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

TL;DR

This paper introduces RISE, a novel method that synthesizes virtual trajectories and negative samples to improve intent alignment in tool-using LLM agents, significantly enhancing task completion and intent accuracy.

Contribution

RISE provides a new data synthesis approach using verified tool primitives and mutation, effectively addressing intent deviation and improving LLM performance in real-world tool-using scenarios.

Findings

Achieved 35.28% improvement in task completion.

Achieved 23.27% improvement in intent alignment.

Outperformed state-of-the-art methods by up to 54.93%.

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.