Research map / Label-efficient and robust learning / Recognition and long-tail
Leveraging Grammar and Reinforcement Learning for Neural Program Synthesis
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Optimize correctness over beam-search candidates instead of reference likelihood | MLE program synthesizer · Karel full dataset | Generalization Top-1: 71.91 -> 75.72 | Helped |
| Optimize correctness over beam-search candidates instead of reference likelihood | MLE program synthesizer · Karel small dataset | Generalization Top-1: 12.58 -> 25.28 | Helped |
| Use REINFORCE expected correctness reward after supervised pretraining | MLE program synthesizer · Karel full dataset | Generalization Top-1: 71.91 -> 68.39 | Hurt |
| Use REINFORCE expected correctness reward after supervised pretraining | MLE program synthesizer · Karel small dataset | Generalization Top-1: 12.58 -> 0 | Hurt |
| Optimize correctness over beam-search candidates | MLE program synthesizer · Karel full dataset | Exact Match Top-1: 39.94 -> 8.21 | Hurt |
| Add diversity objective to RL_beam (ablation) | RL_beam · Karel full dataset | Generalization Top-5: 79.29 -> 82.09 | Helped |
| Add diversity objective to RL_beam (ablation) | RL_beam · Karel small dataset | Generalization Top-1: 25.28 -> 23.72 | Hurt |
| Add penalty for long-running programs to RL_beam_div (ablation) | RL_beam_div · Karel full dataset | Generalization Top-1: 76.20 -> 77.12 | Helped |
| Add penalty for long-running programs to RL_beam_div (ablation) | RL_beam_div · Karel full dataset | Generalization Top-50: 85.86 -> 85.38 | Hurt |
| Add learned syntax module to MLE (ablation) | MLE · Karel small dataset | Generalization Top-1: 12.58 -> 17.02 | Helped |
| Add learned syntax module to MLE (ablation) | MLE · Karel full dataset | Generalization Top-1: 71.91 -> 69.37 | Hurt |
| Add handwritten syntax checker to MLE (ablation) | MLE · Karel full dataset | Generalization Top-1: 71.91 -> 72.07 | Helped |
Closest papers by meaning
- Reinforcement Learning and Data-Generation for Syntax-Guided Synthesis (AAAI 2024 · similarity 0.90)
- HYSYNTH: Context-Free LLM Approximation for Guiding Program Synthesis (NeurIPS 2024 · similarity 0.89)
- Generating Pragmatic Examples to Train Neural Program Synthesizers (ICLR 2024 · similarity 0.88)
- Program Synthesis via Test-Time Transduction (NeurIPS 2025 · similarity 0.88)
- Online Prompt Selection for Program Synthesis (AAAI 2025 · similarity 0.87)
- $\mathcal{B}$-Coder: Value-Based Deep Reinforcement Learning for Program Synthesis (ICLR 2024 · 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.