Research map / Probabilistic and scientific modeling / Molecules and proteins
Latent Programmer: Discrete Latent Codes for 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 |
|---|---|---|---|
| use learned discrete latent plans and two-level beam search instead of direct program decoding | Latent Programmer · String transformation (synthetic), B=10 | accuracy: 51\% -> 57$\%$ | Helped |
| use learned discrete latent plans and two-level beam search instead of direct program decoding | Latent Programmer · Python code generation, B=10 | BLEU: 15.5 -> 18.6 | Helped |
| use learned discrete latent plans and two-level beam search instead of direct program decoding | Latent Programmer · String transformation (modified DSL), B=100 | accuracy: 62\% -> 67\% | Helped |
| replace discrete VQ-VAE codes with continuous autoencoder codes (ablation) | Latent Programmer · String transformation (synthetic), B=10 | accuracy: 57$\%$ -> 50\% | Hurt |
| replace discrete VQ-VAE codes with continuous VAE codes (ablation) | Latent Programmer · String transformation (synthetic), B=10 | accuracy: 57$\%$ -> 51\% | Hurt |
| increase latent beam size from L=1 to L=3 (ablation) | Latent Programmer · String transformation (synthetic), B=10 | accuracy: 52$\%$ -> 57$\%$ | Helped |
| increase latent beam size from L=3 to L=10 (ablation) | Latent Programmer · String transformation (synthetic), B=10 | accuracy: 57$\%$ -> 56$\%$ | Hurt |
| increase latent beam size from L=1 to L=10 (ablation) | Latent Programmer · String transformation (synthetic), B=10 | distinct 4-grams: 0.28 -> 0.33 | Helped |
| use discrete latent planning rather than a single-level search (L=1) (ablation) | Latent Programmer · String transformation (synthetic), B=10 | accuracy: 52$\%$ -> 57$\%$ | Helped |
| use latent plans rather than direct decoding | Latent Programmer · String transformation by ground-truth program length | accuracy: 94.5$\%$ -> 94.0$\%$; 83.9$\%$ -> 84.6$\%$; 72.8$\%$ -> 72.2$\%$; 63.1$\%$ -> 66.1$\%$; 47.1$\%$ -> 49.8$\%$; 40.6$\%$ - | Mixed |
Closest papers by meaning
- Latent Execution for Neural Program Synthesis Beyond Domain-Specific Languages (NeurIPS 2021 · similarity 0.89)
- Searching Latent Program Spaces (NeurIPS 2025 · similarity 0.88)
- HYSYNTH: Context-Free LLM Approximation for Guiding Program Synthesis (NeurIPS 2024 · similarity 0.87)
- From I/O to Code with Discovery Agent (NeurIPS 2026 · similarity 0.86)
- Generating Pragmatic Examples to Train Neural Program Synthesizers (ICLR 2024 · similarity 0.86)
- Program Synthesis via Test-Time Transduction (NeurIPS 2025 · similarity 0.86)
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.