Tencent's Terminal-RSI lifts agent success from 20.68% to 50.58% via recursive self-improvement
mark_k · x · 2026-11-09
Tencent Hy released Terminal-RSI, a concrete recipe for recursive self-improvement: a Qwen3-30B-based agent tackles real terminal tasks, verified trajectories are rewritten into reproducible demonstrations, then refined with reinforcement learning, and the improved agent generates the next round of training data.
The framework alternates Recursive Self-Rewrite (RSR) with Policy-Refine (RPR), plus TB-GAE, Family Dense Reward (FDR), Success-Prioritized Reward (SPR), and a warm-up curriculum model. After five stages:
- Average success across four terminal benchmarks: 20.68% → 50.58%
- Terminal-Bench 4.0: 0 → 19.84%
- Terminal-Bench Hard: 0.21 → 0.37; strictly completed LHTB tasks 0/46 → 5/46
- Terminal-Bench 2.1: 86.52%
The open question is whether a better agent can keep discovering better training examples on subsequent passes.
More from coding & agent
- The fence metaphor for agent trust: draw boundaries, then let agents roam free inside — manosaie · 2026-11-09
- Claude Code project structure explained: when every config file loads — alex_verem · 2026-11-09
- AWS case study: DSPy RLM mines years of Slack history into an AI IT help desk knowledge base — lateinteraction · 2026-11-09
- McKinsey: multi-step agents burn 5-30x more tokens than chat — ingliguori · 2026-11-09
- Apple ships official Xcode plugin for Claude Code, plus an AI playbook to find and build iOS apps — udmrzn · 2026-11-09
- Deaf developer builds Signal Dog, an AI agent plugin that hears door knocks and alarms — Scobleizer · 2026-11-09