Research Papers
Architect-Ant Automates Furniture Layouts for Real Floor Plans
Hugging Face researchers introduce AntPlan, a dataset of 505 annotated professional floor plans, and Architect-Ant, a vision-langu...
LoopCD Boosts Looped Transformer Decoding Without Extra Training
A new training-free contrastive decoding method, LoopCD, uses earlier recurrent passes of looped Transformers as weak references t...
4Director Uses Rigid 3D Geometry to Give Video World Models Precise Control
Hugging Face paper introduces 4Director, a video world model conditioned on an explicit 4D scene representation where each object...
SILSA Cuts 3D Generation Tokens by 98% While Boosting Topology Fidelity
Hugging Face researchers propose SILSA, a topology-aware 3D generation framework that replaces expensive voxel tokens with compact...
GPT-6 Astra Robot Agents Cut Token Use 65% While Boosting Success Rate
A new Hugging Face paper introduces PyRUA-Lean, a code-execution framework that lets GPT-6 Astra robot agents solve more tasks wit...
Hugging Face Paper: BeyondSCe Grasps Objects Referenced by Past Events, Even When Hidden
A new zero-shot robotic grasping system, BeyondSCe, lets a robot grasp objects referenced by the role they played in past events,...
Hugging Face Releases EgoTools, a 100-Hour Benchmark for Egocentric Tool-Use Reasoning
EgoTools introduces 100 hours of tool-centric egocentric video with dense captions and 3D data, plus a 1,000-question diagnostic b...
InterEvolve Lets Humanoid Robots Learn New Tasks at Test Time Without Retraining
A new Hugging Face paper introduces InterEvolve, a framework that evolves reward programs at test time so a humanoid controller ca...
Hugging Face Paper Proposes RASO to Reuse Existing Agent Skills
A new paper introduces Retrieval-Augmented Skill Optimization (RASO), a framework that treats publicly shared agent skills as prio...
AutoGUIWorld Uses Image Generators as Visual World Models to Train GUI Agents
Hugging Face researchers introduce AutoGUIWorld, a framework that synthesizes GUI interaction trajectories using image generators...
DARA Reweights Sparse Rewards to Speed Up Multi-Reward RL
A new Hugging Face paper introduces Density-Aware Reward Aggregation (DARA), a method that calibrates reward contributions in mult...
ROWBench Tests Whether Video Models Render Exactly What a Program Specifies
A new benchmark called ROWBench (PROWBench) evaluates how faithfully video and world models reproduce fine-grained, program-specif...