Do Large Code Models Understand Programming Concepts? Counterfactual Analysis for Code Predicates

2024-02-08

International Conference on Machine Learning

AI Summary

Existing evaluation methods inadequately capture whether large code models genuinely comprehend core programming logic concepts—such as data flow and control flow—at the semantic level. Method: To address this gap, we propose CACP, the first counterfactual black-box evaluation framework targeting programming concept predicates. CACP requires no access to model internals; instead, it systematically quantifies conceptual understanding via program predicate modeling and abstract representations of control and data flow. Contribution/Results: CACP evaluates 10 mainstream large code models across four critical programming concept categories. Experimental results reveal a fundamental deficiency in all models’ grasp of basic logical concepts, with performance substantially below human programmer expectations. This work establishes a novel benchmark and methodology for interpretability analysis and capability assessment of large language models in code-related tasks.

Technology Category

• Cognitive Modeling & Cognitive Systems: Conceptual Inference and Reasoning
• Natural Language Processing: Interpretability, Analysis, and Evaluation of NLP Models
• Computer Vision: Large Vision Models

Application Category

• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking

Abstract

Large Language Models' success on text generation has also made them better at code generation and coding tasks. While a lot of work has demonstrated their remarkable performance on tasks such as code completion and editing, it is still unclear as to why. We help bridge this gap by exploring to what degree auto-regressive models understand the logical constructs of the underlying programs. We propose Counterfactual Analysis for Programming Concept Predicates (CACP) as a counterfactual testing framework to evaluate whether Large Code Models understand programming concepts. With only black-box access to the model, we use CACP to evaluate ten popular Large Code Models for four different programming concepts. Our findings suggest that current models lack understanding of concepts such as data flow and control flow.

Problem

• Evaluate Large Code Models' understanding of programming concepts.
• Propose Counterfactual Analysis for Programming Concept Predicates.
• Assess models' grasp of data flow and control flow.

Innovation

• Counterfactual Analysis for Programming
• Evaluate Large Code Models
• Black-box access technique