Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

ToolACE-MT: Non-Autoregressive Generation for Agentic Multi-Turn Interaction

Xingshan Zeng, Weiwen Liu, Lingzhi Wang, Liangyou Li, Fei Mi, Yasheng Wang, Lifeng Shang, Xin Jiang, Qun Liu

TL;DR

ToolACE-MT tackles the challenge of generating high-quality multi-turn, tool-using dialogues for agentic LLMs by adopting a non-autoregressive pipeline inspired by NAT and diffusion models. The method decomposes generation into coarse initialization, iterative refinement with mask-and-fill, and offline verification to ensure coherence, tool executability, and factual consistency. Empirical results on BFCL, ACEBench, and tau-Bench show substantial improvements in multi-turn accuracy and data efficiency over autoregressive MAS baselines and even exceed some larger models. The approach offers scalable, budget-aware data construction for tool-augmented LLM systems and demonstrates strong generalization across backbones.

Abstract

Agentic task-solving with Large Language Models (LLMs) requires multi-turn, multi-step interactions, often involving complex function calls and dynamic user-agent exchanges. Existing simulation-based data generation methods for such scenarios rely heavily on costly autoregressive interactions between multiple LLM agents, thereby limiting real-world performance of agentic tasks. In this paper, we propose ToolACE-MT, a novel Non-Autoregressive Iterative Generation framework for constructing high-quality multi-turn agentic dialogues. ToolACE-MT generates full conversational trajectories through three stages: coarse-grained initialization, iterative refinement, and offline verification. The initialization phase builds a structurally complete yet semantically coarse dialogue skeleton; the iterative refinement phase introduces realistic complexities and continued refinement via mask-and-fill operations; and the offline verification phase ensures correctness and coherence via rule- and model-based checks. Experiments demonstrate that ToolACE-MT enables efficient, effective and generalizable agentic data generation, offering a new paradigm for high-quality data construction in tool-augmented LLM scenarios.

ToolACE-MT: Non-Autoregressive Generation for Agentic Multi-Turn Interaction

TL;DR

ToolACE-MT tackles the challenge of generating high-quality multi-turn, tool-using dialogues for agentic LLMs by adopting a non-autoregressive pipeline inspired by NAT and diffusion models. The method decomposes generation into coarse initialization, iterative refinement with mask-and-fill, and offline verification to ensure coherence, tool executability, and factual consistency. Empirical results on BFCL, ACEBench, and tau-Bench show substantial improvements in multi-turn accuracy and data efficiency over autoregressive MAS baselines and even exceed some larger models. The approach offers scalable, budget-aware data construction for tool-augmented LLM systems and demonstrates strong generalization across backbones.

Abstract

Agentic task-solving with Large Language Models (LLMs) requires multi-turn, multi-step interactions, often involving complex function calls and dynamic user-agent exchanges. Existing simulation-based data generation methods for such scenarios rely heavily on costly autoregressive interactions between multiple LLM agents, thereby limiting real-world performance of agentic tasks. In this paper, we propose ToolACE-MT, a novel Non-Autoregressive Iterative Generation framework for constructing high-quality multi-turn agentic dialogues. ToolACE-MT generates full conversational trajectories through three stages: coarse-grained initialization, iterative refinement, and offline verification. The initialization phase builds a structurally complete yet semantically coarse dialogue skeleton; the iterative refinement phase introduces realistic complexities and continued refinement via mask-and-fill operations; and the offline verification phase ensures correctness and coherence via rule- and model-based checks. Experiments demonstrate that ToolACE-MT enables efficient, effective and generalizable agentic data generation, offering a new paradigm for high-quality data construction in tool-augmented LLM scenarios.

Paper Structure

This paper contains 28 sections, 15 figures, 9 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Method
  4. Problem Formulation
  5. Overview
  6. Coarse-Grained Initialization
  7. Task Initialization.
  8. Trajectory Initialization.
  9. Iterative Refinement
  10. Complexity Injection.
  11. Reasonability Refinement.
  12. Offline Verification
  13. Experiments
  14. Experimental Setup
  15. Data Quality Assessment
  16. ...and 13 more sections

Figures (15)

  • Figure 1: Multi-Agent Simulation v.s. our proposed Non-Autoregressive Generation.
  • Figure 2: Overall workflow for our ToolACE-MT framework.
  • Figure 3: Illustration figure for Iterative Refinement process.
  • Figure 4: Statistics of assistant turn counts for MAS and NAIG, measured on both the training data and successful inference cases in $\tau$-Bench.
  • Figure 5: The accuracy results of our method on BFCL-v3 when scaling Iterative Refinement times, with and without offline verification.
  • ...and 10 more figures