Research map / Large language models / Agents, code and math
EditBench: Evaluating LLM Abilities to Perform Real-World Instructed Code Edits
Wayne Chi, Valerie Chen, Ryan Shar, Aditya Mittal, Jenny Liang, Wei-Lin Chiang, Anastasios Nikolas Angelopoulos, Ion Stoica, Graham Neubig, Ameet Talwalkar, Chris Donahue
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 highlighted code to the prompt (ablation) | top model of each of 7 families · EDIT-Bench core | pass@1 (%): claude-sonnet-4 62.41 -> 64.81; o3-mini 60.00 -> 56.30 | Mixed |
| add cursor position on top of highlighted code (ablation) | top model of each of 7 families · EDIT-Bench core | pass@1 (%): glm-4.6 56.48 -> 44.81; kimi-k2-0905 56.48 -> 58.15 | Mixed |
| add highlighted code for o3-mini (ablation) | o3-mini · EDIT-Bench core | pass@1 (%): 60.00 -> 56.30 | Hurt |
| add highlighted code for glm-4.6 (ablation) | glm-4.6 · EDIT-Bench core | pass@1 (%): 52.96 -> 56.48 | Improved |
| add cursor position for o3-mini and qwen3-coder (ablation) | o3-mini, qwen3-coder · EDIT-Bench core | pass@1 (%): o3-mini 56.30 -> 55.19; qwen3-coder 53.89 -> 53.89 | No clear effect |
| higher reasoning effort or default reasoning in gpt-5 vs gpt-5-mini | gpt-5 · EDIT-Bench core | pass@1 | Hurt |
| use a fully agentic pipeline to generate test harnesses instead of human annotators | test harness creation · EDIT-Bench | pass@1 (%): 0% | Hurt |
| compare easy vs hard problems across models | 40 evaluated LLMs · EDIT-Bench | pass@1 gap: average gap 58.8% | Mixed |
Closest papers by meaning
- StackEval: Benchmarking LLMs in Coding Assistance (NeurIPS 2024 · similarity 0.88)
- EditLord: Learning Code Transformation Rules for Code Editing (ICML 2025 · similarity 0.88)
- AgentBench: Evaluating LLMs as Agents (ICLR 2024 · similarity 0.88)
- TaskBench: Benchmarking Large Language Models for Task Automation (NeurIPS 2024 · similarity 0.87)
- Unsupervised Evaluation of Code LLMs with Round-Trip Correctness (ICML 2024 · similarity 0.87)
- Copilot Arena: A Platform for Code LLM Evaluation in the Wild (ICML 2025 · similarity 0.87)
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.
