SMARTCAL: An Approach to Self-Aware Tool-Use Evaluation and Calibration

TL;DR

This work investigates the reliability of LLMs in using tools for open-domain QA and identifies prevalent tool-misuse and overconfidence. It introduces SMARTCAL, a recalibration framework comprising Self-Evaluation (SE), Confidence Prior Collection (CPC), and Augmented Reasoning (AR) to mitigate tool abuse and improve confidence calibration. Across Mintaka, PopQA, and Entity Questions, evaluated on DSP and ART frameworks with multiple models, SMARTCAL delivers an average QA performance gain of 8.6% and reduces Expected Calibration Error (ECE) by 21.6%, demonstrating enhanced accuracy and more reliable tool usage. The approach underscores the value of collaborative, calibrated, multi-agent reasoning for trustworthy tool-enabled LLM deployments in real-world tasks.

Abstract

The tool-use ability of Large Language Models (LLMs) has a profound impact on a wide range of industrial applications. However, LLMs' self-control and calibration capability in appropriately using tools remains understudied. The problem is consequential as it raises potential risks of degraded performance and poses a threat to the trustworthiness of the models. In this paper, we conduct a study on a family of state-of-the-art LLMs on three datasets with two mainstream tool-use frameworks. Our study reveals the tool-abuse behavior of LLMs, a tendency for models to misuse tools with overconfidence. We also find that this is a common issue regardless of model capability. Accordingly, we propose a novel approach, \textit{SMARTCAL}, to mitigate the observed issues, and our results show an average of 8.6 percent increase in the QA performance and a 21.6 percent decrease in Expected Calibration Error (ECE) compared to baseline models.

Figures (8)

• Figure 1: Comparison between ART paranjape2023art, ART (V), and SMARTCAL on the complex QA task. ART (V) introduces verbalized confidence elicitation. SMARTCAL includes three steps to mitigate tool-abuse.
• Figure 2: Distribution of entity popularity for Mintaka, PopQA, and Entity Questions dataset.
• Figure 3: ECE and QA accuracy trend comparison on Mintaka dataset. ECE scores remain stable despite increasing model capability.
• Figure 4: Tracking of tool usage from Entity Questions data sciavolino-etal-2021-simple on GPT-4. Diagram on the left is the original ART tool usage, on the right is SMARTCAL tool usage breakdown.
• Figure 5: Calibration performance comparison of GPT-4 on Entity Questions data with ART (V) on the left and SMARTCAL on the right.