Automated transpilation of imperative to functional code using neural-guided program synthe
summary
A tool called NGST2 implements a neural-guided synthesizer that converts imperative Java and Python snippets into Stream and functools pipelines by assuming trace-compatibility between source and target expressions. It uses a novel cognate grammar network to bias generation toward copied source subexpressions and a concolic, bi-directional pruning pass to eliminate impossible partial programs. On 91 GitHub and StackOverflow benchmarks NGST2 solves about 80% within five minutes while prior baselines stay below 20%, and ablations show both components are critical.
AI-written summary · about
chapters (15)
- 0:00Introduction
- 1:25Problem definition
- 3:11Efficient synthesis
- 4:56Cognate grammar network
- 8:27Trace compatibility pruning
- 15:07Benchmark results
- 17:56Ablation study
- 18:37Summary
- 19:01Q&A
- 22:11Cross-language applicability
- 23:20Unsolved benchmark challenges
- 24:24Benchmark complexity
- 25:10Formal correctness
- 26:12New domains
- 27:17Future directions
related talks
15mGauss: Program Synthesis by Reasoning Over Graphs
23mType-Preserving, Dependence-Aware Guide Generation for Sound, Effective Amortize...
45mKeynote: Writing Code? Pfft... Evolve it Instead!
14mNeural Reverse Engineering of Stripped Binaries using Augmented Control Flow Graphs
38mProgram Synthesis: A Dream Realized?
25mInductive Synthesis of Structurally Recursive Functional Programs from Non-recur...
18mActive Learning for Neurosymbolic Program Synthesis
01h 10mNeurosymbolic Program Synthesis: Bridging Perception and Reasoning in Real-World(…)