Paper·arxiv.org

llmevaluationresearchai-agentsmachine-learningtrace-(tool-for-rubric-analysis-in-code-evaluation)

Comparing Developer and LLM Biases in Code Evaluation

Implement the TRACE framework to rigorously evaluate Large Language Models (LLMs) used as code judges. This pack guides you in comparing LLM biases against human developer biases in realistic scenarios, ensuring your AI-assisted development tools accurately predict human judgments and foster reliable software processes.

intermediate1-2 hours6 steps

The play

Understand the Need for Human-Centric LLM Evaluation
Recognize that traditional LLM evaluation often misses realistic interactive scenarios, partial context, and ambiguous intent. Acknowledge the critical need for robust, human-centric evaluation methodologies for AI systems in sensitive applications like code assessment.
Define Your Code Evaluation Rubric
Establish clear, structured criteria (a rubric) for evaluating code quality, correctness, style, and intent. This rubric will be used consistently by both human developers and the LLM under evaluation. Consider factors like functionality, readability, efficiency, and adherence to best practices.
Gather Human Developer Judgments
Select a representative set of code snippets or solutions. Have multiple human developers independently evaluate these code samples against your defined rubric, capturing their scores and qualitative feedback. This forms your 'ground truth' for human judgment.
Prompt the LLM for Code Judgments
Configure your LLM to act as a judge. Provide the LLM with the same code snippets and the exact evaluation rubric used by human developers. Prompt the LLM to provide its judgment (e.g., scores, feedback) according to the rubric.
Compare LLM and Human Biases
Analyze the judgments from the LLM against the human developer judgments. Identify discrepancies, systematic biases, and areas where the LLM consistently deviates from human consensus. Focus on understanding *why* the LLM's judgments differ, considering context and intent.
Iterate and Refine Your LLM or Evaluation Process
Based on the identified biases, refine your LLM's prompting, fine-tuning data, or the evaluation rubric itself. The goal is to improve the LLM's ability to align with human judgments, thereby building more trustworthy and reliable AI-assisted development tools.

Starter code

{
  "evaluation_rubric": {
    "code_functionality": {
      "description": "Does the code correctly implement the specified requirements?",
      "scale": "1-5",
      "criteria": [
        "All test cases pass",
        "Handles edge cases correctly",
        "Produces expected output"
      ]
    },
    "code_readability": {
      "description": "Is the code easy to understand and maintain?",
      "scale": "1-5",
      "criteria": [
        "Clear variable names",
        "Sufficient comments",
        "Consistent formatting"
      ]
    },
    "code_efficiency": {
      "description": "Does the code use resources effectively (time/memory)?",
      "scale": "1-5",
      "criteria": [
        "Optimal algorithm chosen",
        "Avoids unnecessary computations"
      ]
    }
  },
  "code_snippet_id": "example_snippet_001",
  "code_to_evaluate": "def add(a, b):\n    return a + b"
}

Source

Paperarxiv.org