Research Papers
DyPES-VLA: A New Cross-Embodiment VLA Model with Shared Dynamics Priors
Researchers propose DyPES-VLA, a cross-embodiment VLA model that learns shared dynamics priors via future prediction and uses an e...
Hugging Face Unveils W2-VLA: Task-Conditioned Future Wrist Modeling Boosts Fine-Grained Robot Manipulation
Hugging Face researchers introduce World-to-Wrist VLA (W2-VLA), a vision-language-action model that predicts future wrist states c...
Hugging Face Adapts NVIDIA's Nemotron Stack for Modern Greek RAG, Launches HERA Benchmark
Researchers present an end-to-end adaptation of NVIDIA's Nemotron retrieval stack for Modern Greek, introducing the HERA benchmark...
Invisible Shortcuts: How Vision Encoders Learn Camera Fingerprints
A new study reveals that vision models exploit invisible metadata traces in pixels, learning camera fingerprints and processing cu...
Continual Learning in Transition: From Parameters to System-Level Adaptation
A new research paper from Hugging Face proposes a tri-axial framework—When, How, and Where—to characterize the evolution of contin...
Hugging Face Unveils MameLoshnLM, First Open-Source Yiddish LLM and Benchmark
Researchers introduce MameLoshnLM, an 8B-parameter model tailored for Yiddish, along with new corpora and benchmarks to counter th...
Hugging Face Unveils KVAE: A New Family of Multimodal Tokenizers for Generative Models
KVAE introduces a family of tokenizers for audio, image, and video, designed for latent diffusion models, achieving competitive or...
PaDoc: Parallel Decoding Framework Speeds Up Document Parsing While Preserving Full-Page Context
Hugging Face researchers introduce PaDoc, a layout-grounded document parser that enables parallel decoding of layout and content b...
Deterministic Screen-Activity Compiler Turns Agent Memory into Auditable, Replayable Frames
A new zero-model pipeline compiles passively captured screen activity into typed, byte-identical memory frames, cutting context si...
Hugging Face Unveils DataSpace: A New Benchmark for Verifiable Data Agents in Complex Workspaces
Hugging Face introduces DataSpace, a benchmark with 410 cross-language tasks and 7,439 artifacts (15.01 GB) across six formats, ch...
Hugging Face Paper Outlines Blueprint for 'Economic World Models' as Generative AI Engines
A new paper from Hugging Face proposes a six-level capability ladder for building Economic World Models (EWMs) — generative simula...
New Benchmark HarnessOpt-Bench Measures How Well LLMs Optimize Agent Harnesses
Researchers introduce HarnessOpt-Bench, a benchmark for evaluating LLMs' ability to optimize agent harnesses under budgeted, stoch...