Research map / Architectures and efficiency / RNNs and normalization
RobustFill: Neural Program Learning under Noisy I/O
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 |
|---|---|---|---|
| use attention instead of the basic sequence-to-sequence architecture (ablation) | RobustFill synthesis · FlashFillTest | generalization accuracy: 51% -> 83% (Beam=100); 56% -> 86% (Beam=1000) | Helped |
| add double attention (Attention-B) to Attention-A (ablation) | RobustFill synthesis · FlashFillTest | generalization accuracy: roughly 2-5% absolute accuracy | Helped |
| use bidirectional encoders (Attention-C) (ablation) | RobustFill synthesis · FlashFillTest | generalization accuracy: roughly 2-5% absolute accuracy | Helped |
| prune partial programs whose outputs are not prefixes of observed outputs (DP-Beam) (ablation) | RobustFill synthesis · FlashFillTest | generalization accuracy: roughly 5% | Helped |
| use synthesis rather than direct neural induction | RobustFill · FlashFillTest | all-example generalization accuracy: 53% -> 81% | Helped |
| train with synthetic character noise and select programs by character edit rate rather than exact match (ablation) | RobustFill synthesis · FlashFillTest with noisy observed examples | accuracy: 80% vs. 6% | Helped |
| use a hand-engineered exact-string-matching synthesis system rather than the neural model under noisy I/O | FlashFill · FlashFillTest with noisy observed examples | accuracy: 80% vs. 6% | Hurt |
| use GetSpan() in the DSL | RobustFill synthesis · FlashFillTest | instances solvable by the DSL: approximately 20% | Helped |
Closest papers by meaning
- Leveraging Grammar and Reinforcement Learning for Neural Program Synthesis (ICLR 2018 · similarity 0.89)
- PLANS: Neuro-Symbolic Program Learning from Videos (NeurIPS 2020 · similarity 0.89)
- Learning to Compile Programs to Neural Networks (ICML 2024 · similarity 0.87)
- Program Synthesis via Test-Time Transduction (NeurIPS 2025 · similarity 0.87)
- Generating Pragmatic Examples to Train Neural Program Synthesizers (ICLR 2024 · similarity 0.86)
- HYSYNTH: Context-Free LLM Approximation for Guiding Program Synthesis (NeurIPS 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.