Latest stories
Argo-Bench Tests Whether AI Agents Can Actually Run Enterprise Data Workflows
Hugging Face researchers introduce Argo-Bench, a 210-task benchmark built on a simulated food-delivery platform with 81 million or...
Ego2Act Benchmark Tests Whether Video Models Can Actually Complete Everyday Tasks
Hugging Face researchers introduce Ego2Act, a goal-directed benchmark of 2,640 egocentric videos spanning 110 real-world tasks, pl...
CorrGRPO: Correlation-Normalized GRPO Balances Multi-Reward Learning
Researchers propose CorrGRPO, a variant of Group Relative Policy Optimization that normalizes pairwise reward covariances into Pea...
Multimodal Flow: A Fully Continuous Approach to Unified Language and Vision Modeling
Researchers introduce Multimodal Flow, a fully continuous generative model that represents language and vision as ordered continuo...
AutoDataBench Isolates Data Intelligence as the Key Variable in Auto-Research Agents
A new Hugging Face paper introduces AutoDataBench, a controlled testbed that isolates data intelligence—diagnosis, organization, a...
PixelDense Splits Semantic and Geometric Teachers to Boost Pixel Diffusion
A new Hugging Face paper introduces PixelDense, a representation-alignment method that routes semantic and geometric teachers thro...
Omni-Embed-Mini Binds Six Modalities Without Forgetting Text, at Just 0.9B Parameters
A new 0.9B-parameter omni-modal embedding model maps text, speech, audio, images, video and visually-rich documents into one share...
Smaller Models Make Better Rejects: New Study Challenges Preference Distillation Assumptions
A new Hugging Face paper finds that smaller frozen models can generate rejected responses that are cheaper to produce yet train st...
PhysVista Benchmark Tests Whether VLMs Truly Understand Physical Reality
Hugging Face researchers introduce PhysVista, a benchmark that evaluates physical intelligence in vision-language models through a...
HeteroFold Lets Different LLM Families Share KV Caches Without Prefill
A new paper introduces HeteroFold, a prefill-free method that transfers key-value caches between heterogeneous LLMs such as Llama,...
Hugging Face Paper Tests Whether Explanations Improve Forecasts
A new Hugging Face paper introduces "predictive credit," a protocol for measuring whether scientific explanations add value to exp...
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...