Research map / Architectures and efficiency / RNNs and normalization
Automatic Program Synthesis of Long Programs with a Learned Garbage Collector
Amit Zohar, Lior Wolf
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 PCCoder instead of DeepCoder | DSL program synthesis · program length 14 | total solved: 2.0% -> 60.4% | Improved |
| remove auxiliary function prediction task (ablation) | PCCoder · DSL program length 8 | total solved: 83% -> 70% | Hurt |
| remove variable dropping and randomly discard variables instead (ablation) | PCCoder · DSL program length 8 | total solved: 83% -> 77% | Hurt |
| remove variable dropping and function prediction tasks (ablation) | PCCoder · DSL program length 8 | total solved: 83% -> 66% | Hurt |
| learn garbage collection after fixing the representation (ablation) | PCCoder · DSL program length 8 | total solved: 83% -> 79% | Hurt |
| compute garbage collection separately rather than as an auxiliary task (ablation) | PCCoder · DSL program length 8 | total solved: 83% -> 80% | Hurt |
| replace dense block with 3 feed-forward layers (ablation) | PCCoder · DSL program length 8 | total solved: 83% -> 73% | Hurt |
| replace dense block with ResNet blocks (ablation) | PCCoder · DSL program length 8 | total solved: 83% -> 76% | Hurt |
| replace CAB with DFS search (ablation) | PCCoder · DSL program length 8 | total solved: 83% -> 67% | Hurt |
| train and test using 10 examples rather than 5 (ablation) | PCCoder · DSL program length 8 | total solved: 83% -> 78% | Hurt |
| train with 10 examples but provide 5 at test time (ablation) | PCCoder_Ten · DSL program length 8 | total solved: 78% -> 84% | Improved |
| use PCCoder without intermediate variable values (ablation) | PCCoder · DSL program length 8 | total solved: 83% -> 4% | Hurt |
Closest papers by meaning
- RobustFill: Neural Program Learning under Noisy I/O (ICML 2017 · similarity 0.87)
- Leveraging Grammar and Reinforcement Learning for Neural Program Synthesis (ICLR 2018 · similarity 0.87)
- Execution-Guided Neural Program Synthesis (ICLR 2019 · similarity 0.86)
- Adaptive Neural Compilation (NeurIPS 2016 · similarity 0.86)
- Program Synthesis via Test-Time Transduction (NeurIPS 2025 · similarity 0.85)
- Learning to Compile Programs to Neural Networks (ICML 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.
