Research Papers
Hugging Face Paper: GPD Distills 3D Geometry Cues to Sharpen Spatial Reasoning in VLMs
A new paper introduces GPD, a geometry-privileged distillation method that routes question-relevant depth, semantic, and bird's-ey...
Hugging Face Paper: High Accuracy Doesn't Mean LLM Agents Admit Uncertainty
A new Hugging Face paper proposes evaluating LLM agents on "epistemic humility" — whether they recognize, act on, and communicate...
BrickBench Tests Whether AI Agents Can Actually Build With LEGO
Hugging Face researchers introduce BrickBench, a benchmark that asks coding agents to design buildable LEGO assemblies from text p...
Hugging Face Paper Unveils AgenticBBO-Bench for LLM-Driven Black-Box Optimization
A new cross-domain benchmark, AgenticBBO-Bench, standardizes how LLM agents are evaluated on black-box optimization, finding they...
Hugging Face Paper: One Recurrent Transformer Block Rivals Full-Depth Vision Encoders
A new paper introduces reViT, a recurrent vision Transformer that reuses a single block across depths and programs its FFN as a co...
Hugging Face Paper Tests Whether AI Agents Can Drive Open-Ended Science
A new study introduces Station, a multi-agent simulated scientific ecosystem, and reports that it rediscovered 62.7% of findings f...
Hugging Face Paper Unveils Hebero Benchmark and DGPO for Multi-Task Robot Learning
A new GPU-parallel Isaac Lab benchmark, Hebero, enables joint training and evaluation of a single policy across 40 heterogeneous m...
V-CoLA Brings Training-Free Vision Token Compression to Linear-Attention VLMs
A new Hugging Face paper proposes V-CoLA, a training-free framework that compresses vision tokens specifically for hybrid vision-l...
Hugging Face Paper Introduces SpatialOPSD, Distilling Tool-Free Spatial Reasoning into MLLMs
A new Hugging Face paper proposes SpatialOPSD, an on-policy self-distillation framework that internalizes the spatial reasoning of...
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...