Search - metacurate.io

www.analyticsinsight.net Tools and repositories

Vision-Language-Action (VLA) Models

VLA models fuse vision, language, and action to enable robots to make real-time decisions in changing environments, allowing them to adjust to obstacles and unexpected events.

1 min read Listed 6 days ago in RSS: analyticsinsight

arxiv.org Scientific research

[2610.32216] PEAR: Adaptive Advantage Reshaping for Perception and Exploration in Large Vision-Language Models

PEAR introduces an adaptive entropy-based exploration gain on top of visual dependency to better balance perception and reasoning exploration in multimodal models, improving performance on multimodal and textual benchmarks.

1 min read Listed 19 days ago in Bluesky

tencent.github.io Tools and repositories

WeVisOne-VL: Unified Visual Perception from Objects to Pixels

WeVisOne-VL is a unified vision-language model that combines object- and pixel-level perception using structured next-token prediction, enabling tasks such as detection, localization, counting, pointing, OCR, segmentation, and matting at varying granularities.

1 min read Listed 15 days ago in reddit/computervision/Vision artifacts this week: pixel grounding, long-document OCR, streaming state and robot planning

luma.com Tutorial

Erfan Shayegani - Visual Grounding Safety in Vision-Language Models · Luma

The document discusses safety gaps in visual grounding for vision-language models and how targeted training can improve safety while preserving visual capabilities, presented by Erfan Shayegani during an Apple internship.

1 min read Listed 16 days ago in X

arxiv.org Scientific research

[2610.40636] Probing Spatial State Representations in Visual Foundation Models

The paper investigates whether visual foundation models maintain a usable spatial state under self-motion by probing four capability levels with streaming egocentric views; results show readout of the visible scene often fails to guarantee object persistence or future feature retrieval, and goal changes alone do not guarantee improved action, highlighting limits in current spatial representations for persistent, predictive, and actionable use.

2 mins read Listed 16 days ago in Hugging Face Daily Papers

arxiv.org Scientific research

[2610.26625] EgoXperience: Learning World-Grounded Egocentric Representations from Human Experience

EgoXperience introduces a full-stack approach for learning world-grounded egocentric representations, featuring an automatic annotation engine, a large annotated dataset, and a model that jointly reconstructs camera, scene, hands, and objects with strong empirical gains over baselines.

2 mins read Listed 27 days ago in Hugging Face Daily Papers

arxiv.org Scientific research

[2610.30173] Neural Networks with Differentiable Belief-Function Layers

A deep evidential neural network uses Yager's rule to propagate Dempster–Shafer masses, making m(Θ) the explicit measure of uncertainty. It contrasts with Dempster's normaliser, which can create unstable gradients, and demonstrates that Yager's vacuity better captures disagreement and supports competitive out-of-distribution detection without needing annotator distributions during training.

1 min read Listed 22 days ago in Bluesky

arxiv.org Scientific research

[2611.11632] Perturbation Profiles of Scanpath Classifiers: What Reading, Code-Comprehension and Video Decoders Are Sensitive To

The study tests which scanpath features support eye-movement classification by altering fixation order, timing, orientation, and spatial structure, then refitting classifiers. Across reading, code-comprehension, and viewing tasks, the profiles reveal sensitivities to spatial spread and stimulus layout, alongside task-specific results for recurrence and graph-based descriptors.

2 mins read Listed 4 days ago in Bluesky

sciherald.com News

Trending Vision-Language Models This Week (November 9, 2026) | Sci Herald

The article lists 10 vision-language models trending on Hugging Face, with brief descriptions, popularity figures, licenses, and estimated hardware needs. Their uses include OCR, document analysis, coding, reasoning, and agent tasks; the ranking reflects popularity, not quality.

3 mins read Listed 5 days ago in Bluesky

docs.venice.ai Tools and repositories

All Models | Venice API Docs

Catalog of AI models with IDs, capabilities, usage limits, and pricing; includes multiple vendors and variants with notes on privacy and retirement.

23 mins read Listed 19 days ago in X