@HuggingPapers: Recursive Self-Rewrite scales model capability on complex tasks A single base model (Qwen-3.8-27B) discovers solutions …
Summary
HuggingPapers 介绍了一篇关于 Recursive Self-Rewrite 的研究：单一基础模型 Qwen-3.8-27B 在多种 harness 下发现解法并将其改写为可复用的训练轨迹，使 Terminal-Bench 2 的 Pass@3 从 57.0% 提升到 74.2%。
View Cached Full Text
Cached at: 10/06/26, 03:49 PM
Recursive Self-Rewrite scales model capability on complex tasks
A single base model (Qwen-3.8-27B) discovers solutions across diverse harnesses, then rewrites them into reusable training trajectories under a general harness. Pass@3 on Terminal-Bench 2 jumps from 57.0% to 74.2%. https://t.co/lbtTF5P6YC
Similar Articles
Scaling Trajectories for Complex Tasks through Recursive Self-Rewrite
The paper proposes Recursive Self-Rewrite (RSR), a framework where a single base model (Qwen-3.8-27B) discovers successful solutions under diverse harnesses and reconstructs them into training trajectories under a general harness via a planner-critic-executor loop, yielding significant pass@3 gains on terminal-bench tasks after supervised finetuning.
Self-Supervised Scaling of Terminal Environments for Scientific Domains
提出软件在环重构（software-in-the-loop reconstruction）的自监督框架，从现有科学软件工作流中自动生成可执行参考行为与验证目标，用于训练终端智能体；使用 Qwen3.8-Max 生成训练数据并对 Qwen3.8-27B 微调，使 Terminal-Bench2 平均性能从 47.94% 提升至 53.56%。
Terminal-RSI: Recursive Self-Improving Terminal Agents via Recursive Self-Rewrite and Policy-Refine
Terminal-RSI 提出通过递归轨迹重写（RSR）与继承批评器的策略精炼（RPR）交替进行，使终端代理从部分进展推进到严格验证，并在 Terminal-Bench 多项基准上取得显著成绩，发布了 Terminal-RSI-27B 模型。
@Yucheng__Shi: We used an agent to help train a stronger successor. Then we used that successor to help train the next. Our new paper,…
Tencent Hy Frontier and academic collaborators introduce Terminal-RSI, a recursive self-improvement loop for terminal agents that alternates verified-trajectory rewriting with critic-inheriting policy refinement, lifting a 27B model's Terminal-Bench 2.1 success from 51.7% to 86.5% and reaching a new 19.8% SOTA on Terminal-Bench 4.0.
@omarsar0: Banger paper from Tencent on recursive self-improvement for terminal agents. (bookmark it) This is a great read if you …
腾讯等机构联合发表 Terminal-RSI 论文，通过交替进行轨迹重写（RSR）与批评家继承式策略优化（RPR），将基于 Qwen3.8-27B 的终端智能体在 Terminal-Bench 系列基准上的平均成绩从 20.68% 提升至 50.58%，并展示了奖励设计审计对区分严格成功与部分奖励的重要性。