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

2026-06-01 · Source: Artificial Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, quick

Summary

WRIT (Write-Read Intensive Trajectory Synthesis) is a novel pipeline designed to generate complex training trajectories for multi-turn user-facing agents. It addresses a gap where existing methods primarily focus on "write-intensive" tasks, which train sequential execution, but neglect the "read-heavy" challenge of gathering and comparing substantial evidence before making a single decision. WRIT synthesizes tasks along two complexity axes: the number of write decisions and the evidence burden per decision. The pipeline generates both write-intensive and read-heavy tasks, diversifies user behavior instructions, and simulates agent-user interactions in an executable environment. Training a 4B model on only 2K WRIT-synthesized trajectories resulted in performance exceeding GPT-5.1 no-think on τ²-bench, significantly reducing inference-time token usage. This demonstrates that compact Supervised Fine-Tuning (SFT) data can convert expensive test-time reasoning into efficient agent behavior.

Key takeaway

For Machine Learning Engineers developing multi-turn user-facing agents, consider integrating WRIT's trajectory synthesis approach. Your training data should explicitly incorporate both "write-intensive" sequential tasks and "read-heavy" evidence-gathering scenarios to build more robust and efficient agents. This method can significantly improve agent performance on benchmarks like τ²-bench and reduce inference-time token usage, converting expensive reasoning into efficient behavior.

Key insights

WRIT synthesizes agent training trajectories balancing sequential execution and evidence-gathering complexity for robust decision-making.

Principles

• Single write decisions can be difficult due to evidence burden.
• Training data should reflect realistic conversational variation.
• Compact SFT data can convert test-time reasoning into efficient behavior.

Method

WRIT generates write-intensive and read-heavy tasks, diversifies user behavior instructions, then simulates agent-user interactions in an executable environment.

In practice

• Train agents for longer task execution.
• Develop agents for robust, evidence-grounded decision making.
• Reduce inference-time token usage with targeted SFT data.

Topics

• Multi-turn Agents
• Trajectory Synthesis
• Agent Training
• Evidence-Grounded Decision Making
• Large Language Models
• Supervised Fine-Tuning

Best for: AI Engineer, Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• Skills on the Fly: Test-Time Adaptive Skill Synthesis for LLM Agents — Dynamically synthesizing task-specific skills from retrieved trajectories improves LLM agent performance without parameter updates.
• Gyms for Them, Mirrors for Us — Prioritize read-only "mirror" AIs for human feedback and controlled "gym" environments for model training over risky write-enabled "butler" agents.
• ACC: Compiling Agent Trajectories for Long-Context Training — Agent Context Compilation (ACC) transforms multi-turn agent trajectories into long-context QA pairs, directly supervising long-range reasoning.
• Active Exploring like a Pigeon: Reinforcing Spatial Reasoning via Agentic Vision-Language Models — An agentic VLM pipeline uses dynamic cognitive maps and Spatial Assertion Codes for robust spatial reasoning.
• Same Evidence, Different Answers: Canonical-Context On-Policy Distillation for Multi-Turn Language Models — Aligning multi-turn LLM behavior with full-context teacher responses mitigates self-anchored drift.
• When Does Memory Help Multi-Trajectory Inference for Tool-Use LLM Agents? — The effectiveness of memory methods in tool-use LLM agents is highly dependent on the chosen multi-trajectory inference strategy.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.