RightSignal · Thursday, October 22

Open-source models build practice worlds for training computer-use agents

Zhaoyang Wang, Qianhui Wu, Haorui Wang, et al. · Submitted 2026-10-19 · 45 upvotes yesterday

Preprint · not peer reviewedCode releasedTrending

The authors built a pipeline that generates working practice environments where agents can click through screens, call tools and run commands. Automatic checks make sure each task can be solved, is hard enough and is graded reliably.

Key finding: It produced 341 long tasks, and GPT-6-Astra solved only 60% of them. Two Qwen models trained only in these worlds improved on 8 benchmarks they had not seen.

Why it matters: Agents can be trained safely on hard tasks without touching real systems or relying on closed models.

Read on arXiv

Systems

AI agents can trim their own working memory without re-reading the history

Tianyu Fu, Ruoxi Wang, Zihan Min, et al. · Submitted 2026-10-20 · 34 upvotes yesterday

Preprint · not peer reviewedCode releasedTrending

Long agent sessions are costly because the model's cached memory of earlier text keeps growing. Summarizing the history forces the model to re-read it. This paper lets agents drop selected cached parts directly, and pairs that with a serving engine called AKV that reuses the cache correctly.

Key finding: A policy that drops old tool responses raised benchmark scores by 1.7 points and cut cost by 27.7% compared with summarization.

Why it matters: It points to cheaper long-running agents without losing accuracy.

Read on arXiv

Extra training lifts a small robot planner close to a much larger model

Xinxin Liu, Bingxin Xu, Wenrui Bao, et al. · Submitted 2026-10-20 · 16 upvotes yesterday

Preprint · not peer reviewedTrending

The authors applied the post-training methods used for coding agents to robot planners: supervised fine-tuning followed by multi-turn reinforcement learning. Adapters let one training pipeline work across different robot control setups.

Key finding: On LIBERO-PRO, a 9-billion-parameter planner's success rate rose from 18.1% to 71.0% without external memory, close to the 72.1% that Codex reached with memory.

Why it matters: Robot planners may not need large closed models if smaller ones are trained this way. The authors say they will release code later.

Read on arXiv

Writing the missing reasoning steps helps a student model learn from a teacher

Yan Wang · Submitted 2026-10-21 · 11 upvotes yesterday

Preprint · not peer reviewedTrending

Some AI providers return only answers, not the reasoning behind them. TraceForge trains a generator to write a plausible reasoning trace for a teacher's answer, in the style of the student model's own family, and a critic to judge how real a trace looks.

Key finding: Using the same 5,000 questions, the generated traces improved coding scores by 3.7 points on HumanEval+ and 7.1 on MBPP+ over the teacher's own summarized thinking.

Why it matters: It offers a way to learn from answer-only models without breaking providers' terms of service.

Read on arXiv

Riddles from 62 cultures show where AI models still fail

My Chiffon Nguyen, Tianyi Hu, Vinod Anbalagan, et al. · Submitted 2026-10-21 · 1 upvotes yesterday

Preprint · not peer reviewedTrending

Communities wrote original riddles rooted in their own cultures, covering 45 countries and 51 languages. The authors tested 28 model configurations with both automatic and human evaluation.

Key finding: Even frontier models reached at most 22% accuracy for some communities.

Why it matters: Most tests are centred on English. This one shows how unevenly models understand different cultures.

Read on arXiv

Your RightSignal

Topics in News

Open-source models build practice worlds for training computer-use agents

AI agents can trim their own working memory without re-reading the history

Extra training lifts a small robot planner close to a much larger model

Writing the missing reasoning steps helps a student model learn from a teacher

Riddles from 62 cultures show where AI models still fail

That’s every paper.