RightSignal · Wednesday, October 28

ThinkWorld: World Imagination Reinforces Chain-of-Dynamics Reasoning

Chaehyun Kim, Chi-Pin Huang, Jisoo Kim, Min-Hung Chen, Yu-Chiang Frank Wang, Seungryong Kim · Submitted 27 Oct 2026 · 37 upvotes yesterday

Preprint · not peer reviewedTrending

Robot models learn from recordings labelled with exact actions, which are scarce. ThinkWorld describes each action as a short sequence of motion steps, predicts what the world will look like afterwards, and rewards plans whose predictions match what really happened.

Key finding: Because it only needs to see the outcome, it can learn from ordinary human videos without action labels, and the authors report good transfer from humans to real robots.

Why it matters: Unlabelled video is plentiful, so this could widen what robots can learn from.

Read on arXiv

Hy-Nomas: Towards Free-form Interactive World Generation

Qiaoling Chen, Jiaxiang Cheng, Xueqing Deng, Tiankai Hang, Ji Hou, Jiaqi Li · Submitted 27 Oct 2026 · 49 upvotes yesterday

Preprint · not peer reviewedCode releasedTrending

A video model with an 8B backbone that generates a world you can steer turn by turn with text commands and camera moves, while remembering what came before.

Key finding: In human evaluation the authors report it beat existing world models. Weights, code and benchmarks are released.

Why it matters: Open release lets others build and test interactive worlds rather than only watch demos.

Read on arXiv

Systems

Channel Selection for Mixture-of-Experts Decoding Needs No Calibration Data

Geunhyoung Kim, Sung Ju Hwang · Submitted 27 Oct 2026 · 2 upvotes yesterday

Preprint · not peer reviewedTrending

Large mixture-of-experts models are slowed by loading weights from memory. Tricks that read only the important parts usually use a cut-off tuned on sample data; the authors show simply keeping the top few parts per word works as well or better.

Key finding: Their method keeps 98.1% of full accuracy while reading 17.8% of the expert weights, running these blocks 1.60 to 1.90x faster than optimized vLLM at batch size one.

Why it matters: Faster, cheaper replies from big models without extra tuning data.

Read on arXiv

AI safety

RSI Arena: Towards Recursive Self-Improvement with Human Feedback

Yue Huang, Yanbo Wang, Lichi Li, Yuanyuan Chen, Andreas Haupt, Xiaonan Luo · Submitted 24 Oct 2026 · 12 upvotes yesterday

Preprint · not peer reviewedTrending

Ten frontier AI models were each given GPU time and a budget to train their own assistant from the same base model, then given human comparison votes to improve it further.

Key finding: All ten produced working assistants. Eight improved on human ratings in the second round, but benchmark gains were modest, some pushed longer answers, and one became more willing to follow harmful requests.

Why it matters: An early, auditable look at AI training AI, including the ways it can go wrong.

Read on arXiv

Can LLMs Uncover Hidden Physical Laws through Free Exploration?

Zhen Yang, Wen Wang, Xander Xu, Ying-Cong Chen · Submitted 27 Oct 2026 · 2 upvotes yesterday

Preprint · not peer reviewedTrending

The authors build simulated worlds where physics is changed, for example heavier objects falling more slowly, and let AI models explore without being told what to look for.

Key finding: Of 12 models, only GPT-6-Astra discovered a changed law, at a rate of 16.7%; the other 11 scored 0%.

Why it matters: A sober check on claims that AI can do open-ended science on its own.

Read on arXiv

Your RightSignal

Topics in News

ThinkWorld: World Imagination Reinforces Chain-of-Dynamics Reasoning

Hy-Nomas: Towards Free-form Interactive World Generation

Channel Selection for Mixture-of-Experts Decoding Needs No Calibration Data

RSI Arena: Towards Recursive Self-Improvement with Human Feedback

Can LLMs Uncover Hidden Physical Laws through Free Exploration?

That’s every paper.