RightSignal · Tuesday, November 3

Systems

Wave Forcing: Towards Speed-of-Light Streaming Video Generation

Yijie Jin, Yuxuan Liu, Jiajun Xu, et al. · Submitted 31 Oct 2026 · 99 upvotes yesterday

Preprint · not peer reviewedTrending

The authors generate long videos in real time by letting each new chunk start before earlier chunks finish, and by spreading the stages across several copies of the model on different GPUs.

Key finding: On eight H200 GPUs the system reaches 134.8 frames per second end to end, 7.07 times a single GPU, with quality comparable to an earlier method.

Why it matters: Live AI video, such as interactive worlds, needs frames faster than people watch them; this shows a practical way there.

Read on arXiv

CUA-Master: Scaling Computer-use Agents through Open-World Recursive Improvements

Xinlei Yu, Gen Li, Ji Zhao, et al. · Submitted 2 Nov 2026 · 90 upvotes yesterday

Preprint · not peer reviewedTrending

To train AI agents that use computers, the authors mine real workflows from instructional videos, turn them into many varied practice tasks, and retrain the model repeatedly on its weak spots.

Key finding: Their 27B model scores 80.2 on the OSWorld Verified benchmark and keeps improving across rounds.

Why it matters: Good practice data is the main bottleneck for computer-using agents; this offers a repeatable way to make more of it.

Read on arXiv

CoworkOS: Scaling Agent Coworkers for Real-world Automation

Jiajie Jin, Guanting Dong, Chengjun Pan, et al. · Submitted 2 Nov 2026 · 85 upvotes yesterday

Preprint · not peer reviewedTrending

A framework for long-running AI agent work where the model decides what to do and the runtime keeps track of files, decisions and logs, so humans and agents can edit the same documents.

Key finding: Used to automatically fine-tune another model over five rounds, it raised that model's average score on eight benchmarks from 33.19 to 42.95.

Why it matters: It is a concrete test of AI agents doing multi-day work, including improving other AI models.

Read on arXiv

Can Coding Agents Sustain Continual Evolution without Forgetting?

Shangziqi Zhao, Yitong Zhang, Jia Li · Submitted 2 Nov 2026 · 1 upvotes yesterday

Preprint · not peer reviewedCode releasedTrending

The authors let coding agents tune their own tooling on one task type after another and checked whether earlier skills survived.

Key finding: One setup reached 94.0% on SWE-Bench, then fell to 32.0% after later tuning on a different dataset, a 66% relative drop.

Why it matters: Self-improving agents can quietly lose old abilities; teams should re-test earlier tasks after each change.

Read on arXiv

IdeaTaste: Do LLMs Share the Research Taste of Human Researchers?

Shanshan Zhong, Chen Xu, Jingjie Ning, et al. · Submitted 2 Nov 2026 · 1 upvotes yesterday

Preprint · not peer reviewedTrending

Models were given research literature frozen at 2025 and asked to propose work; their proposals were compared with what the same fields actually published in 2026, across more than 90,000 proposals.

Key finding: AI reached about 26% of the human level at matching citations and only 18% at matching ideas.

Why it matters: It suggests AI 'scientists' can spot popular directions but struggle to pick the concrete approaches researchers actually pursue.

Read on arXiv

Your RightSignal

Topics in News

Wave Forcing: Towards Speed-of-Light Streaming Video Generation

CUA-Master: Scaling Computer-use Agents through Open-World Recursive Improvements

CoworkOS: Scaling Agent Coworkers for Real-world Automation

Can Coding Agents Sustain Continual Evolution without Forgetting?

IdeaTaste: Do LLMs Share the Research Taste of Human Researchers?

That’s every paper.