Research map / Large language models / Agents, code and math
ANPL: Towards Natural Programming with Interactive Decomposition
Di Huang, Ziyuan Nan, Xing Hu, Pengwei Jin, Shaohui Peng, Yuanbo Wen, Rui Zhang, Zidong Du, Qi Guo, Yewen Pu, Yunji Chen
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 |
|---|---|---|---|
| use ANPL interactive decomposition instead of ChatGPT with interaction | GPT-3.5-turbo · ARC 400 tasks | solving rate (%): 58.4 -> 75.0 | Improved |
| remove interaction from ANPL (one-shot) vs vanilla GPT one-shot (ablation) | ANPL · ARC 400 tasks | solving rate (%): 16.8 -> 23.5 | Improved |
| allow interaction in ANPL vs one-shot ANPL (ablation) | ANPL · ARC 400 tasks | solving rate (%): 23.5 -> 75.0 | Improved |
| allow interaction with ChatGPT vs one-shot GPT (ablation) | GPT-3.5-turbo · ARC 400 tasks | solving rate (%): 16.8 -> 58.4 | Improved |
| ANPL vs ChatGPT when users cannot see Python code (beginners) | GPT-3.5-turbo · ARC beginner setting, 40 problems | solving rate (%): 37.2 -> 89.9 | Improved |
| ANPL vs ChatGPT in lines of code written | ANPL · ARC 400 tasks | mean lines of code: 26.47 -> 10.29 | Improved |
| ANPL vs ChatGPT in time waiting for GPT | ANPL · ARC 400 tasks | waiting time (%): 23.5 -> 14.6 | Improved |
| ANPL vs ChatGPT in variation across users | ANPL · ARC per-user solving rate | variance and min rate: 0.022 -> 0.038 | Hurt |
| use ANPL for MAGIC card game and LS-8 CPU projects | ANPL · MAGIC and LS-8 CPU | lines of code reduction (%): 67 | Improved |
| use ANPL first (order A-B vs B-A) before ChatGPT | ChatGPT (System B) · ARC order effect | solving rate (%): 54.5 -> 62.5 | Improved |
| order of systems on ANPL | ANPL · ARC order effect | solving rate (%): 74.5 -> 75.5 | No clear effect |
Closest papers by meaning
- CodeChain: Towards Modular Code Generation Through Chain of Self-revisions with Representative Sub-modules (ICLR 2024 · similarity 0.87)
- BigCodeBench: Benchmarking Code Generation with Diverse Function Calls and Complex Instructions (ICLR 2025 · similarity 0.86)
- Is Programming by Example Solved by LLMs? (NeurIPS 2024 · similarity 0.86)
- CodePlan: Unlocking Reasoning Potential in Large Language Models by Scaling Code-form Planning (ICLR 2025 · similarity 0.85)
- CRAFT: Customizing LLMs by Creating and Retrieving from Specialized Toolsets (ICLR 2024 · similarity 0.85)
- LILO: Learning Interpretable Libraries by Compressing and Documenting Code (ICLR 2024 · 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.
