Ant International and Westlake University release Trajectory2Gym for tool-use agent training
ant-intl · hf · 2026-10-27
Ant International and Westlake University released Trajectory2Gym, addressing a core problem in tool-use agent training: successful trajectories aren't necessarily useful supervision — they may underrepresent multi-step dependencies or include actions unjustified by information visible at decision time.
The paper identifies two coupled requirements for effective trajectory supervision:
- Compositional coverage: important observation-action pairs must all be represented
- Decision grounding: each supervised action must be supported by the learner-visible interaction history
Trajectory2Gym constructs executable interactions around under-supported compositional structures and derives supervision only from grounded decisions in realized executions, using structural-support selection to allocate experience and parent-normalized supervised fine-tuning with semantic interactions as training-weight units. This reframes trajectory generation as deliberate construction of learnable experience rather than collection of successful trajectories. Across five Qwen3 scales, it improves results on Tau1. Code and data are open-sourced.
More from coding & agent
- Polylane builds its agent memory system on Git, with nightly self-correction — dinasaur_404 · 2026-10-27
- Manus on rebuilding itself as an AI-native org: expose interfaces, demand evidence, go async — dotey · 2026-10-27
- RL doesn't just reward thinking longer—it incentives parallel agent swarms — rosstaylor90 · 2026-10-27
- A peek at a personal multi-agent setup: Codex, Muse and Grokbot with role-based memory — op7418 · 2026-10-27
- Context Layer V2 evals: ~40% accuracy gain, ~30% latency cut vs MCP alone — shensi · 2026-10-27
- Agent made 133 calls in 41 seconds: engineer argues physics, not protocols, is the bottleneck — shensi · 2026-10-27