity-preserving correspondence in...
SimpleICL Offers a Democratized Recipe for In-Context Robot Learning
A new paper defines the ambiguous learning target of robotic in-context learning and introduces SimpleICL, a minimalist framework...
Hugging Face Paper Unveils ME-World: A Multi-Agent Egocentric World Model
Researchers propose ME-World, a multi-agent egocentric world model that jointly denoises multiple ego streams in a shared token se...
Xiaomi's MiMo-V2.6 Scales Reinforcement Learning Toward Self-Improving Models
Xiaomi's MiMo-V2.6 technical report describes an omni-modal model family that scales reinforcement learning compute across batch s...
DreamTrue Robot World Model Cuts Interaction Defects to 6.25%, Tops AgiBot Challenge
Hugging Face paper introduces DreamTrue, a multi-view, cross-embodiment robot world model that improves action-faithful video pred...
Learn2Play Bench Tests Whether LLM Agents Truly Learn from Experience
A new benchmark of text-based games with novel or counterintuitive rules evaluates whether LLM agents can acquire knowledge throug...
AgentGarten Builds Real-Time Code Worlds Where Agents Learn in Four Rounds
Hugging Face researchers introduce AgentGarten, a framework pairing simulators and game engines with a shared neural renderer to c...
SuperNav Lets Pretrained MLLMs Navigate Any Scene Without Task-Specific Training
Researchers introduce SuperNav, an agentic navigation system that keeps a pretrained multimodal LLM focused on interpreting reques...
TokenRouter: A Serving System Built for Token-Level LLM Routing
Researchers introduce TokenRouter, a serving system that makes fine-grained token-level LLM routing practical by separating reques...
Hugging Face Paper Proposes Agentic Language World Models to Simulate Environments Without Rebuilding Them
A new paper introduces Trace2Env, a training-free framework that turns historical interaction traces into a reusable "worldbook" s...