EvoSyn framework improves agent task success by 13.3 points via agentic data synthesis
Which summary reads better? Pick one — models revealed after.Both summaries are AI-generated.
EvoSyn reports a 13.3-point average task-success gain over the strongest synthetic-data baseline on τ²-Bench and CAR-Bench by making the synthesis space itself editable from learner feedback. The practical shift is that agent training data generation stops being just “sample more failed cases” and becomes a closed-loop controller that can add new controllable interaction factors, which matters if your bottleneck is coverage of rare, task-specific failure modes rather than raw trajectory volume.
EvoSyn closes the loop on agent training data by converting learner failures directly into synthesis directives, yielding a 13.3-point average task-success gain over the strongest synthesis baseline on τ²-Bench and CAR-Bench. The key shift is that it not only targets weak regions within a fixed generation space but expands that space—surfacing controllable interaction factors that static synthesis pipelines never discover. If you're generating synthetic trajectories to fine-tune agents, this argues for a feedback-driven, revisable synthesis state rather than one-shot bulk generation, since your current pipeline is likely blind to coverage gaps it can't yet represent.