Skip to content

Conduct related work & novelty audit to anticipate reviewer objections #18

Description

@nv78

Why This Matters

Code generation benchmarks are the most crowded subfield in NLP. Reviewers will challenge "how is this different from HumanEval, MBPP, SWE-bench, or BigCodeBench?" A systematic audit is essential to articulate the exact differentiators — especially the security dimension, which is the primary novelty.

Audit Tasks

1. Comprehensive Code Benchmark Survey

Catalog all existing code generation and evaluation benchmarks. For each, document:

  • Name, venue, year
  • Task types (generation, completion, repair, security)
  • Languages covered
  • Whether it evaluates security/vulnerability properties
  • Contamination risk level
  • Key gaps relative to our contribution

Key works to cover: HumanEval, MBPP, APPS, CodeContests, DS-1000, SWE-bench, SWE-bench Verified, BigCodeBench, LiveCodeBench, EvalPlus, CRUXEval, CyberSecEval, CWE-bench, SecurityEval, LLMSecEval

2. Security Benchmark Novelty

The security evaluation angle is our primary differentiator. Document exactly which security task types are NOT covered by CyberSecEval, SecurityEval, and CWE-bench. Our novelty must be in this gap.

3. Specification Gaming Novelty

Specification gaming and intent specification compliance evaluation is underexplored. Search specifically for papers evaluating whether code models follow specifications vs. exploit them.

4. Anticipate the "Already Done" Review

Strongest objections: "BigCodeBench already covers diverse coding tasks" and "CyberSecEval already covers security." Write rebuttals focusing on our integrated security + specification compliance evaluation and contamination-resistant design.

5. Citation Completeness Check

  • Search Semantic Scholar for code generation evaluation papers from 2022–2025
  • Ensure all evaluated models (GPT-4o, Claude, Gemini, Codestral, etc.) are cited
  • Check that SWE-bench, EvalPlus, and LiveCodeBench (which updated HumanEval) are cited and distinguished

Deliverable

A related_work_audit.md file with comparison table, novelty matrix, strongest objections + rebuttals, and refined novelty statement.

Metadata

Metadata

Assignees

No one assigned

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions