Research map / Large language models / Agents, code and math
Parsel🐍: Algorithmic Reasoning with Language Models by Composing Decompositions
Eric Zelikman, Qian Huang, Gabriel Poesia, Noah Goodman, Nick Haber
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 |
|---|---|---|---|
| Decompose tasks into Parsel functions and search combinations of their implementations with tests | Code generation on APPS competition-level · APPS competition-level | pass rate: 14.5% to 25.5% | Improved |
| Translate high-level plans directly to Python instead of using Parsel synthesis (ablation) | Codex code generation · APPS competition-level, 100 randomly sampled problems | pass rate: 25.5% to 6% | Hurt |
| Use Parsel decomposition and synthesis rather than direct GPT-4 generation | GPT-4 code generation · HumanEval | pass@1: 67% to 85% | Improved |
| Use Parsel with generated tests and CodeT selection instead of GPT-4 with CodeT alone | GPT-4 code generation · HumanEval, generated tests | pass rate: 81.1% to 85.1% | Improved |
| Generate combinations of function implementations with Parsel rather than direct programs | GPT-4 code generation · HumanEval, pass@any | pass rate: 82.3% to 91.5%; 84.1% to 93.9% | Improved |
| Present plans generated with Parsel rather than the baseline planning approach | VirtualHome robotic planning · VirtualHome, human accuracy rankings | comparisons preferred over baseline: 69% | Improved |
| Generate plans with Parsel rather than the baseline planning approach | VirtualHome robotic planning · VirtualHome, human clarity rankings | relative likelihood of preference: 70% more likely | Improved |
| Show indented function names alongside step-by-step output (ablation) | Parsel robotic plans · VirtualHome, human accuracy and clarity rankings | preference relative to non-indented Parsel | No clear effect |
| Increase the number of program evaluations for generated function combinations | Parsel · APPS competition-level | pass rate | Improved |
| Use Parsel composition to combine function implementations instead of sampling only complete programs | Code generation on chained-component tasks · Chained components | problems solved: 1x2; 16 programs | Improved |
| Use GPT-4 alone instead of GPT-4 with Parsel (ablation) | GPT-4 code generation · HumanEval, pass@any | problems solved: two problems | Mixed |
Closest papers by meaning
- BigCodeBench: Benchmarking Code Generation with Diverse Function Calls and Complex Instructions (ICLR 2025 · similarity 0.88)
- Steering Large Language Models between Code Execution and Textual Reasoning (ICLR 2025 · similarity 0.88)
- Benchmarking Large Language Models with Integer Sequence Generation Tasks (NeurIPS 2025 · similarity 0.88)
- Reasoning-as-Logic-Units: Scaling Test-Time Reasoning in Large Language Models Through Logic Unit Alignment (ICML 2025 · similarity 0.88)
- THOR: Tool-Integrated Hierarchical Optimization via RL for Mathematical Reasoning (ICLR 2026 · similarity 0.88)
- Scaling Code-Assisted Chain-of-Thoughts and Instructions for Model Reasoning (NeurIPS 2025 · 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.
