Saved in:
Bibliographic Details
Main Authors: Wang, Chenpeng, Cheng, Xiaojie, Wang, Chunye, Yang, Linfeng, Zhang, Lei
Format: Preprint
Published: 2025
Subjects:
Online Access:https://arxiv.org/abs/2510.06825
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866915540765769728
author Wang, Chenpeng
Cheng, Xiaojie
Wang, Chunye
Yang, Linfeng
Zhang, Lei
author_facet Wang, Chenpeng
Cheng, Xiaojie
Wang, Chunye
Yang, Linfeng
Zhang, Lei
contents Tool-augmented language models have demonstrated strong capabilities, but their reliance on live API access creates scalability and reliability challenges during training and deployment. We propose MTR, a simulation-first training framework for tool-augmented reasoning. Instead of relying on live APIs, MTR learns from complete ReAct traces with schema-validated, simulated observations. Our approach operates through a multi-agent architecture where a ToolMaker generates task-specific, OpenAI-compatible tool interfaces, an AutoAgent produces structured think-act-observe sequences, and a ToolActor simulates realistic responses. Training proceeds in two stages: Stage-1 Supervised Fine-Tuning (SFT) teaches 'trace grammar' from complete reasoning sequences; Stage-2 Group Relative Policy Optimization (GRPO) optimizes strategy with a composite trace reward that balances answer correctness and internal consistency. Across four multi-hop QA benchmarks (HotpotQA, MuSiQue, 2WikiMultiHopQA, Bamboogle), MTR attains competitive Exact Match (EM) scores to live-API systems and excels on reasoning-intensive tasks, suggesting that effective tool reasoning can be learned from structured traces without live interactions.
format Preprint
id arxiv_https___arxiv_org_abs_2510_06825
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Adaptive Tool Generation with Models as Tools and Reinforcement Learning
Wang, Chenpeng
Cheng, Xiaojie
Wang, Chunye
Yang, Linfeng
Zhang, Lei
Computation and Language
Tool-augmented language models have demonstrated strong capabilities, but their reliance on live API access creates scalability and reliability challenges during training and deployment. We propose MTR, a simulation-first training framework for tool-augmented reasoning. Instead of relying on live APIs, MTR learns from complete ReAct traces with schema-validated, simulated observations. Our approach operates through a multi-agent architecture where a ToolMaker generates task-specific, OpenAI-compatible tool interfaces, an AutoAgent produces structured think-act-observe sequences, and a ToolActor simulates realistic responses. Training proceeds in two stages: Stage-1 Supervised Fine-Tuning (SFT) teaches 'trace grammar' from complete reasoning sequences; Stage-2 Group Relative Policy Optimization (GRPO) optimizes strategy with a composite trace reward that balances answer correctness and internal consistency. Across four multi-hop QA benchmarks (HotpotQA, MuSiQue, 2WikiMultiHopQA, Bamboogle), MTR attains competitive Exact Match (EM) scores to live-API systems and excels on reasoning-intensive tasks, suggesting that effective tool reasoning can be learned from structured traces without live interactions.
title Adaptive Tool Generation with Models as Tools and Reinforcement Learning
topic Computation and Language
url https://arxiv.org/abs/2510.06825