RightSignal · Thursday, November 5

FrontierPhysics: a test of whether AI agents can do real physics research from start to finish

Bingran You, Wenjun Ke, Ria Rosenauer, Bo Liu, Wenbo Chen, Min Chen · Submitted 4 Nov 2026 · 27 upvotes yesterday

Preprint · not peer reviewedTrending

42 physicists built 56 tasks from their own research. Agents must read the literature, plan, run experiments and write a paper, which is checked by automatic tests and expert rubrics.

Key finding: Across 22 models in 26 setups, none passed more than 46 of its 168 attempts. The main weakness was loose reasoning and limited expert intuition, not missing physics knowledge.

Why it matters: It gives a sober measure of how far AI 'research agents' really are from doing publishable science.

Read on arXiv

Robot control models do better when they skip the separate image encoder

Weichen Fan, Haiwen Diao, Penghao Wu, Siyuan Hu, Ziqi Huang, Tianqi Liu · Submitted 4 Nov 2026 · 51 upvotes yesterday

Preprint · not peer reviewedTrending

Most robot models first compress camera images with a separate vision encoder. NEO-VLA instead feeds lightweight image patches straight into the main model alongside the instructions.

Key finding: Under controlled comparisons this design beat encoder-based versions, and the signals most useful for actions often peaked in earlier layers, not the final one.

Why it matters: It suggests fine details like edges and depth get lost when images are compressed too early, which matters for robots handling objects.

Read on arXiv

An AI coding agent makes robot training data instead of a human operator

Kewei Zhang, Juncheng Ma, Guodong Lv, Duomin Wang, Feixiang Ruan, Ruiyang Si · Submitted 3 Nov 2026 · 20 upvotes yesterday

Preprint · not peer reviewedTrending

From one real recording, a coding agent rebuilds the task in simulation, writes a controller and a success check, and a video model renders realistic camera footage.

Key finding: On five real-robot tasks, policies trained only on this data scored 93 vs. 84 successes in 100 trials, and held up better under new lighting (87 vs. 46).

Why it matters: Collecting robot demonstrations by hand is slow and costly; this points to a cheaper way to get training data.

Read on arXiv

Training AI agents with reinforcement learning mostly helps them recover from mistakes

Kejian Zhu, Zhuoran Jin, Shangqing Tu, Tianyi Men, Yupu Hao, Jiachun Li · Submitted 4 Nov 2026 · 25 upvotes yesterday

Preprint · not peer reviewedTrending

The authors measured agents across different numbers of tries and allowed steps to see where reinforcement learning actually helps.

Key finding: The gains shrink sharply when agents get fewer turns, and come from error recovery rather than better single decisions. Varying the turn budget in training (BudgetMix RL) made results steadier.

Why it matters: Agents trained with generous step limits may disappoint when deployed with tight ones.

Read on arXiv

Systems

SpecOPD trains the small 'drafting' models that speed up AI text generation faster

Tianyu Liu, Jun Zhang, Wenjie Huang, Yuhao Shen, Tianqi Xu, Xinyue Hao · Submitted 2 Nov 2026 · 7 upvotes yesterday

Preprint · not peer reviewedTrending

Speculative decoding uses a small model to guess several words that a big model then checks. SpecOPD trains that small model on many independent branches in parallel instead of one long sequence.

Key finding: It reports up to 41.45% more accepted guesses, about 16 times faster training supervision, and up to 3.56 times faster generation than normal word-by-word decoding.

Why it matters: Faster drafters mean cheaper and quicker AI answers without changing the big model.

Read on arXiv

Your RightSignal

Topics in News

FrontierPhysics: a test of whether AI agents can do real physics research from start to finish

Robot control models do better when they skip the separate image encoder

An AI coding agent makes robot training data instead of a human operator

Training AI agents with reinforcement learning mostly helps them recover from mistakes

SpecOPD trains the small 'drafting' models that speed up AI text generation faster

That’s every paper.