PaperlayerView exact revision on arXiv

Source-linked AI summary

FunReason-MT Technical Report: Advanced Data Synthesis Solution for Real-world Multi-Turn Tool-use

Zengzhuang Xu, Bingguang Hao, Zechuan Wang, Yuntao Wen, Xinyi Xu, Yang Liu, Long Chen, Dong Wang, Maolin Wang, Tong Zhao, Yicheng Chen, Cunyin Peng, Jinjie Gu, Leilei Gan, Xiangyu Zhao, Chenyi Zhuang, Shi Gu

arXiv:2510.24645v2cs.AI

TL;DR

Existing random sampling and multi-agent role-playing struggle with targeted synthesis, hard-query construction, and multi-turn logical dependencies, limiting the quality of multi-turn function-calling data. FunReason-MT uses environment-API dependencies to generate targeted trajectories, synthesizes advanced tools and hard queries, and iteratively refines CoT with feedback. After RL, FunReason-MT reaches a 57.75 (+42.00) Multi-Turn score on BFCLv3, achieving state-of-the-art results, and a 15.10 Overall score on BFCLv4.

  • Problem

    Existing random sampling and multi-agent role-playing struggle with targeted synthesis, hard-query construction, and multi-turn logical dependencies, limiting the quality of multi-turn function-calling data.

    Source: §1 · p. 2

  • Method

    FunReason-MT uses environment-API dependencies to generate targeted trajectories, synthesizes advanced tools and hard queries, and iteratively refines CoT with feedback.

    Source: §1 · p. 2, §4.2 · p. 5

  • Results

    After RL, FunReason-MT reaches a 57.75 (+42.00) Multi-Turn score on BFCLv3, achieving state-of-the-art results, and a 15.10 Overall score on BFCLv4.

    Source: §5.2 · pp. 6–7, §5.3 · p. 7

  • Takeaways & Limitations

    The authors present FunReason-MT as a robust source of agentic-learning data, with reported BFCLv4 gains even though its training data is unrelated to Web Search and Memory tasks.

    Source: §5.3 · p. 7

Abstract

from arXiv · show

Function calling (FC) empowers large language models (LLMs) and autonomous agents to interface with external tools, a critical capability for solving complex, real-world problems. As this ability becomes increasingly central to advanced AI systems, the need for high-quality, multi-turn training data to develop and refine it cannot be overstated. Existing data synthesis methods, such as random environment sampling or multi-agent role-playing, are not powerful enough to generate high-quality data in real-world environments. Practical challenges come in three folds: targeted data synthesis, hard query construction, and multi-turn logical dependency. To address these structural deficiencies, we present FunReason-MT, a novel data synthesis framework for real-world multi-turn tool use. FunReason-MT resolves the complexity barrier in multi-turn FC data by employing 1) Environment-API Graph Interactions to gather varied high-quality trajectories with targeted tool, 2) Advanced Tool-Query Synthesis to simplify hard query construction, and 3) Guided Iterative Chain for sophisticated CoT generation. Evaluations on Berkeley Function-Calling Leaderboard (BFCLv3) demonstrate the power of our framework: a 4B model built upon FunReason-MT generated data achieves state-of-the-art performance among comparable-sized models. Further performance improvements on BFCLv4 confirm that FunReason-MT provides a reliable and robust source for agentic learning.

1 Introduction

The introduction argues that existing synthesis approaches struggle to produce diverse, complex multi-turn tool-use data. FunReason-MT combines targeted trajectory generation, hard-query construction, and guided reasoning, and is reported to achieve strong benchmark results.

Problem

  • Random environment sampling misses rare, complex events, while multi-agent role-playing can produce simplistic or insufficiently diverse scenarios, limiting training data effectiveness.The introduction identifies targeted synthesis, hard-query construction, and multi-turn logical dependency as connected challenges.Source: §1 · p. 2

FunReason-MT components

  • Environment-API Graph Interactions use tool and environment dependencies to gather varied multi-turn trajectories centered on a targeted complex API.The stated aim is to support targeted tool mastery and collaborative use.Source: §1 · p. 2
  • Advanced Tool-Query Synthesis uses collected tool-state pairs to construct an advanced tool, making prerequisite sub-tools implicit in generated queries.The introduction describes the reverse construction as a way to simplify hard-query generation.Source: §1 · p. 2
  • Guided Iterative Chain generates CoT from the hard query, sub-tools, and advanced-tool description, using corrective feedback until the function call is correct or attempts are exhausted.The feedback targets reasons the generated call fails to match ground truth.Source: §1 · p. 3

Results

  • FunReason-MT achieves state-of-the-art performance among comparable-sized models on BFCLv3 and improves performance on BFCLv4, including Web Search and Memory.The BFCLv3 evaluation includes single-turn and multi-turn tests.Source: §1 · p. 3

2 Motivation

The motivation is that data-generation methods relying on complexity emerging from simple interactions have limited controllability and reliability. FunReason-MT instead directs synthesis toward specific complex tools and scenarios.

2 Motivation

  • Random sampling and role-playing are described as bottom-up methods that hope complexity emerges from simple interactions, with low controllability and reliability.The paper frames the performance ceiling as a limitation of the generation paradigm, not merely data quantity.Source: §2 · p. 3
  • The proposed top-down methodology explicitly directs the model toward mastery of a specific complex tool in a targeted complex scenario.The shift is motivated by the difficulty of obtaining such trajectories through less directed interactions.Source: §2 · p. 3
  • The approach models environment dependencies, generates logical-jump queries in reverse, and uses guided feedback to refine CoT accuracy.These mechanisms target complexity and reliability in data generation.Source: §2 · p. 3

3 Notation

The section introduces the notation and concepts used to model FunReason-MT’s system components and operations.

3 Notation

  • The report introduces key symbols and concepts for its system model, with core concepts and models defined in Table 1.The table is identified as covering FunReason-MT’s core concepts and models.Source: §3 · p. 3, Table 1 · p. 3

4 Methodology

FunReason-MT constructs multi-turn tool-use trajectories in three phases: dependency-aware trace generation, advanced-tool and hard-query synthesis, and iterative CoT refinement against ground truth.

4 Methodology

  • The methodology comprises three phases: build an execution trace, synthesize an advanced tool and hard query, then refine CoT for consistency with ground truth.Figure 3 depicts the three-phase framework.Source: §4 · pp. 3–4, Figure 3 · p. 4

4.1 Phase I: Environment-API Graph Interactions

  • Phase I balances executable tool sampling with goal-directed prioritization of a target tool.Execution correctness requires respecting tool dependencies; sampling is also directed toward the target.Source: §4.1 · p. 4

Formalizing Tool Call Legality

  • A tool is callable only when its prerequisites are among the tools already executed; the indicator returns 1 when that condition holds and 0 otherwise.A directed sampler prioritizes legal tools nearer the target, and each executed call updates the environment state and API graph recorded in the trace.Source: Formalizing Tool Call Legality · p. 4, Tool Sampling with Priority · p. 4

4.2 Phase II: Advanced Tool-Query Synthesis

  • Phase II abstracts an execution trace into an advanced tool that captures the sub-tools’ functionality and dependencies, then generates a hard query requiring that abstraction.Synthesis noise ε increases task difficulty and encourages generalization.Source: §4.2 · p. 5

4.3 Phase III: Guided Iterative Chain

  • Phase III uses feedback-driven refinement to improve CoT quality and coherence until reasoning converges to a reliable solution.The phase addresses the increased risk of logical inconsistency or incomplete CoT after hard-query construction.Source: §4.3 · p. 5

Initial Reasoning Attempt and Correction Loop

  • The reasoning agent generates a function call and CoT for the hard query, and the call is validated against the known ground-truth execution.On failure, a critiquing agent identifies the error and supplies a corrective hint for the next attempt; only validated samples retained before the attempt cap pass.Source: Initial Reasoning Attempt and… · p. 5

4.4 Scaling from Single-Turn Trace to Multi-Turn Trajectory

  • The scaling process repeats the three phases for N rounds, adding each successful trace, advanced-tool/query pair, and ground-truth-verified sample toward a multi-turn trajectory.The synthesis agents form the advanced tool and hard query from the successful trace; reasoning and critiquing agents verify the sample against the ground truth.Source: §4.4 · pp. 5–6

5 Experiments

Experiments evaluate FunReason-MT on BFCLv3 and held-out BFCLv4. The trained 4B model achieves strong BFCLv3 results and gains on BFCLv4 agentic tasks.

5.1 Experimental Setup

  • The experiments train on 17,000 multi-turn samples alongside APIGen function-calling data and evaluate BFCLv3 single- and multi-turn tasks plus held-out BFCLv4 Web Search and Memory.Qwen3-4B-Instruct-2507 is the base model; the report gives accuracy as the evaluation measure.Source: §5.1 · p. 6

5.2 Main Results

  • On BFCLv3, FunReason-MT raises Qwen3-4B-Instruct’s Multi-Turn score from 15.75 to 57.75 (+42.00) after RL, and its RL-trained model achieves state-of-the-art results.Source: §5.2 · pp. 6–7

5.3 Agentic Results

  • On BFCLv4, RL raises Overall from 8.85 to 15.10 (+6.25), Web Search Overall from 5.00 to 16.00 (+11.00), and Memory Overall from 12.69 to 14.19.The RL-trained model outperforms competing specialized models on the main Overall metric.Source: §5.3 · p. 7, Table 3 · p. 7

6 Conclusion

The conclusion presents FunReason-MT as a framework for addressing bottlenecks in multi-turn function-calling data generation. It reports state-of-the-art performance among comparable-sized models on BFCLv3 and promising out-of-distribution generalization on BFCLv4.

  • FunReason-MT is designed to address bottlenecks in multi-turn function-calling data generation, where existing methods produce insufficiently diverse and logically dependent trajectories.Source: §6 · p. 7
  • The framework combines Environment-API Graph Interactions, Advanced Tool-Query Synthesis, and Guided Iterative Chain for targeted data, complex queries, and CoT generation.Source: §6 · p. 7
  • The report claims state-of-the-art performance among comparable-sized models on BFCLv3 and promising out-of-distribution generalization on BFCLv4.Source: §6 · p. 7

A Supplementary Case

The case study illustrates multi-turn file operations and a hard ticket-purchase query that combines prerequisite lookups with a purchase. It shows the requested tool sequences and their returned outputs.

A Supplementary Case

  • A hard ticket query compresses two zipcode lookups and a zipcode-based purchase into one request; the trace returns ticket_id 14589.The trace gives zipcodes 83214 and 74532 for Rivermist and Stonebrook, then records the ticket purchase.Source: A Supplementary Case · p. 9
  • The directory-search turn uses pwd() followed by find(path=".") to locate entries under the current directory.The returned current directory is /workspace, with files and directories listed beneath it.Source: A Supplementary Case · p. 11
  • The follow-up reads unggv.json from data and tails seven lines of dtpsp.json from data/raw, returning product id 94 and user id 27.The calls navigate into data, read unggv.json, then enter raw and tail dtpsp.json.Source: A Supplementary Case · pp. 12–13

Related papers

Loading 2510.24645v2…