Tencent's Terminal-RSI lifts a 27B model from 51.7% to 86.5% on terminal agent tasks
GabGarrett · x · 2026-10-19
Tencent Hunan's open-source Terminal-RSI paper shows recursive self-improvement (RSI) can steadily boost terminal agents built on Qwen3.8-27B.
Method: Alternates Recursive Self-Rewrite (RSR) of verified trajectories with critic-inheriting Recursive Policy-Refine (RPR) across five stages (Base→RSR-1→RPR-1→RSR-2→RPR-2), where each stage's model generates training data for the next. Supporting techniques include TB-GAE turn-boundary credit assignment, Family Dense Reward, and Success-Prioritized Reward to convert partial progress into strict task completion.
Results:
- Terminal-Bench 2.1: 51.70% → 86.52% (+34.82pp)
- Terminal-Bench 4.0: 0% → 19.84%
- Terminal-Bench Hard: 79.00%
- Long-horizon LHTB process reward 0.21→0.37, strict completions 0/46→5/46; SWR100 success doubled to 6.00%
The takeaway: a 27B model with RSI can match or beat larger models on terminal agent benchmarks.
More from coding & agent
- MyGo ships: build tiny Go desktop apps with native GPU-drawn UI or web frontend — dotey · 2026-10-19
- MOSS-Image ecosystem roundup: Spectrum speeds up to 12 steps on an RTX 3060 — optimisticalish · 2026-10-19
- One character sheet to a 44s AI-made promo: full workflow with Claude, ElevenLabs and HeyGen — azed_ai · 2026-10-19
- How to Build Your Own Algorithmic Trading Analyzer with Claude Code in 5 Steps — mdancho84 · 2026-10-19
- Open-source MCP for verifying AI outputs missed a 9,000 mg/day medication error — Practical_Green5775 · 2026-10-19
- Team launches India's first equity research MCP for Claude covering 5,600+ NSE/BSE stocks — viewmind · 2026-10-19