Tsinghua + Qwen paper: rebuilding agent workspaces beats imitating trajectories, lifts Terminal-Bench to 58.1%
rohanpaul_ai · x · 2026-09-10
A new Tsinghua + Qwen team paper flips the usual agent-data recipe: instead of imitating recorded trajectories, reconstruct the actual code workspace so you can inject new bugs, new features, or stronger agents and generate many fresh verified training tasks.
- Trajectory = one frozen solution; environment = a reusable codebase producing many tasks.
- Training Qwen3.5-27B on this data moved Terminal-Bench 2.1 from 46.2% to 58.1% and EvoCode-Bench v2 MT@4 from 6.3 to 20.1 (Terminus2 setup).
- Takeaway: collect reusable execution environments, not more traces to imitate.
More from coding & agent
- macOS tip: store secrets and tokens in Apple Keychain instead of env files — HankYeomans · 2026-09-10
- Dev ships six agents inside a self-hosted CRM instead of a chat window, shares lessons — PleasantBid4665 · 2026-09-10
- The Software Factory Stack: Open, Composable, Defined-in-Code — vikvang1 · 2026-09-10
- Browserbase: browser agents that write code beat pixel-clicking CUA models — adnan_hashmi · 2026-09-10
- Codex CLI 0.154.0 ships GPT-6-Astra, experimental worktree support — github-actions[bot] · 2026-09-10
- Gradium TTS lands on LiveKit Inference with voice cloning, sub-250ms latency, free until Oct 9 — alexcovo_eth · 2026-09-10