Inductive Program Synthesis Guided by Observational Program Similarity
summary
Introduces SyMetric, a tool that replaces observational equivalence with a domain-specific observational similarity metric to cluster programs into bounded-radius version spaces and then applies repair via distance-guided rewrites (hill-climbing over rewrite rules) to turn “close but wrong” representatives into exact solutions. Implements bottom-up enumeration with clustering, ranking, extraction, and incremental repair. Evaluated on inverse CSG (25 random + 22 handwritten benchmarks), regex inference, and a tower-building task; outperforms domain‑agnostic equivalence-based syntheses and matches or nears domain-specific baselines. metric program synthesis
AI-written summary · about
chapters (12)
- 0:00Inductive program synthesis intro
- 1:05Bottom-up synthesis recap
- 2:51Limitations of equivalence
- 3:53Extending to observational similarity
- 5:39Algorithm overview
- 7:25Parameterizing algorithm
- 8:27Inverse CSG results
- 9:31Results in regular expression domain
- 10:35Tower building results
- 11:17Ablation study
- 12:18Conclusion and future work
- 13:20Q&A
