Qwen's Terminal-Universe paper: reconstructing agent training environments from trajectories
dair_ai · x · 2026-09-04
- New paper from the Qwen team on where agent training environments actually come from: terminal agent trajectories have accumulated at scale while realistic executable environments remain scarce.
- Key insight: environments are what post-training needs — each can be re-queried into many verifiable tasks with execution feedback, while a trajectory is a single frozen demonstration.
- Terminal-Universe reconstructs the environment from the trajectory instead of generating one from scratch: the tool-execution history exposes the environment's structure; replaying recorded file operations restores files to pre-modification state, and a completion agent fills in missing files and dependencies.
- Recovered workspaces are scaled along two axes (text truncated).
More from coding & agent
- GPT-6 Astra impresses dev by finding and patching bugs in upstream dependencies — heyneighbor · 2026-09-05
- Blender clay renders + video diffusion: a workflow idea for OpenAI's Astra agent — OdinLovis · 2026-09-05
- Agents coordinating via a shared message board as persistent state — novasarc01 · 2026-09-05
- Perplexity to detail SPACE, the Rust sandbox platform behind Perplexity Computer — AravSrinivas · 2026-09-05
- Open-Source Writeup on RL Training Knowledge-Work Agents at Scale Goes Viral — heghbalz · 2026-09-05
- Neon Fall: a CRT-style Tetris clone built with GPT few-shot prompting — intellectronica · 2026-09-05