Research map / Reinforcement learning / Visual and representation RL
Learning to Synthesize Programs as Interpretable and Generalizable Policies
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 |
|---|---|---|---|
| learn a program embedding space and search it with CEM instead of synthesizing tokens from scratch (ablation) | LEAPS · Program behavior reconstruction | avg reward: 0.56 -> 0.84 | Helped |
| add latent behavior reconstruction loss L to the VAE loss (ablation) | LEAPS-P · Program behavior reconstruction | avg reward: 0.67 -> 0.75 | Helped |
| add program behavior reconstruction loss R to the VAE loss (ablation) | LEAPS-P · Program behavior reconstruction | avg reward: 0.67 -> 0.72 | Helped |
| add both R and L losses (full objective) (ablation) | LEAPS-P · Program behavior reconstruction | avg reward: 0.67 -> 0.84 | Helped |
| use CEM instead of random sampling of 64 latents (ablation) | LEAPS · Program behavior reconstruction | avg reward: 0.58 -> 0.84 | Helped |
| use CEM instead of random sampling of 64 latents (ablation) | LEAPS · Karel CleanHouse | mean return: 0.03 -> 0.18 | Helped |
| use CEM instead of random sampling of 64 latents (ablation) | LEAPS · Karel TopOff | mean return: 0.33 -> 0.81 | Helped |
| add latent behavior reconstruction loss L (ablation) | LEAPS-P · Embedding smoothness, testing | R_mat of 10 nearest programs: 0.22 -> 0.28 | Helped |
| add program behavior reconstruction loss R (ablation) | LEAPS-P · Embedding smoothness, testing | R_mat of 10 nearest programs: 0.22 -> 0.22 | No clear effect |
| program synthesis via LEAPS instead of DRL | Karel tasks · Karel Harvester | mean return: 0.90 -> 0.45 | Hurt |
| LEAPS programs instead of DRL when transferred to larger grid | Karel Maze 100x100 · Karel Maze 100x100 | mean return: 0.04 -> 1.00 | Helped |
Closest papers by meaning
- Programmatic Reinforcement Learning without Oracles (ICLR 2022 · similarity 0.91)
- Programmatically Interpretable Reinforcement Learning (ICML 2018 · similarity 0.91)
- Synthesizing Programmatic Policies that Inductively Generalize (ICLR 2020 · similarity 0.90)
- $\mathcal{B}$-Coder: Value-Based Deep Reinforcement Learning for Program Synthesis (ICLR 2024 · similarity 0.89)
- Synthesizing Programmatic Reinforcement Learning Policies with Large Language Model Guided Search (ICLR 2025 · similarity 0.89)
- Hierarchical Programmatic Option Framework (NeurIPS 2024 · 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.