HuggingFace

IBM introduces CUGA, an open-source agent harness that handles plumbing for state, tool calls, and orchestration, allowing developers to focus on defining tools and prompts. The article showcases two dozen single-file example apps built with CUGA, demonstrating how it eliminates repetitive framework setup.

This guest post explores the proposed Cross-Origin Storage API to improve caching of AI model resources in Transformers.js, enabling efficient reuse across origins while maintaining privacy and integrity for in-browser inference.

Hugging Face describes how they built a weekly release pipeline for their huggingface_hub library using AI, open-source tools, and human oversight, enabling faster and more reliable releases.

PP-OCRv6 is the latest generation of PaddleOCR's universal OCR model family, offering three tiers from 1.5M to 34.5M parameters, supporting 50 languages, and achieving significant accuracy improvements over previous versions.

ShotcreteDepth is a bi-modal dataset of stereo RGB and LiDAR data from construction environments, designed to support research in depth perception under challenging conditions. The dataset includes 11,252 samples with 220 annotated, and is accompanied by a lightweight annotation tool.

TROPT is an open-source framework that unifies discrete text-trigger optimization, standardizing development and execution across domains like LLM jailbreaking and model interpretability. It includes over 15 optimizers and 30 recipes, lowering barriers for adoption and advancement.

This paper introduces AgentCIBench, a benchmark to evaluate privacy risks in computer-use agents, finding that 11 of 15 frontier agents leak information in over 50% of scenarios.

Vera is a layered diffusion model for video editing that preserves source content by generating edit layers and alpha mattes, using a Mixture-of-Transformers architecture.

This paper introduces representational commitment, a cross-run hidden-state convergence that diagnoses when an LLM agent has locked onto a trajectory prematurely. It shows that commitment predicts trajectory consistency but not correctness, and proposes monitoring to detect when an agent is confidently settled rather than assuming consistency equals trust.

Arbor introduces explicit geometric control for 3D asset generation by using constraint meshes (hull, avoidance, touch regions) to condition latent generation, improving spatial constraint adherence without sacrificing object quality.

HAKARI-Bench is a lightweight benchmark for comparing retrieval methods across multiple configurations and languages, enabling efficient model selection and performance analysis. It reproduces full benchmarks like MTEB at high correlation while being faster to run.

MeshFlow introduces an equivariant optimal-transport flow matching model for direct triangle mesh generation, achieving state-of-the-art quality while providing approximately 18x inference speedup over autoregressive methods.

Foresight is a failure detection framework for long-horizon robotic manipulation that uses action-conditioned world model latents and functional conformal prediction to monitor trajectories, trained only with final task labels. It demonstrates state-of-the-art performance across simulation and real robot tasks.

The blog post describes using local open-weight models like Gemma and Qwen in an agent harness to automatically triage issues and pull requests in the OpenClaw repository, enabling real-time notifications without relying on costly closed API models.

KaLM-Reranker-V1 is a fast reranker that decouples query and passage computation using an encoder-decoder architecture with Matryoshka embedding pooling and cross-attention, achieving state-of-the-art reranking performance on BEIR and competitive results on multilingual benchmarks.

VESFlow is a training-free safety method for flow matching-based text-to-image generation that edits velocity fields to ensure safe output while maintaining prompt integrity.

EnterpriseClawBench presents a benchmark for enterprise agents based on real-world workplace sessions, offering 852 reproducible tasks and comprehensive evaluation metrics beyond single performance scores.

This paper argues that language model agents should assist causal discovery workflows by providing contextual support and explanations rather than generating causal conclusions, and introduces causal-learn+ platform to demonstrate this principle.

DR-MV3D presents a map-grounded learning framework with dense rewards to improve multi-view 3D visual question answering through global map construction, view-trajectory planning, and egocentric grounding.

SelfCompact is a scaffolding approach that lets language models autonomously decide when and how to compact long agent traces, achieving better performance with reduced token costs compared to fixed-interval methods.