Research map / Large language models / Agents, code and math
Coder Reviewer Reranking for Code Generation
Tianyi Zhang, Tao Yu, Tatsunori Hashimoto, Mike Lewis, Wen-tau Yih, Daniel Fried, Sida Wang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add Reviewer likelihood to Coder-only reranking | Code generation reranking across six datasets and model families · 48 dataset-model pairs | accuracy: 43 out of 48 | Improved |
| Use length-normalized Coder-Reviewer scores instead of unnormalized scores (ablation) | Coder-Reviewer · HumanEval, Codex002 | accuracy: 56.7 -> 62.5 | Improved |
| Use the alternate MMI formulation based on subtracting the marginal program likelihood (ablation) | Coder-Reviewer · HumanEval | accuracy: 56.7 -> 55.30; 62.5 -> 50.73 | Hurt |
| Use the alternate MMI formulation based on subtracting the marginal program likelihood (ablation) | Coder-Reviewer · MBPP-S | accuracy: 64.4 -> 63.0; 61.6 -> 62.4 | Hurt |
| Tune the Coder/Reviewer mixing ratio instead of fixing it at 0.5 (ablation) | Coder-Reviewer · Most experimental settings | accuracy: less than 1% | Improved |
| Tune the Coder/Reviewer mixing ratio (ablation) | InCoder6B on Plotting and CodeGen16B on NL2Bash · Plotting; NL2Bash | accuracy | Improved |
| Remove degenerate-solution rejection (ablation) | Codex002 reranking methods · HumanEval | accuracy: 61.5 -> 53.2 | Hurt |
| Remove degenerate-solution rejection (ablation) | Codex002 reranking methods · MBPP-S | accuracy: 60.8 -> 60.5 | Hurt |
| Filter empty, trivial, repetitive, and surface-overlap-prone programs before reranking (ablation) | Coder, Reviewer, and Coder-Reviewer · HumanEval and MBPP-S | accuracy: 38.1 with 7.1 degradation; 55.3 with 4.5 degradation | Improved |
| Apply executability filtering before reranking (ablation) | Most ranking methods · Most datasets | accuracy | Improved |
| Increase the number of sampled programs reranked (ablation) | Coder-Reviewer versus Coder-only ranking · Zero-shot datasets | accuracy | Mixed |
| Use Reviewer-only ranking instead of Coder-only ranking | CodeGen models · Across datasets | accuracy: 5 exceptions out of 48 | Mixed |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- CodeT — CodeT: Code Generation with Generated Tests (ICLR 2023) · better on 0 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- CoRNStack: High-Quality Contrastive Data for Better Code Retrieval and Reranking (ICLR 2025 · similarity 0.87)
- Reranking Laws for Language Generation: A Communication-Theoretic Perspective (NeurIPS 2024 · similarity 0.86)
- Repoformer: Selective Retrieval for Repository-Level Code Completion (ICML 2024 · similarity 0.86)
- CPRet: A Dataset, Benchmark, and Model for Retrieval in Competitive Programming (NeurIPS 2025 · similarity 0.86)
- NextCoder: Robust Adaptation of Code LMs to Diverse Code Edits (ICML 2025 · similarity 0.85)
- Copilot Arena: A Platform for Code LLM Evaluation in the Wild (ICML 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
