Research map / Large language models / Agents, code and math
Learning to superoptimize programs
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 |
|---|---|---|---|
| Replace uniform proposals with a learned unconditioned bias distribution | Stoke · Hacker's Delight | average relative score: 57.01% -> 36.45 % | Helped |
| Replace uniform proposals with a learned program-conditioned MLP distribution | Stoke · Hacker's Delight | average relative score: 57.01% -> 35.96 % | Helped |
| Condition the learned proposal distribution on input-program features rather than using an unconditioned bias | Learned proposal distribution · Hacker's Delight | average relative score: 36.45 % -> 35.96 % | No clear effect |
| Replace uniform proposals with a learned unconditioned bias distribution | Stoke · Automatically generated programs | average relative score: 76.63% -> 61.81% | Helped |
| Replace uniform proposals with a learned program-conditioned MLP distribution | Stoke · Automatically generated programs | average relative score: 76.63% -> 60.13% | Helped |
| Condition the learned proposal distribution on input-program features rather than using an unconditioned bias | Learned proposal distribution · Automatically generated programs | average relative score: 61.81% -> 60.13% | Helped |
| Use a parameterized categorical proposal instead of a uniform proposal | MCMC search · MCMC proposal distribution | iterations per second: 60 000 /second -> 20 000 /second | Hurt |
Closest papers by meaning
- Simple Baselines are Competitive with Code Evolution (NeurIPS 2026 · similarity 0.84)
- Reinforcement Learning and Data-Generation for Syntax-Guided Synthesis (AAAI 2024 · similarity 0.83)
- Training Language Models on Synthetic Edit Sequences Improves Code Synthesis (ICLR 2025 · similarity 0.83)
- Program Synthesis with Best-First Bottom-Up Search (Abstract Reprint) (AAAI 2024 · similarity 0.83)
- Program Synthesis via Test-Time Transduction (NeurIPS 2025 · similarity 0.83)
- PEARL: Solver-in-the-Loop Interactive Optimization Modeling from Natural Language (NeurIPS 2026 · similarity 0.83)
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.