Dynamic Scenario Synthesis Framework Improves LLM Agent Adaptability

01 ABSTRACT

To address the gap that most training environments remain static and fail to reflect external events that invalidate ongoing plans, the authors propose DynaWeave, a framework that uses directed acyclic graphs to represent event dependencies, activates events during agent interaction, and employs a dynamic verifier that revises only checks affected by events that actually occur. Using this framework, they collect 2,095 verified trajectories across 36 environments and perform supervised fine-tuning of Qwen3.5 models; experiments on Gaia2, CostBench, and ToolMaze show consistent overall performance improvements across all evaluated model sizes. The authors argue that dynamic scenario synthesis is a practical approach to training agents to adapt to changing environments, but the conclusions are based on specific models and benchmarks, and cross-environment generalization requires further validation.

02 KEY FINDINGS

1. DynaWeave automatically converts static tasks into verifiable dynamic scenarios involving requirement changes, environment state changes, and external information updates.
2. The framework uses directed acyclic graphs to manage exogenous event dependencies, and a dynamic verifier revises only evaluation checks affected by events that actually occur.
3. After fine-tuning Qwen3.5 on 2,095 verified trajectories collected from 36 environments, overall performance improves across all model sizes on three benchmarks.
4. Task completion rate on CostBench increases by 67.34 percentage points, and on ToolMaze it even exceeds the performance of Gemini-3.1-Pro-Preview.