Research map / Reinforcement learning / Visual and representation RL
Improving the Universality and Learnability of Neural Programmer-Interpreters with Combinator Abstraction
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 combinator abstraction with a fixed core rather than learning a growing set of individual programs | CNPI-4 versus RNPI-4x2 and RNPI-4x3 · Combinator auxiliary tasks, final stage | success rate (%): 91 vs 0 | Helped |
| Add adaptive sampling rather than uniform sampling | CNPI core and combinators · Three combinator auxiliary tasks, gradual curriculum | success rate (%): 49 -> 91 | Helped |
| Add a gradual curriculum rather than no curriculum | CNPI core and combinators with adaptive sampling · Three combinator auxiliary tasks | success rate (%): 31 -> 91 | Helped |
| Initialize the LSTM state with the combinator embedding instead of feeding the embedding at every time step (ablation) | CNPI core · Full combinator set (57 combinators) | prediction accuracy (%) | Helped |
| Keep the core fixed while adding new combinators | CNPI-4 · Full combinator set, old combinators after learning a new set | test accuracy (%): 100 vs 12.9, 3.3, 6.5, 1.7 | Helped |
| Learn individual programs with RNPI instead of reusing combinators | RNPI-4x2 and RNPI-4x3 · Combinator auxiliary tasks, final stage | success rate (%): 91 vs 0 | Hurt |
| Use adaptive sampling and a gradual curriculum instead of uniform sampling and no curriculum | CNPI core and combinators · Three combinator auxiliary tasks | success rate (%): 7 -> 91 | Helped |
Closest papers by meaning
- Representing Partial Programs with Blended Abstract Semantics (ICLR 2021 · similarity 0.88)
- Gradient-Based Program Synthesis with Neurally Interpreted Languages (ICLR 2026 · similarity 0.88)
- Making Neural Programming Architectures Generalize via Recursion (ICLR 2017 · similarity 0.88)
- Neural Program Lattices (ICLR 2017 · similarity 0.88)
- Interpreting Neural Combinatorial Optimization via Evolving Programmatic Bottlenecks (NeurIPS 2026 · similarity 0.86)
- HYSYNTH: Context-Free LLM Approximation for Guiding Program Synthesis (NeurIPS 2024 · similarity 0.84)
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.