腾讯论文:终端智能体递归自我改进与RL训练
Banger paper from Tencent on recursive self-improvement for terminal agents.
腾讯在终端智能体 RL 训练上的硬核实践,给出了完整的五阶段交替训练流程与详细 Benchmark 数据,对做 Coding Agent 的同学极具参考价值。
Banger paper from Tencent on recursive self-improvement for terminal agents.
腾讯关于终端代理递归自我改进的论文。
(bookmark it)
(收藏它)
This is a great read if you are tracking research on training terminal or coding agents with RL
如果你正在关注使用强化学习训练终端或代码代理的研究,这是一篇值得阅读的好文章。
Starting from Qwen3.8-27B, they alternate two phases over five stages. One phase fine-tunes on rewritten versions of trajectories that passed the verifier. The other runs PPO with a critic inherited from the previous stage.
从 Qwen3.8-27B 开始,他们在五个阶段中交替进行两个阶段。一个阶段在通过验证器的轨迹重写版本上进行微调。另一个阶段运行 PPO,并使用从上一阶段继承的评论家模型。
The four-suite Terminal-Bench average goes from 20.68% to 50.58%, with 86.52% on Terminal-Bench 2.1 and 79.00% on Terminal-Bench Hard.
Terminal-Bench 四套件的平均分从 20.68% 提升到 50.58%,其中 Terminal-Bench 2.1 达到 86.52%,Terminal-Bench Hard 达到 79.00%。
The reward audit is just as useful. Under a dense progress reward, failed runs collected 50.09% of the reward, and 48 failures passed every executed check but still missed the final test. Separating strict success from partial credit cut that share to 28.72%.
奖励审计同样有用。在密集进度奖励下,失败的运行获得了 50.09% 的奖励,且有 48 次失败通过了所有执行的检查,但仍未能通过最终测试。将严格成功与部分得分区分开来后,该比例降至 28.72%。
https://arxiv.org/abs/2610.22192
Chat with Paper: https://academy.dair.ai/papers/terminal-rsi-recursive-self-improving-terminal-agents-via-recursive-self-rewrite-2610.22192
与论文对话:https://academy.dair.ai/papers/terminal-rsi-recursive-self-improving-terminal-agents-via-recursive-self-rewrite-2610.22192
更进一步:量化金融体系
看懂新闻只是起点——沿量化金融路径,把它变成能交付的工程能力