Latest stories
ReSPO Targets Gradient Starvation in Off-Policy RL for LLMs
Hugging Face paper ReSPO replaces clipping in policy optimization with a smooth two-branch sequence-level kernel to prevent sign-d...
ViSkill Lets VLM Agents Learn From Visual Skill Cards, Not Just Text
A new framework from ZJU-REAL encodes successful agent interactions as reusable visual skill cards that guide both inference and r...
Study Finds LLMs Can Recognize Sequential Strategies but Fail to Simulate Them Faithfully
A controlled study using Rock-Paper-Scissors and n-gram continuation tasks finds that longer context does not improve strategy ide...
Mara Chain Turns Rejected AI Configurations Into Stepping Stones for Auto-Evolution
A new refinement procedure from Ant Research keeps rejected candidate configurations and iteratively improves them, delivering lar...
WorldGuide Turns Video Generation Into Closed-Loop Task Execution
A new Hugging Face paper from MBZUAI-ORYX introduces WorldGuide, a goal-directed video world model that plans and executes procedu...
Hugging Face Paper Unveils Compo: A Spatial Canvas That Turns Poster Generation Into Composing
A new Hugging Face paper proposes a Spatial Canvas Interface that lets users express two-dimensional poster composition directly i...
Hugging Face Paper Unveils SpaceCast-Bench for Predictive Spatial Reasoning in VLMs
A new benchmark, SpaceCast-Bench, directly evaluates whether vision-language models can predict unseen outcomes of spatial transfo...
Hugging Face Paper Proposes 'Retrieval-Centric Deep Learning' With Growing Neural Networks
A new paper develops principled functional-gradient learning rules for kernelized attention layers, replacing fixed-size weight ma...
REMORY: Hugging Face Paper Adds Learned Residual Memory Tokens to Compress Agent Context
A new paper introduces REMORY, a neural memory network that appends a bounded sequence of soft memory tokens to a compacted summar...
Opera: A Verbal Critic That Follows Feedback Until Coding Agents Actually Fix the Problem
Hugging Face researchers propose Opera, a verbal critic framework for long-horizon coding agents that treats each correction as a...
OmniCapBench Turns Audio-Visual Captioning Into a Diagnostic Test for Multimodal AI
A new benchmark from Hugging Face researchers reframes audio-visual caption evaluation as structured prediction, using atomic veri...
SanSi Turns Looped Language Models Into Typed Decision Makers
A new Hugging Face paper introduces SanSi, a looped typed decision model that returns option probabilities without generating reas...