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
- Harness bug masked Gemini 4 Argon: rerun with Antigravity beats Astra, nears Opus 5.5 — vitaliychiley · 2026-10-27
- Design Is Becoming Game Dev: Build Your Own Tools With AI as the Executor — hewarsaber · 2026-10-27
- Dev turns DJI Mavic into local AI security drone built with Opus 5.5 — alexcovo_eth · 2026-10-27
- Microsoft Research: ForkPilot cuts long-horizon agent compute by up to 59% — EliasEskin · 2026-10-27
- Cursor launches Agent SDK: agents as code with Markdown prompts and TypeScript tools — soleio · 2026-10-27
- Vibecheck, the 'usable formal methods' hackathon, sells out two floors in SF — spikedoanz · 2026-10-27