Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Neural Program Meta-Induction
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 |
|---|---|---|---|
| Initialize task training from the best related-task portfolio model | Karel program induction · Karel, except when the new task has 10k+ examples | accuracy | Helped |
| Use meta-learning conditioned on demonstration I/O examples | Karel program induction · Karel with fewer than ten task examples | accuracy | Helped |
| Use portfolio adaptation rather than meta adaptation | Karel program induction · Karel with 1000+ task examples | accuracy | Helped |
| Fine-tune the meta model on the new task (ablation) | Meta Program Induction · Karel with small numbers of task examples | accuracy | No clear effect |
| Fine-tune the meta model on the new task (ablation) | Meta Program Induction · Karel with large numbers of task examples | accuracy | Helped |
| Increase the portfolio from one model to 25 models (ablation) | Portfolio-Adapted Program Induction · Karel, especially when n<100 | accuracy | No clear effect |
| Train portfolio models on more examples, d=100000 rather than d=1000 (ablation) | Portfolio-Adapted Program Induction · Karel, particularly with 50 to 100 new-task examples | accuracy | Helped |
| Mix new-task examples with random examples from other tasks during adaptation (ablation) | Adapted Meta Program Induction · Karel | held-out loss | Helped |
| Randomly reinitialize all weights before adapting the meta network (ablation) | Adapted Meta Program Induction · Karel with many new-task examples | accuracy | Helped |
| Increase model capacity from Small or Medium to Large (ablation) | Karel program induction · Karel across values of n | accuracy | Helped |
| Train the meta model on fewer background tasks (ablation) | Meta Program Induction · Karel new-task generalization | accuracy | Hurt |
Closest papers by meaning
- A large-scale benchmark for few-shot program induction and synthesis (ICML 2021 · similarity 0.87)
- Neuro-Symbolic Program Synthesis (ICLR 2017 · similarity 0.87)
- Meta-learning with an Adaptive Task Scheduler (NeurIPS 2021 · similarity 0.86)
- Meta-Learning without Memorization (ICLR 2020 · similarity 0.86)
- Gradient-Based Program Synthesis with Neurally Interpreted Languages (ICLR 2026 · similarity 0.86)
- Learning to Learn with Contrastive Meta-Objective (NeurIPS 2025 · 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.