Probe-Guided Training Improves LLM Tool Call Parameter Accuracy

A new study addresses a critical challenge in large language model (LLM) agents: accurately filling the parameters of tool calls. While previous research focused on tool selection and orchestration, this work highlights that correct parameter generation is equally vital, noting that even frontier models struggle with this task in complex domains like cloud networking. To tackle this, researchers developed ParamBench, a difficulty-graded benchmark built from real cloud-network APIs, categorizing instances by parameter nesting, dependencies, and reasoning complexity. The core innovation is a probe-guided framework, which leverages a strong correctness signal found in LLM hidden states during parameter generation. A simple linear probe can predict whether a generated value is correct. This signal is used in two ways: Probe-Filtered Bootstrapped Training (PBT) filters reliable self-generated calls for fine-tuning, and Probe-Guided Reranking (PGR) selects better candidates during inference. Extensive experiments show this method dramatically improves parameter generation, increasing the average exact match rate from 19.7% to 59.6% across various open models and benchmarks.