HalluJudge: A Reference-Free Hallucination Detection for Context Misalignment in Code Review Automation

TL;DR

HalluJudge tackles the problem of hallucinations in LLM-generated code review comments by reframing detection as context alignment with the code diff. It introduces a grounding function and four assessment strategies, including a Tree-of-Thoughts approach, and evaluates them on Atlassian’s enterprise data. The study demonstrates that reference-free detection can achieve an $F1$ of $0.85$ with a modest average cost of $0.009$ per inference, and that about $0.67$ of judgments align with developer preferences in production. These findings suggest HalluJudge can serve as a practical safeguard to reduce ungrounded feedback and increase trust in AI-assisted code reviews, enabling safer large-scale deployment. The work also highlights the trade-offs between detection performance and cost, informing practitioners about when more elaborate reasoning (e.g., ToT) is warranted in production pipelines.

Abstract

Large Language models (LLMs) have shown strong capabilities in code review automation, such as review comment generation, yet they suffer from hallucinations -- where the generated review comments are ungrounded in the actual code -- poses a significant challenge to the adoption of LLMs in code review workflows. To address this, we explore effective and scalable methods for a hallucination detection in LLM-generated code review comments without the reference. In this work, we design HalluJudge that aims to assess the grounding of generated review comments based on the context alignment. HalluJudge includes four key strategies ranging from direct assessment to structured multi-branch reasoning (e.g., Tree-of-Thoughts). We conduct a comprehensive evaluation of these assessment strategies across Atlassian's enterprise-scale software projects to examine the effectiveness and cost-efficiency of HalluJudge. Furthermore, we analyze the alignment between HalluJudge's judgment and developer preference of the actual LLM-generated code review comments in the real-world production. Our results show that the hallucination assessment in HalluJudge is cost-effective with an F1 score of 0.85 and an average cost of $0.009. On average, 67% of the HalluJudge assessments are aligned with the developer preference of the actual LLM-generated review comments in the online production. Our results suggest that HalluJudge can serve as a practical safeguard to reduce developers' exposure to hallucinated comments, fostering trust in AI-assisted code reviews.

Figures (3)

• Figure 1: An example of a hallucinated code review comment (i.e., claims about an SQL injection risk) that is context-misaligned with the code diff (i.e., no area of code related to SQL or user-controlled queries).
• Figure 2: (RQ2) The distribution of token counts and monetary costs for each inference of hallucination judgment.
• Figure 3: An example of HalluJudge's explanation of a hallucinated code review comment.