AI Paper Atlas

Research map / Large language models / Agents, code and math

CodeT: Code Generation with Generated Tests

Bei Chen, Fengji Zhang, Anh Nguyen, Daoguang Zan, Zeqi Lin, Jian-Guang Lou, Weizhu Chen

ICLR 2023 · Large language models › Agents, code and math · approach: Execution-based code verification · 60 citations (OpenAlex)

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 changedWhereMeasuredEffect
Use CODET dual execution agreement (generated tests + solution consensus) instead of baseline sampling selectioncode-davinci-002 · HumanEvalpass@1: +18.8Improved
Use CODET dual execution agreement instead of baseline selectioncode-davinci-002 · MBPPpass@1: +9.6Improved
Use CODET dual execution agreement instead of baseline selectioncode-davinci-002 · APPS INTRODUCTORYpass@1: +7.4Improved
Use CODET dual execution agreement instead of baseline selectioncode-davinci-002 · APPS COMPETITION and CodeContestspass@1: +0.4 (COMPETITION), +1.4 (CodeContests)No clear effect
Use CODET dual execution agreement instead of baseline selectionINCODER-6B and CODEGEN-MONO-16B · HumanEval and MBPPpass@1: +4.2 to +13.1Improved
Score consensus sets by number of code solutions only (f' ablation) instead of joint solutions x tests (ablation)code-davinci-002 · HumanEvalpass@1: -9.9 vs CODETHurt
Score consensus sets by number of test cases only (f'' ablation) instead of joint solutions x tests (ablation)code-davinci-002 · HumanEvalpass@1: -7.4 vs CODETHurt
Reduce generated test cases from 500 to about 10 per problem (ablation)code-davinci-002 CODET · HumanEvalpass@1: 65.8 to 56.5Hurt
De-duplicate code solutions and test cases before CODET (ablation)code-cushman-001 CODET · HumanEval and MBPPpass@1: HumanEval 44.5 to 42.7; MBPP 55.4 to 54.7No clear effect
Use test cases generated by code-davinci-002 instead of model's own tests (ablation)code-cushman-001, code-davinci-001, INCODER, CODEGEN · HumanEval and MBPPpass@1: absolute gains of 1.8% to 4.3% for Codex cushman/davinci-001; 6.2% to 15.9% for INCODER and CODEGENImproved
Remove trivial code solutions before CODET (ablation)code-davinci-002 CODET · APPS INTRODUCTORY and CodeContestspass@1: +0.3 (INTRODUCTORY), +0.6 (CodeContests)No clear effect
Remove example input-output cases from HumanEval context (ablation)baseline and CODET · HumanEval original vs modified contextpass@1Mixed

Later papers that compared with this method

14 later accepted papers report CODET as a baseline in their results tables; 13 of them report a better number on most of the results they share with it.

6 more not listed.

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

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy