RightSignal · Monday, November 2

Cheap screening before expensive AI reasoning could make scaling more efficient

Qi Wang, Yun Qu, Lizhou Cai, Heming Zou, Haoang Chi, Yiqin Lv · Submitted 29 Oct 2026 · 44 upvotes yesterday

Preprint · not peer reviewedTrending

The authors pair a large, costly AI system with a small, fast predictor that first guesses which actions are worth trying. Only those that pass a threshold go to the big system, and the results teach the predictor.

Key finding: With ideal screening, the compute spent drops in proportion to how rare the useful actions are; prediction mistakes and exploration reduce the saving.

Why it matters: It argues that progress may come from spending compute more wisely, not only from more of it.

Read on arXiv

R^2Orch: an AI agent coordinator that trains itself by picking its own tasks

Mingda Zhang, Yun Huang, Qiang Huang, Yanjin Li, Rui Qian, Qichao Zhao · Submitted 22 Oct 2026 · 36 upvotes yesterday

Preprint · not peer reviewedCode releasedTrending

An agent that directs other tools and agents improves by setting itself tasks and practising on them. An outer loop removes each learned skill once tests show the agent does just as well without it.

Key finding: Across twelve datasets in question answering, maths, code and interactive tasks, it beat the baselines.

Why it matters: It reduces the need for hand-designed training plans for agent systems.

Read on arXiv

CaPtain trains a small open AI model to control robots by writing code

Yijia Fan, Xilin He, Yilei Jiang, Dongming Wu, Jiaxu Wang, Xiangyu Yue · Submitted 29 Oct 2026 · 15 upvotes yesterday

Preprint · not peer reviewedCode releasedTrending

The team fine-tuned a 4-billion-parameter open model on example runs, then trained it further with trial and error, so it writes code that moves a robot arm. A helper scores each step by how much it changes the chance of success.

Key finding: The model went from 0% to 73.8% success on the LIBERO robot test, and ran on a real robot after training only in simulation.

Why it matters: Small open models could do robot tasks that so far needed large closed models; code, data and weights are released.

Read on arXiv

RoboArchitect: an AI rewrites the control code around a robot instead of retraining it

Shengguang Wu, Suning Huang, Cherie Ho, Ria Doshi, Hao Zhu, Serena Yeung-Levy · Submitted 30 Oct 2026 · 4 upvotes yesterday

Preprint · not peer reviewedTrending

A coding model watches a robot's attempts, works out why it failed, and keeps revising the code that decides when to call the robot's model, use vision tools or move the arm directly. No model weights change.

Key finding: Average success on unseen scenes rose from 9.7% to 43.3%, and on unseen tasks from 12.7% to 40.2%.

Why it matters: Much of a robot's skill may sit in the surrounding code, which can be improved automatically and cheaply.

Read on arXiv

AI safety

HALO: experts find AI assistants most often fail by stepping in too early or too much

Wenqi Pei, Xinyu Wang, Jinrui Liu, Cynthia Zhuge, Jie Sheng, Zihan Wang · Submitted 29 Oct 2026 · 6 upvotes yesterday

Preprint · not peer reviewedTrending

The authors built 400 scenarios where an assistant must decide on its own whether and how much to help, and had 20 experts judge 8,000 answers from 10 models.

Key finding: The best closed model's answers were acceptable 72.88% of the time and the best open model's 56.63%; acting too early or too forcefully caused 61.20% of rejected answers.

Why it matters: As assistants start acting without being asked, knowing when to hold back matters as much as being able to help.

Read on arXiv

Your RightSignal

Topics in News

Cheap screening before expensive AI reasoning could make scaling more efficient

R^2Orch: an AI agent coordinator that trains itself by picking its own tasks

CaPtain trains a small open AI model to control robots by writing code

RoboArchitect: an AI rewrites the control code around a robot instead of retraining it

HALO: experts find AI assistants most often fail by stepping in too early or too much

That’s every paper.