Terminal-RSI boosts 27B model's terminal-agent success from 20.7% to 50.6% via recursive self-improvement
Dr_Singularity · x · 2026-11-09
- 腾讯 Hy Frontier 联合新加坡国立大学、马里兰大学等发布论文 Terminal-RSI,提出通过「递归自我改写(RSR)+ 策略精炼(RPR)」交替循环让终端 agent 递归自我改进：将验证通过的解题轨迹重写为更通用策略，再由继承 critic 的精炼阶段固化行为。
- 技术要点包括：Turn-Boundary Skip-Observation GAE、家族级密集奖励(FDR)校准语义进度、成功优先奖励(SPR)区分严格成功与部分得分、Warm-up Critic 与阶段性学习率协调两阶段训练。
- 基于 Qwen3.8-27B(应为 Qwen3-27B 系)的五阶段流程后，Terminal-RSI-27B 在 Terminal-Bench 4.0 达 19.84%、Terminal-Bench 3.0 达 16.95%、Terminal-Bench 2.1 达 86.52%、Terminal-Bench Hard 达 79.00%;四个基准平均成功率从 20.68% 升至 50.58%。长时程任务 LHTB 严格完成从 0/46 提升到 5/46,SWR100 成功率从 3% 翻倍到 6%。
More from coding & agent
- David Sacks: Salesforce made Claude its front end — SaaS is turning into an agent database — 2C_ornot2C · 2026-11-09
- Building agents is hill climbing: early results are noisy and gains may not generalize — andreisavu · 2026-11-09
- Evolver turns agent exploration into reusable know-how: +4.4pp success, -24% tokens — omarsar0 · 2026-11-09
- People are vibe coding synthesizers now — AI-built synths are a thing — billyjhowell · 2026-11-09
- Dev builds free browser tycoon game Mall Bats with Claude writing nearly all the code — jacobjr23 · 2026-11-09
- Open-Source Mac Status Light App Monitors Multiple Claude Code Sessions — stateofteddy · 2026-11-09