WRIT: Write-Read Intensive Trajectory Synthesis for Multi-Turn User-Facing Agents

In three linesWRIT is a trajectory synthesis pipeline for multi-turn agent training. It generates complex tasks along two axes: number of write decisions and evidence burden per decision. With 2K synthesized trajectories, a 4B model outperforms GPT-5.1 no-think on τ²-bench while reducing inference-time token usage.

Summary generated by Claude — human-verified