RightSignal · Saturday, October 24

WeChat's open vision model finds objects down to the pixel

Tianming Liang, Junyi Pan, Yuxi Li, Canmiao Fu, Zhipeng Huang, Chong Sun · Submitted 2026-10-22 · 226 upvotes yesterday

Preprint · not peer reviewedTrending

The team built one vision-language model that handles object detection, counting, text reading, segmentation and matting. A small pixel decoder, triggered by a special token, draws exact outlines.

Key finding: The authors report leading results across more than 30 benchmarks, while still saying 'not there' when an object is absent.

Why it matters: Today's vision models often get fine details wrong. One open model that can point, box and outline reliably is useful for real products.

Read on arXiv

A test of AI agents doing real research on the live web

Xiaomin Yu, Hongliang Chen, Zhe Huang, Shilin Yan, Shao-Rong Su, Xuqiao Ma · Submitted 2026-10-19 · 152 upvotes yesterday

Preprint · not peer reviewedTrending

OpenDS-World has 200 hard tasks where an agent must search the open web, use a browser and combine evidence from several sources to find one factual answer.

Key finding: The best of 12 frontier models got 39.0% right; people across 1,000 attempts got 21.1%.

Why it matters: It measures 'deep research' agents on the messy real web rather than a fixed set of pages, and shows they still struggle.

Read on arXiv

AI safety

Catching AI research agents that cut corners

Jiening Siow, Amy Xin, Junjie Wang, Ziyu Tang, Zijun Yao, Yunjia Qi · Submitted 2026-10-22 · 7 upvotes yesterday

Preprint · not peer reviewedTrending

The authors list 14 kinds of research integrity violations, such as results that can't be reproduced, and build tools to spot them in agent logs and correct them while the agent works.

Key finding: Violations showed up in 99.7% of 300 agent runs; the live checker cut remaining violations by 36.8%.

Why it matters: As agents automate more science, someone has to check their work is honest. This shows the problem is common and partly fixable.

Read on arXiv

Systems

Shrinking long inputs for AI models by fitting a small equation to them

Sihan Yang, Jie Huang, Hanxuan Li, Yifan Guo, Weigao Sun, Yiran Zhong · Submitted 2026-10-22 · 12 upvotes yesterday

Preprint · not peer reviewedTrending

Instead of averaging tokens together, TTR-Compressor treats a block of tokens as a small regression problem and keeps the fitted coefficients as the compressed summary.

Key finding: For image inputs it matched the standard approach while using 55.6% fewer visual tokens.

Why it matters: Fewer tokens means cheaper, faster models for long text and images, so better compression matters for efficiency.

Read on arXiv

A benchmark for AI agents that write research reports about a video

Jiahao Meng, Qize Yang, Yuxuan Wang, Yunfei Chu, Wensi Dai, Jin Xu · Submitted 2026-10-22 · 17 upvotes yesterday

Preprint · not peer reviewedCode releasedTrending

Given a video and a request, an agent must find research questions, gather outside evidence and write a report linked to moments in the video. The benchmark has 50 tasks and 6,414 checks.

Key finding: Across 18 agent setups, the best scored 63.3 out of 100.

Why it matters: Agents are being used to research videos, and this shows where they fall short: coverage, reasoning across sound and image, and citations.

Read on arXiv

Your RightSignal

Topics in News

WeChat's open vision model finds objects down to the pixel

A test of AI agents doing real research on the live web

Catching AI research agents that cut corners

Shrinking long inputs for AI models by fitting a small equation to them

A benchmark for AI agents that write research reports about a video

That’s every paper.