More successful trajectories do not necessarily make better training data: an action can succeed while relying on information the agent never saw, and common workflows can leave important multi-step dependencies underrepresented. The key insight is to construct interactions that expose those dependencies, then supervise only decisions grounded in the agent’s visible history.
Key Findings
- The method targets under-supported sequences of actions and observations, rather than sampling trajectories uniformly. This gives rare but important structures a deliberate path into the training set.
- Native environment execution supplies authentic observations, state changes, and errors. Actions dependent on collector-private information are excluded from imitation targets, while useful error context remains available for learning recovery.
- Across five Qwen3 scales, results improve on Tau1, Tau2, and BFCL, with the equal-weight three-benchmark mean rising by 16.9–29.8 percentage points. Gains are strongest in stateful and multi-turn interactions, where later actions depend on earlier feedback.
- Training weights semantic interactions as units, helping prevent trajectories with many actions from dominating simply because they are longer.
Method in Context
The approach treats trajectory generation as a way to design learnable experience, not merely to collect successful demonstrations. Each interaction is built around a target sequence of capabilities, executed in a native environment, and certified against the realized log before its grounded decisions contribute to training. This connects coverage of compositional behavior with evidence that the agent could actually have made each supervised decision.
Who It’s For—and the Trade-Off
A strong fit for researchers training tool-use agents that must handle multi-step, stateful workflows, especially when ordinary demonstrations underrepresent recovery or observation-dependent actions. The evidence is specific to the tested Qwen3 scales, benchmarks, and complete training recipe; it does not establish that every synthesis choice or scale transfers unchanged to other models and environments.