Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Correctness Assessment of Code Generated by Large Language Models Using Internal Representations

Tuan-Dung Bui, Thanh Trong Vu, Thu-Trang Nguyen, Son Nguyen, Hieu Dinh Vo

TL;DR

OPENIA introduces a white-box framework that leverages in-process internal representations from open-source code LLMs to predict the correctness of generated code. It demonstrates that hidden activations encode reliability signals and trains a lightweight probing classifier to predict correctness during generation. Across HumanEval, MBPP, and DeviEval benchmarks, OPENIA outperforms post-hoc baselines and LLM-as-a-judge settings in both standalone and repo-level tasks, with substantial improvements in F1 and accuracy. The results suggest practical benefits for proactive quality assurance and potential for correctness-guided generation workflows in AI-assisted software development.

Abstract

Ensuring the correctness of code generated by Large Language Models (LLMs) presents a significant challenge in AI-driven software development. Existing approaches predominantly rely on black-box (closed-box) approaches that evaluate correctness post-generation, failing to utilize the rich insights embedded in the LLMs' internal states during code generation. In this paper, we introduce OPENIA, a novel white-box (open-box) framework that leverages these internal representations to assess the correctness of LLM-generated code. OPENIA systematically analyzes the intermediate states of representative open-source LLMs specialized for code, including DeepSeek-Coder, CodeLlama, and MagicCoder, across diverse code generation benchmarks. Our empirical analysis reveals that these internal representations encode latent information, which strongly correlates with the correctness of the generated code. Building on these insights, OPENIA uses a white-box/open-box approach to make informed predictions about code correctness, offering significant advantages in adaptability and robustness over traditional classification-based methods and zero-shot approaches. Experimental results demonstrate that OPENIA consistently outperforms baseline models, achieving higher accuracy, precision, recall, and F1-Scores with up to a 2X improvement in standalone code generation and a 46% enhancement in repository-specific scenarios. By unlocking the potential of in-process signals, OPENIA paves the way for more proactive and efficient quality assurance mechanisms in LLM-assisted code generation.

Correctness Assessment of Code Generated by Large Language Models Using Internal Representations

TL;DR

OPENIA introduces a white-box framework that leverages in-process internal representations from open-source code LLMs to predict the correctness of generated code. It demonstrates that hidden activations encode reliability signals and trains a lightweight probing classifier to predict correctness during generation. Across HumanEval, MBPP, and DeviEval benchmarks, OPENIA outperforms post-hoc baselines and LLM-as-a-judge settings in both standalone and repo-level tasks, with substantial improvements in F1 and accuracy. The results suggest practical benefits for proactive quality assurance and potential for correctness-guided generation workflows in AI-assisted software development.

Abstract

Ensuring the correctness of code generated by Large Language Models (LLMs) presents a significant challenge in AI-driven software development. Existing approaches predominantly rely on black-box (closed-box) approaches that evaluate correctness post-generation, failing to utilize the rich insights embedded in the LLMs' internal states during code generation. In this paper, we introduce OPENIA, a novel white-box (open-box) framework that leverages these internal representations to assess the correctness of LLM-generated code. OPENIA systematically analyzes the intermediate states of representative open-source LLMs specialized for code, including DeepSeek-Coder, CodeLlama, and MagicCoder, across diverse code generation benchmarks. Our empirical analysis reveals that these internal representations encode latent information, which strongly correlates with the correctness of the generated code. Building on these insights, OPENIA uses a white-box/open-box approach to make informed predictions about code correctness, offering significant advantages in adaptability and robustness over traditional classification-based methods and zero-shot approaches. Experimental results demonstrate that OPENIA consistently outperforms baseline models, achieving higher accuracy, precision, recall, and F1-Scores with up to a 2X improvement in standalone code generation and a 46% enhancement in repository-specific scenarios. By unlocking the potential of in-process signals, OPENIA paves the way for more proactive and efficient quality assurance mechanisms in LLM-assisted code generation.
Paper Structure (30 sections, 5 equations, 8 figures, 6 tables)

This paper contains 30 sections, 5 equations, 8 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Internal Representations in LLMs for Code Generation
  3. The Code Generation Process
  4. Internal Representations
  5. A Framework Leveraging Internal Representations for LLM Generated Code Quality Assessment
  6. Task Formulation
  7. Framework Overview
  8. Evaluation Design
  9. Dataset Construction
  10. Studied Code LLMs
  11. Experimental Procedure
  12. Metrics
  13. Experimental Results
  14. Performance Analysis
  15. Performance Comparison
  16. ...and 15 more sections

Figures (8)

  • Figure 1: Openia's overview
  • Figure 2: An example of independent-unit code generation for a task in the HumanEval benchmark
  • Figure 3: An example of repo-level code generation for a task in DevEval benchmark
  • Figure 4: Impact of the length of the generated code on Openia's performance
  • Figure 5: Impact of layer selection on Openia's performance
  • ...and 3 more figures