Feedback-Driven Semi-supervised Synthesis of Program Transformations
summary
Semi-supervised synthesis combines input-output examples with extra unlabeled inputs (cursor locations and mined candidates) and uses anti-unification modulo provenance to infer outputs for those inputs, then invokes Refazer to synthesize transformations. Implemented as a Visual Studio plugin called Blue-pencil and integrated with IntelliCode, the feedback loop automatically adds missed inputs and refines programs, raising recall from 27% to 100% on evaluated edits and, across 295 cursor scenarios, yielding 1 false positive and 3 false negatives.
AI-written summary · about
chapters (10)
- 0:00Introduction
- 1:10Demo
- 7:13Additional Input
- 8:00Synthesis
- 8:44Provenance
- 9:22Antiunification
- 10:31Unification
- 11:50User Feedback
- 13:45Experiments
- 15:00Conclusion
related talks
38mProgram Synthesis: A Dream Realized?
47mBig Ideas in Program Synthesis
40mType-Driven Program Synthesis
18mActive Learning for Neurosymbolic Program Synthesis
29mSyntia: Synthesizing the Semantics of Obfuscated Code
15mAbstraction Refinement-guided Program Synthesis for Robot Learning from Demonstrations
37mMANTIS: Semantics-driven Inductive Program Synthesis
26mData-Driven Lemma Synthesis for Interactive Proofs