Learning to Generate Unit Tests for Automated Debugging

TL;DR

This paper introduces UTGen and UTDebug, novel methods for generating and validating unit tests to improve automated debugging and code correctness assessment using large language models.

Contribution

The paper presents UTGen and UTDebug, new techniques for generating error-revealing unit tests and validating their outputs to enhance LLM-based debugging and code evaluation.

Findings

UTGen outperforms baselines by 7.59% in error-revealing test generation.

Using UTGen with UTDebug improves pass@1 accuracy on HumanEvalFix and MBPP+ datasets.

UTGen enhances code correctness judgment, surpassing a state-of-the-art reward model by 4.43%.

Abstract

Unit tests (UTs) play an instrumental role in assessing code correctness as well as providing feedback to large language models (LLMs), motivating automated test generation. However, we uncover a trade-off between generating unit test inputs that reveal errors when given a faulty code and correctly predicting the unit test output without access to the gold solution. To address this trade-off, we propose UTGen, which teaches LLMs to generate unit test inputs that reveal errors along with their correct expected outputs based on task descriptions. Since model-generated tests can provide noisy signals (e.g., from incorrectly predicted outputs), we propose UTDebug that (i) scales UTGen via test-time compute to improve UT output prediction, and (ii) validates and backtracks edits based on multiple generated UTs to avoid overfitting, and helps LLMs debug effectively. We show that UTGen outperforms other LLM-based baselines by 7.59% based on a metric measuring the presence of both error-revealing UT inputs and correct UT outputs. When used with UTDebug, we find that feedback from UTGen's unit tests improves pass@1 accuracy of Qwen2.5 32B on HumanEvalFix and our own harder debugging split of MBPP+ by over 3.17% and 12.35% (respectively) over other LLM-based UT generation baselines. Moreover, we observe that feedback from Qwen2.5 32B-based UTGen model can enhance debugging with frontier LLMs like GPT-4o by 13.8%. Lastly, we demonstrate that UTGen is a better judge for code correctness, outperforming a state-of-the-art trained 8B reward model by 4.43% on HumanEval+ with best-of-10 sampling using Qwen2.5 7B.