CoHalLo: code hallucination localization via probing hidden layer vector

TL;DR

CoHalLo is a novel method that localizes code hallucinations at the line level by probing hidden-layer vectors, improving the precision of identifying hallucinated code segments in AI-generated code.

Contribution

This paper introduces CoHalLo, a fine-grained hallucination localization technique using hidden-layer probing and syntactic analysis, advancing beyond coarse detection methods.

Findings

Achieves up to 83.33% Top-10 accuracy in localization

Outperforms baseline methods in hallucination detection metrics

Effectively identifies key syntactic structures associated with hallucinations

Abstract

The localization of code hallucinations aims to identify specific lines of code containing hallucinations, helping developers to improve the reliability of AI-generated code more efficiently. Although recent studies have adopted several methods to detect code hallucination, most of these approaches remain limited to coarse-grained detection and lack specialized techniques for fine-grained hallucination localization. This study introduces a novel method, called CoHalLo, which achieves line-level code hallucination localization by probing the hidden-layer vectors from hallucination detection models. CoHalLo uncovers the key syntactic information driving the model's hallucination judgments and locates the hallucinating code lines accordingly. Specifically, we first fine-tune the hallucination detection model on manually annotated datasets to ensure that it learns features pertinent to code syntactic information. Subsequently, we designed a probe network that projects high-dimensional latent vectors onto a low-dimensional syntactic subspace, generating vector tuples and reconstructing the predicted abstract syntax tree (P-AST). By comparing P-AST with the original abstract syntax tree (O-AST) extracted from the input AI-generated code, we identify the key syntactic structures associated with hallucinations. This information is then used to pinpoint hallucinated code lines. To evaluate CoHalLo's performance, we manually collected a dataset of code hallucinations. The experimental results show that CoHalLo achieves a Top-1 accuracy of 0.4253, Top-3 accuracy of 0.6149, Top-5 accuracy of 0.7356, Top-10 accuracy of 0.8333, IFA of 5.73, Recall@1% Effort of 0.052721, and Effort@20% Recall of 0.155269, which outperforms the baseline methods.