Sample, Predict, then Proceed: Self-Verification Sampling for Tool Use of LLMs

2025-06-03

AI Summary

Existing LLM-based tool-use frameworks struggle with reliable function calling in stateful environments, as current test-time computation strategies rely on iterative environment interaction—rendering them impractical in real-world deployments. To address this, we propose DyMo (Dynamic Modeling) and SVS (Self-Validating Sampling), a synergistic framework enabling zero-shot, zero-trial reliable tool planning without runtime environment feedback. DyMo introduces lightweight internal state modeling to predict the outcomes of candidate actions; SVS performs real-time confidence estimation and proactive refusal over function-call candidates during generation. We further enhance the framework via state-prediction-aware instruction tuning and function-call-specific post-training. Evaluated on the Berkeley Function Calling Leaderboard V2, our method achieves significant improvements in success rate and pass@k, while substantially reducing hallucination rates—demonstrating enhanced output reliability and robustness.

Technology Category

• Planning, Routing, and Scheduling: Planning with Language Models
• Machine Learning: Large Multimodal Models (LMMs)
• Search and Optimization: Sampling/Simulation-based Search

Application Category

• Search and Retrieval-Augmented AI: Search Tool Learning with LLM: Teaching LLMs to invoke search and make use of retrieved information
• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions
• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems

Abstract

Tool use in stateful environments presents unique challenges for large language models (LLMs), where existing test-time compute strategies relying on repeated trials in the environment are impractical. We propose dynamics modelling (DyMo), a method that augments LLMs with a state prediction capability alongside function calling during post-training. This enables LLMs to predict the future states of their actions through an internal environment model. On the Berkeley Function Calling Leaderboard V2, DyMo improves success rates and significantly reduces hallucinations. We further integrate the internal environment model into self-verification sampling (SVS), and show that this substantially improves pass^k over number of trials k, and allows the model to refuse unreliable outputs. Together, DyMo and SVS greatly enhance the effectiveness and reliability of LLMs for tool use. We believe this work charts a path towards scalable planning RL methods for LLM inference without repeatedly querying the oracle environment.

Problem

• Enhancing LLMs' tool use in stateful environments
• Reducing hallucinations in function calling tasks
• Improving reliability via self-verification sampling

Innovation

• Augments LLMs with state prediction capability
• Integrates internal environment model into self-verification
• Reduces hallucinations and improves success rates