Research map / Large language models / Agents, code and math
Latent Execution for Neural Program Synthesis Beyond Domain-Specific Languages
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 |
|---|---|---|---|
| add latent executor (partial latent execution traces) (ablation) | LaSynth · Restricted C | generalization accuracy: 38.6% -> 55.2% | Helped |
| execute at every step rather than only regularizing the final state (partial latent execution) (ablation) | LaSynth · Restricted C | generalization accuracy: 42.9% -> 55.2% | Helped |
| add operation predictor (ablation) | LaSynth · Restricted C | generalization accuracy: 53.7% -> 55.2% | Helped |
| add attention over decoded program tokens (ablation) | LaSynth · Restricted C | generalization accuracy: 53.5% -> 55.2% | Helped |
| LaSynth vs RobustFill | program synthesizer · Restricted C | generalization accuracy: 37.6% -> 55.2% | Helped |
| increase RobustFill model size | RobustFill · Restricted C | generalization accuracy | No clear effect |
| retrain on programs synthesized by the model (1 then 2 iterations) | LaSynth · Karel | generalization accuracy: 86.04% -> 89.28% -> 89.36% | Helped |
| retrain on programs synthesized by the model (1 then 2 iterations) | LaSynth · Restricted C | generalization accuracy: 55.2% -> 56.0% -> 56.5% | Helped |
| learned latent executor vs interpreter-based Exec | LaSynth · Karel | generalization accuracy: 83.68% vs 86.04% | Mixed |
Closest papers by meaning
- Write, Execute, Assess: Program Synthesis with a REPL (NeurIPS 2019 · similarity 0.90)
- Latent Programmer: Discrete Latent Codes for Program Synthesis (ICML 2021 · similarity 0.89)
- HYSYNTH: Context-Free LLM Approximation for Guiding Program Synthesis (NeurIPS 2024 · similarity 0.89)
- Program Synthesis via Test-Time Transduction (NeurIPS 2025 · similarity 0.89)
- Training Language Models on Synthetic Edit Sequences Improves Code Synthesis (ICLR 2025 · similarity 0.87)
- From I/O to Code with Discovery Agent (NeurIPS 2026 · 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.