From Self-Evolving Synthetic Data to Verifiable-Reward RL: Post-Training Multi-turn Interactive Tool-Using Agents

TL;DR

This paper introduces EigenData, a hierarchical multi-agent framework that synthesizes tool-using dialogues and employs verifier-based reinforcement learning to improve multi-turn interactive agents, achieving high performance without extensive human data.

Contribution

The paper presents a novel unified framework combining self-evolving synthetic data generation with verifier-based RL for training complex tool-using agents.

Findings

Achieved 73.0% pass rate on Airline and 98.3% on Telecom benchmarks.

Demonstrated scalable bootstrapping of tool-using behaviors without human annotation.

Improved generation reliability through closed-loop self-evolving prompts and workflows.

Abstract

Interactive tool-using agents must solve real-world tasks via multi-turn interaction with both humans and external environments, requiring dialogue state tracking, multi-step tool execution, while following complex instructions. Post-training such agents is challenging because synthesis for high-quality multi-turn tool-use data is difficult to scale, and reinforcement learning (RL) could face noisy signals caused by user simulation, leading to degraded training efficiency. We propose a unified framework that combines a self-evolving data agent with verifier-based RL. Our system, EigenData, is a hierarchical multi-agent engine that synthesizes tool-grounded dialogues together with executable per-instance checkers, and improves generation reliability via closed-loop self-evolving process that updates prompts and workflow. Building on the synthetic data, we develop an RL recipe that first fine-tunes the user model and then applies GRPO-style training with trajectory-level group-relative advantages and dynamic filtering, yielding consistent improvements beyond SFT. Evaluated on tau^2-bench, our best model reaches 73.0% pass^1 on Airline and 98.3% pass^1 on Telecom, matching or exceeding frontier models. Overall, our results suggest a scalable pathway for bootstrapping complex tool-using behaviors without expensive human annotation.