Terminal-Universe turns agent traces into training worlds
Terminal-Universe reconstructs executable workspaces from terminal-agent trajectories, then expands them into verifiable single- and multi-round training tasks. Its 37.3K-environment corpus improved Qwen3.5-27B by 11.9 points on Terminal-Bench 2.1 and 13.8 points on EvoCode-Bench v2 MT@4.
Terminal-Universe makes a compelling case that executable environments—not just more trajectories—are the missing ingredient in coding-agent post-training.
- –Replays file operations to recover workspaces, then uses agentic completion to restore missing files and dependencies.
- –Converts one frozen trajectory into many tasks through intent recovery, cross-workspace queries, and simulated multi-round feedback.
- –Reported gains suggest environment diversity can matter more than simply expanding imitation data.
- –The approach could create a self-reinforcing data flywheel as stronger agents generate richer, more reusable workspaces.
- –The results remain primarily a research demonstration, with teacher-generated tasks and verification creating potential coverage and evaluation blind spots.
DISCOVERED
1h ago
2026-09-06
PUBLISHED
1h ago
2026-09-06
RELEVANCE
AUTHOR
Discover AI