Explainable Program Synthesis by Localizing Specifications
summary
Introduces subspecifications that summarize what a subexpression must satisfy given a global spec, provides a two stage algorithm that constructs then iteratively simplifies them, and ships an implementation called S3. Evaluations include a 20‑participant user study showing about a 91% improvement in answer accuracy with subspecs, compression results where two thirds of cases compress >80% and most DreamCoder cases compress >50%, and timing where 80% of subspecs synthesize as fast or faster than the original task.
AI-written summary · about
chapters (11)
- 0:00Explainable program synthesis
- 1:02Understanding synthesized code
- 3:28Sub-specifications concept
- 5:58Sub-specifications insights
- 7:00User study results
- 9:06Algorithm for sub-specifications
- 10:54Summary and future work
- 11:40Q&A: Applicability limits
- 12:45Q&A: Readability tips
- 13:46Q&A: Complex specs handling
- 14:07Q&A: User study stats
related talks
47mBig Ideas in Program Synthesis
38mProgram Synthesis: A Dream Realized?
40mType-Driven Program Synthesis
14mGeneralizable Synthesis Through Unification
16mJust-in-Time Learning for Bottom-Up Enumerative Synthesis
16mFeedback-Driven Semi-supervised Synthesis of Program Transformations
17mSynthesizing MILP Constraints for Efficient and Robust Optimization
15mAbstraction Refinement-guided Program Synthesis for Robot Learning from Demonstrations