Title: WRIT: Synthesizing Write-Read Intensive Trajectories for Multi-Turn User-Facing Agents

Abstract:

Multi-turn agents designed for user interaction are tasked with deducing user intent from partial prompts, acquiring missing details via dialogue and tool usage, and performing valid actions. These processes are captured in training trajectories, which document the interleaved sequence of user inputs, agent replies, and tool invocations. While synthesizing complex trajectories is now a primary method for training agents, current pipelines typically escalate difficulty by chaining multiple user requests into extended tasks. This approach yields write-intensive trajectories that primarily train for sequential execution. However, we contend that individual write decisions can be inherently difficult when an agent must accumulate and evaluate significant evidence from read-tools before its arguments can be determined—a nuance that write-heavy data alone fails to capture.

Building on this premise, we introduce WRIT (\uline{W}rite-\uline{R}ead \uline{I}ntensive \uline{T}rajectory Synthesis), a framework designed to generate multi-turn agent training trajectories across two dimensions of complexity: the volume of write decisions within a task and the evidentiary load required for each decision. The WRIT pipeline begins by creating tasks that are both write-intensive and read-heavy. It subsequently enhances diversity by varying user behavioral instructions to mimic realistic conversational shifts, and finally, it simulates agent-user interactions within an executable environment to yield full training trajectories.

This resulting dataset equips agents to handle not only extended task execution but also robust, evidence-based decision-making under heavy information loads. In experiments, a 4B model trained on just 2,000 WRIT-synthesized trajectories surpassed GPT-5.1 no-think on the $\tau^2$-bench. Furthermore, it significantly lowered inference-time token consumption, demonstrating that compact supervised fine-tuning (SFT) data can effectively replace costly test-time reasoning with efficient agent performance.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC