Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Emerging from Ground: Addressing Intent Deviation in Tool-Using Agents via Deriving Real Calls into Virtual Trajectories

Qian Xiong, Yuekai Huang, Bo Yang, Yujia Zheng, Tianhao Li, Ziyou Jiang, Zhiyuan Chang, Zhaoyang Li, Huanxiang Feng, Mingyang Li

TL;DR

This work tackles the subtle problem of intent deviation in tool-using LLM agents by introducing RISE, a Real-to-Virtual data synthesis framework anchored to verified real tool primitives. RISE generates realistic virtual trajectories and diverse negative samples through ICP mutations, then applies a two-stage training process (SFT followed by DPO) to align model behavior with user intent. Across data-quality, intent-alignment, and generalization experiments, RISE yields substantial improvements (e.g., average Acc_task +35.28% and Acc_intent +23.27%) and strong performance on unseen domains, outperforming state-of-the-art baselines. The approach provides a principled, scalable path to robust tool use in LLMs and includes open-source resources to facilitate future research and benchmarking.

Abstract

LLMs have advanced tool-using agents for real-world applications, yet they often lead to unexpected behaviors or results. Beyond obvious failures, the subtle issue of "intent deviation" severely hinders reliable evaluation and performance improvement. Existing post-training methods generally leverage either real system samples or virtual data simulated by LLMs. However, the former is costly due to reliance on hand-crafted user requests, while the latter suffers from distribution shift from the real tools in the wild. Additionally, both methods lack negative samples tailored to intent deviation scenarios, hindering effective guidance on preference learning. We introduce RISE, a "Real-to-Virtual" method designed to mitigate intent deviation. Anchoring on verified tool primitives, RISE synthesizes virtual trajectories and generates diverse negative samples through mutation on critical parameters. With synthetic data, RISE fine-tunes backbone LLMs via the two-stage training for intent alignment. Evaluation results demonstrate that data synthesized by RISE achieve promising results in eight metrics covering user requires, execution trajectories and agent responses. Integrating with training, RISE achieves an average 35.28% improvement in Acctask (task completion) and 23.27% in Accintent (intent alignment), outperforming SOTA baselines by 1.20--42.09% and 1.17--54.93% respectively.

Emerging from Ground: Addressing Intent Deviation in Tool-Using Agents via Deriving Real Calls into Virtual Trajectories

TL;DR

This work tackles the subtle problem of intent deviation in tool-using LLM agents by introducing RISE, a Real-to-Virtual data synthesis framework anchored to verified real tool primitives. RISE generates realistic virtual trajectories and diverse negative samples through ICP mutations, then applies a two-stage training process (SFT followed by DPO) to align model behavior with user intent. Across data-quality, intent-alignment, and generalization experiments, RISE yields substantial improvements (e.g., average Acc_task +35.28% and Acc_intent +23.27%) and strong performance on unseen domains, outperforming state-of-the-art baselines. The approach provides a principled, scalable path to robust tool use in LLMs and includes open-source resources to facilitate future research and benchmarking.

Abstract

LLMs have advanced tool-using agents for real-world applications, yet they often lead to unexpected behaviors or results. Beyond obvious failures, the subtle issue of "intent deviation" severely hinders reliable evaluation and performance improvement. Existing post-training methods generally leverage either real system samples or virtual data simulated by LLMs. However, the former is costly due to reliance on hand-crafted user requests, while the latter suffers from distribution shift from the real tools in the wild. Additionally, both methods lack negative samples tailored to intent deviation scenarios, hindering effective guidance on preference learning. We introduce RISE, a "Real-to-Virtual" method designed to mitigate intent deviation. Anchoring on verified tool primitives, RISE synthesizes virtual trajectories and generates diverse negative samples through mutation on critical parameters. With synthetic data, RISE fine-tunes backbone LLMs via the two-stage training for intent alignment. Evaluation results demonstrate that data synthesized by RISE achieve promising results in eight metrics covering user requires, execution trajectories and agent responses. Integrating with training, RISE achieves an average 35.28% improvement in Acctask (task completion) and 23.27% in Accintent (intent alignment), outperforming SOTA baselines by 1.20--42.09% and 1.17--54.93% respectively.
Paper Structure (26 sections, 6 equations, 4 figures, 1 table)

This paper contains 26 sections, 6 equations, 4 figures, 1 table.

Table of Contents

  1. Introduction
  2. Definition of Critical Tool and Parameter
  3. Methodology
  4. Tool Ground Initialization
  5. Real-to-Virtual Data Synthesis
  6. Tool Primitive Generation and Executability Validation
  7. Trajectory Synthesis via Multiple Planing Patterns
  8. Request Synthesis and Correctness Validation
  9. Negatives Expansion via Multi-type Mutations
  10. ICP Identification via Semantic Mapping
  11. Negative Generation via Multi-Type Mutations on ICPs
  12. Negatives Subsampling via Stratified Binning
  13. Two-Stage Training for Intent Alignment
  14. Experimental Design
  15. Research Questions
  16. ...and 11 more sections

Figures (4)

  • Figure 1: The difference between real tools and virtual tools simulated by GPT-4o, exemplified by the githubSearchPullRequests tool (Description: "Search GitHub pull requests with comprehensive filtering and analysis"). (1) Element Absence, where virtual tools lack the comprehensive information elements essential for the completion of specific tasks; (2) Pattern Degradation, where rich parameter patterns (e.g., unions, regex) are simplified into a data primitive; (3) General Security Missing, where general security constraints, like sensitive information protection, are out of consideration.
  • Figure 2: Overview of the RISE for intent-aligned tool-using agents.
  • Figure 3: Results of quality evaluation of the synthetic data
  • Figure 4: Performance Comparison of the original model (ORG) and enhanced by RISE (Ours) on OOD datasets