TACO: Trust Assessment of Large Language Models in Coding Assistance Tasks

Large Language Models (LLMs) have rapidly become integral to software development workflows, particularly in coding assistance tasks (CAT) such as debugging, implementation, and code optimization. However, the trustworthiness of LLM-generated responses remains a critical concern, as hallucinations (incorrect or misleading outputs) can severely hinder developer productivity and software reliability. In this paper, we introduce TACO, a comprehensive framework for trust assessment of LLMs in CAT scenarios. TACO jointly evaluates both the code quality and the alignment with user intent of LLM responses, enabling fine-grained and interpretable trust evaluation. We construct two new benchmarks: TACO-Judge, a human-annotated dataset for validating evaluation methods, and TACO-Eval, a large-scale benchmark for assessing LLM performance on real-world CAT problems. Through extensive experiments, we (1) demonstrate the accuracy of TACO as an automated evaluator, (2) investigate whether LLMs exhibit self-preference when scoring their own outputs, and (3) compare the trustworthiness of several state-of-the-art LLMs in CAT scenarios. Our results highlight both the promise and limitations of current LLMs, and establish TACO as a reliable tool for their evaluation in software engineering contexts.