腾讯论文：用递归自我改进训练终端智能体，Terminal-Bench 从 20.68% 升至 50.58%
腾讯发了篇终端智能体训练论文，五阶段交替微调加 PPO，Terminal-Bench 成绩直接翻倍多，奖励审计那段很值得细看。
腾讯发布论文（arXiv 2610.22192），以 Qwen3.8-27B 为起点，在五个阶段中交替执行两个环节：对通过验证器的轨迹改写后微调，以及用上一阶段继承的 critic 运行 PPO。四套 Terminal-Bench 平均成绩从 20.68% 提升到 50.58%，其中 Terminal-Bench 2.1 达 86.52%，Terminal-Bench Hard 达 79.00%。奖励审计显示，在使用稠密进度奖励时，失败运行仍拿到 50.09% 的奖励；将严格成功与部分得分分离后，这一比例降至 28.72%。
Banger paper from Tencent on recursive self-improvement for terminal agents. (bookmark it) This is a great read if you are tracking research on training terminal or coding agents with RL 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. 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. 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%. arxiv.org/abs/2610.22192 Chat with Paper: academy.dair.ai/papers/termina… 💬 10 🔄 1 ❤️ 16 👀 1309 📊 11 ⚡
