HuggingFace

This paper introduces Harness-Zero, a method for distilling optimized agent harnesses into large language models via agent-as-harness, significantly boosting task performance in knowledge work, tool use, and science domains even after specialized harnesses are removed.

onPanda is an interactive tool that uses token-level correction to efficiently annotate LLM alignment data and agent trajectories, reducing median annotation time by 52%.

The paper introduces an information-efficiency ratio (IER) for optimizing token selection in sparse on-policy distillation, demonstrating that using only 1% of tokens can achieve performance comparable to full supervision.

CARE is a framework for Vision-Language-Action policies that enhances robotic manipulation by learning from execution failures to generate corrective actions, demonstrating improved success rates in simulations and real-world tasks.

Document Retrieval-Aware Chunking (D-RAC) is a method that normalizes enterprise documents to PDF, converts them to retrieval-optimized Markdown using a multimodal LLM, and chunks them efficiently, significantly reducing token usage and costs compared to agentic chunking.

WorldCrafter is a video world model that learns a camera-queryable implicit 3D-aware memory for consistent and camera-controllable streaming scene exploration from a single image or text prompt.

This paper introduces Regularized Recursive Self-Improvement (RRSI) for AI agent harnesses, which applies regularization to prevent overfitting during recursive evolution, demonstrating performance gains on multiple benchmarks.

Hugging Face releases tokenizers v1, a major performance update for the tokenization library, with benchmarks showing significant speed improvements over previous versions.

GGUF quantizations of the Qwen-Image-2.1 model for local image generation using ComfyUI, with recommended quantizations and setup instructions for deployment.

Hemmingway-1 is a 27B-parameter open-source AI model specialized for everyday writing tasks, outperforming leading models on benchmarks for human-like communication.

This paper proposes Grounded Action Models (GAMs), a new paradigm for robot foundation models that integrates 3D grounding, achieving state-of-the-art performance on manipulation tasks.

Mira-Scene introduces a compositional 3D scene reconstruction framework using pixel-aligned canonical coordinate maps for accurate object layouts, achieving significant improvements in layout accuracy over existing methods.

This paper introduces a category-aware expert training framework for software engineering agents to mitigate uneven progress across task categories, using iterative training and multi-teacher distillation, with significant performance gains on Pro-618 and SWE-bench Multilingual benchmarks.

This paper presents an interpretability study on video diffusion models, revealing that Rotary Position Embedding (RoPE) induces excessive spatial attention decay, causing physics violations, and proposes a lightweight architectural modification to enhance physical coherence in generated videos.

The paper proposes Calibrated Clipping to stabilize FP8 quantization in reinforcement learning for LLMs by aligning clipping bounds with high-precision distributions, eliminating entropy surges and restoring performance.

UltraTex is an efficient framework for high-resolution multi-view diffusion-based 3D texturing, introducing techniques to reduce redundancy and achieve significant speedups in training and inference.

OmniEdu introduces an open family of foundation models for K-12 education, trained on a curated corpus to improve problem-solving, curriculum grounding, and pedagogical tutoring capabilities.

This paper presents RoboDawn, a method to transfer Vision-Language Model intelligence to robotic control, achieving state-of-the-art results on benchmarks with zero-shot and one-shot learning and successful real-world applications.

Laya is an open-source non-autoregressive decision model that provides typed answers with calibrated probabilities, designed for tasks like email triage and conversational AI, showing significant performance improvements over existing models.

This paper presents HuRo, a pipeline for robotizing human videos to create scalable VLA pretraining data, showing significant improvements in task completion and robustness on real-world manipulation tasks.