Tencent's Terminal-RSI lifts terminal agents from 20.7% to 50.6% on Terminal-Bench via recursive self-improvement
omarsar0 · x · 2026-10-18
Tencent Hybrid Frontier's Terminal-RSI paper trains terminal agents via recursive self-improvement with RL, alternating RSR (fine-tuning on rewritten verifier-passed trajectories) and RPR (PPO with a critic inherited from the previous stage) over five stages from Qwen3.8-27B.
- Four-suite Terminal-Bench average rises from 20.68% to 50.58%; 86.52% on Terminal-Bench 2.1, 79.00% on Terminal-Bench Hard; LHTB strict completion goes from 0/46 to 5/46.
- Reward audit: under dense progress reward, failed runs collected 50.09% of the reward, and 48 failures passed every executed check but missed the final test; separating strict success from partial credit cut that share to 28.72%.
More from coding & agent
- Grok Build v1.0.63 adds multi-agent visibility, MEMORY.md index and shell-free structured hooks — XFreeze · 2026-10-18
- OpenCode confirms breach, stolen customer API keys already incurring charges — jedisct1 · 2026-10-18
- AnyPS5 runs PS5 games on PC without an emulator, 27.8k GitHub stars — Vjeux · 2026-10-18
- JevAgentHarness: 4B Decision Model Cuts Agent Cost 14.2% With Zero Task Failures — dair_ai · 2026-10-18
- Free model Space Bunny Max proposes new best bound for integer multiplication — IgorCarron · 2026-10-18
- Ex-Citadel COO analysis: waiting for GPT-7 to agentify your company is too late — ChanceKelch · 2026-10-18