Tencent's Terminal-RSI builds recursively self-improving terminal agents, hitting 19.84% on Terminal-Bench 4.0
tencent · hf · 2026-10-16
Tencent Hy Frontier and universities present Terminal-RSI, a recursive self-improvement framework for terminal agents alternating Recursive Self-Rewrite of verified trajectories with critic-inheriting Recursive Policy-Refine. Techniques include TB-GAE, Family Dense Reward, Success-Prioritized Reward, warm-up critic and stage-specific learning rates. From a Qwen3-8-27B base through five stages, Terminal-RSI-27B scores 19.84% on Terminal-Bench 4.0, 86.52% on 2.1, 16.95% on 3.0 and 79.00% on Hard; on LHTB process reward rises 0.21→0.37 and strict completion 0/46→5/46.
More from coding & agent
- SpecAgent: small model drafts tool calls, large model verifies in parallel to cut latency — Zilin Wang · 2026-10-16
- Salesforce's CLIFT turns sparse judge feedback into rewards for open-source web agents — Salesforce · 2026-10-16
- Agent spots invalid SPF record and autonomously fixes it in Vercel DNS via browser — dkundel · 2026-10-16
- MCPJam engineer asks how to test if ChatGPT plugins actually see latest UI selection — Paoli99 · 2026-10-16
- ChatGPT revives unplayable 2000s game Bookworm on Mac and iPhone in 15 minutes — Shleemy_Pants · 2026-10-16
- Agentflow: open-source collective memory network for agents via FastMCP and pgvector — agentflow0328 · 2026-10-16