Research map / Learning theory and algorithms / Combinatorial and search optimization
Learning to Combine Per-Example Solutions 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 |
|---|---|---|---|
| Aggregate per-example programs with the Cross Aggregator and include PE solution scores (ablation) | N-PEPS · E1, length 4 test programs | success ratio: 86.22 ± 0.25 -> 87.07 ± 0.28 | Helped |
| Use global-set execution states rather than the default PE execution states for keys (ablation) | N-PEPS · E1, length 4 test programs | success ratio: 87.07 ± 0.28 -> 85.19 ± 0.26 | Hurt |
| Add PE solution scores to the global-set key variant (ablation) | N-PEPS-PG · E1, length 4 test programs | success ratio: 85.19 ± 0.26 -> 85.94 ± 0.26 | Helped |
| Use only examples satisfied by each PE solution for key construction instead of global-set execution (ablation) | N-PEPS · E1, length 4 test programs | success ratio: 85.19 ± 0.26 -> 85.97 ± 0.26 | Helped |
| Add PE solution scores to the PE-only key variant (ablation) | N-PEPS-PP · E1, length 4 test programs | success ratio: 85.97 ± 0.26 -> 86.21 ± 0.27 | Helped |
| Add Cross Aggregator attention to direct aggregation of per-example cues (ablation) | PEPS · E1 and E2 program-length test settings | success ratio | Helped |
| Use per-example solutions instead of only global program search | N-PEPS · E1 and E2 program-length test settings | success ratio | Helped |
| Combine multiple per-example solutions rather than rely on individual solutions | PEPS · E1, length 4 test split | success ratio: ind(2): 95.4; tot(2): 97.6 | Helped |
| Increase the PEPS timeout | N-PEPS · E2 validation set | success ratio | Mixed |
| Increase the contribution from the Cross Aggregator relative to GPS | N-PEPS · E2 validation set | success ratio | Helped |
Closest papers by meaning
- Selecting Representative Examples for Program Synthesis (ICML 2018 · similarity 0.88)
- Generating Pragmatic Examples to Train Neural Program Synthesizers (ICLR 2024 · similarity 0.88)
- Solver-Aware Decompositions for Programming-by-Example: When Dividing Requires Knowing how to Conquer (NeurIPS 2026 · similarity 0.85)
- ExeDec: Execution Decomposition for Compositional Generalization in Neural Program Synthesis (ICLR 2024 · similarity 0.85)
- Learning to Compile Programs to Neural Networks (ICML 2024 · similarity 0.85)
- 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.