Research map / Large language models / Agents, code and math
Code Driven Planning with Domain-Adaptive Selector
Zikang Tian, Shaohui Peng, Di Huang, Jiaming Guo, Ruizhi Chen, Rui Zhang, Xishan Zhang, Yuxuan Guo, Zidong Du, Qi Guo, Ling Li, Yewen Pu and 2 more
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 programs evolution with history summarization (ablation) | CoPiC · ALFWorld Clean/Heat/Cool average | success rate: 75.60% -> 91.44% -> 100.00% | Improved |
| add the RL-finetuned domain-adaptive selector instead of random plan choice (ablation) | CoPiC · ALFWorld Heat and Cool | success rate: 20% to 60% higher | Improved |
| replace GPT-4.1 LLM-as-a-Judge with the RL-finetuned selector (ablation) | CoPiC · ALFWorld | success rate and token cost: +19.88% success rate, -27.14% token consumption | Improved |
| use multiple planning programs (n=2 to 4) instead of a single program (ablation) | CoPiC · StarCraft II Easy | success rate: 0.67 -> 1.00 | Improved |
| use multiple planning programs (n=2 to 4) instead of a single program (ablation) | CoPiC · StarCraft II Medium and Hard | success rate: 0.33 -> 1.00 | Improved |
| increase planning programs from n=3 to n=4 beyond optimum (ablation) | CoPiC · StarCraft II Hard | token cost: 0.04M -> 0.08M | Hurt |
| increase number of planning programs n from 2 to 5 (ablation) | CoPiC · StarCraft II SuperHard | success rate: 32.00 -> 96.00 | Improved |
| increase number of planning programs n from 5 to 6 (ablation) | CoPiC · StarCraft II SuperHard | success rate: 96.00 -> 92.00 | No clear effect |
| add Pick-task interaction experience to the initial prompt (ablation) | CoPiC · ALFWorld Clean, Heat, Cool | success rate of initial programs: 38.06 -> 43.87; 39.13 -> 53.91; 0.00 -> 14.29 | Improved |
| online learning on the test set (TSL) instead of learning on training tasks (ablation) | CoPiC · ALFWorld Heat | success rate: 100.00 -> 96.52 | No clear effect |
| vary LLM temperature from 0.2 to 0.8 (0.6 default) (ablation) | CoPiC · ALFWorld Examine | success rate: 90.00 (0.2), 95.56 (0.5), 100.00 (0.6), 100.00 (0.8) | Mixed |
| swap GPT-3.5 for open-source LLMs (DeepSeek-Coder, Qwen2.5-Coder-14B) (ablation) | CoPiC · ALFWorld average | success rate: 100.00 (DeepSeek-Coder), 99.76 (Qwen2.5-Coder-14B-Instruct) | No clear effect |
Closest papers by meaning
- Large Language Models as Planning Domain Generators (Student Abstract) (AAAI 2024 · similarity 0.90)
- Leveraging Environment Interaction for Automated PDDL Translation and Planning with Large Language Models (NeurIPS 2024 · similarity 0.89)
- SayCanPay: Heuristic Planning with Large Language Models Using Learnable Domain Knowledge (AAAI 2024 · similarity 0.89)
- Plan-and-Act: Improving Planning of Agents for Long-Horizon Tasks (ICML 2025 · similarity 0.88)
- LLMs Can Plan Only If We Tell Them (ICLR 2025 · similarity 0.88)
- SPIRAL: Symbolic LLM Planning via Grounded and Reflective Search (AAAI 2026 · similarity 0.88)
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.
