Reproduce “Neuro-Symbolic Program Synthesis” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Emilio Parisotto, Abdel-rahman Mohamed, Rishabh Singh, Lihong Li, Dengyong Zhou, Pushmeet Kohli
Neuro-Symbolic Program Synthesis (NSPS) integrates neural networks with symbolic program generation to synthesize explicit, interpretable programs from input-output examples. This approach, developed by Microsoft Research, demonstrates 91% accuracy on synthesizing unseen string transformation programs and solves 82.7% of real-world FlashFill benchmarks within its trained complexity.
Key insights
- Neural networks can learn to generate explicit, structured symbolic programs, bridging the gap between neural program induction and symbolic program synthesis.
- A recursive-reverse-recursive neural network (R3NN) that explicitly builds program parse trees is significantly more effective than linearized sequence generation for program synthesis.
- Learning to extract relevant features from input-output examples using specialized I/O encoders (e.g., cross-correlation based) helps guide the program synthesis process.
- Sampling multiple candidate programs from the learned distribution can dramatically improve the success rate for synthesizing unseen programs.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.