Terminal-RSI: recursive self-improvement lifts 27B agent from 51.7% to 86.5% on Terminal-Bench
yifeiwang77 · x · 2026-10-17
The Terminal-RSI paper (arXiv:2610.22192) studies a recursive self-improvement loop for terminal agents: rewriting verified trajectories via Recursive Self-Rewrite (RSR), then refining policies through critic-inheriting Recursive Policy-Refine (RPR). After two cycles, a 27B model improves Terminal-Bench 2.1 success from 51.7% to 86.5% and Terminal-Bench 4.0 from 0.0% to 19.8% (new SOTA).
- Terminal-Bench 3.0: 16.95%; Terminal-Bench Hard: 79.00%
- LHTB process reward rises 0.21→0.37; strict completions 0/46→5/46
Techniques include Turn-Boundary Skip-Observation GAE, Family Dense Reward, Success-Prioritized Reward, and a warm-up critic with recursive critic handoff. The authors ask how far such hill climbing can go and what makes it stall.
More from coding & agent
- Reflection AI Launches Mirror, an Open Coding Agent Optimized for Beam — sloppenheimer · 2026-10-17
- Google's Gemini 4 Argon 1M-Token Output Is Undersold for Legacy Modernization — rseroter · 2026-10-17
- RustGIS v0.3.0 Ships: Pure-Rust Open-Source GIS With Built-in AI Assistant — giswqs · 2026-10-17
- No accepted agent loop yet for Jev-style decision models plus LLM, Reddit thread finds — emersoftware · 2026-10-17
- AI influencer hands an agent full control of a $500 day out — and it's going "horribly wrong" — alliekmiller · 2026-10-17
- Microsoft Copilot Studio adds agent accounts, memory and triggers for persistent business-process agents — clamanna · 2026-10-17