Beyond Accuracy: Characterizing Code Comprehension Capabilities in (Large) Language Models

2026-01-19

AI Summary

This work addresses the limitation of existing code understanding benchmarks, which offer only coarse-grained evaluations and fail to reveal the specific capabilities and shortcomings of large language models (LLMs). To enable fine-grained diagnosis, the authors reformulate code understanding as an input–output consistency verification task and construct a diagnostic framework that evaluates both classification and generation models at the instance level. They further investigate the relationship between model performance and human-centric software complexity metrics—such as lexical size, control flow complexity, and abstract syntax tree (AST) structure. Experimental results show that traditional complexity metrics exhibit weak correlation with model performance (AUROC 0.63), whereas predictions derived from shadow models achieve substantially higher correlation (AUROC 0.86), suggesting that LLMs follow non-human-centric patterns in code comprehension and highlighting the urgent need for new evaluation paradigms.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: Interpretability, Analysis, and Evaluation of NLP Models
• Knowledge Representation and Reasoning: Computational Complexity of Reasoning

Application Category

• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems
• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

Abstract

Large Language Models (LLMs) are increasingly integrated into software engineering workflows, yet current benchmarks provide only coarse performance summaries that obscure the diverse capabilities and limitations of these models. This paper investigates whether LLMs'code-comprehension performance aligns with traditional human-centric software metrics or instead reflects distinct, non-human regularities. We introduce a diagnostic framework that reframes code understanding as a binary input-output consistency task, enabling the evaluation of classification and generative models. Using a large-scale dataset, we correlate model performance with traditional, human-centric complexity metrics, such as lexical size, control-flow complexity, and abstract syntax tree structure. Our analyses reveal minimal correlation between human-defined metrics and LLM success (AUROC 0.63), while shadow models achieve substantially higher predictive performance (AUROC 0.86), capturing complex, partially predictable patterns beyond traditional software measures. These findings suggest that LLM comprehension reflects model-specific regularities only partially accessible through either human-designed or learned features, emphasizing the need for benchmark methodologies that move beyond aggregate accuracy and toward instance-level diagnostics, while acknowledging fundamental limits in predicting correct outcomes.

Problem

• code comprehension
• large language models
• software complexity metrics
• model evaluation
• benchmarking

Innovation

• code comprehension
• large language models
• diagnostic framework
• shadow models
• software complexity metrics