Program Synthesis: A Dream Realized?
summary
Claims a semantics-guided pipeline that augments user examples with relational properties and feeds them to existing inductive synthesizers; implements a perturbation engine and SketchAX on top of Sketch and evaluates against benchmarks, showing roughly a 60% improvement (baseline ~50% → ~80% success, bitvector cases much larger). Also adapts idea to LambdaSquare and explores permutation invariance and other semantic biases via augmented examples and partial max SMT property inference.
AI-written summary · about
chapters (14)
- 0:00Introduction to synthesis
- 1:04Defining program synthesis
- 3:24Inductive synthesis approach
- 5:08Synthesis as search
- 7:06Challenges of correctness
- 9:53Semantic biases in synthesis
- 12:56Project Mantis overview
- 14:42Property-based augmentation
- 17:20Applying semantic biases
- 19:43User interfaces for properties
- 24:02Evaluation of sketch-ax
- 27:36Functional programming domain
- 29:46Future research directions
- 32:36Q&A and conclusion
related talks
37mMANTIS: Semantics-driven Inductive Program Synthesis
47mBig Ideas in Program Synthesis
01h 10mNeurosymbolic Program Synthesis: Bridging Perception and Reasoning in Real-World(…)
16mFeedback-Driven Semi-supervised Synthesis of Program Transformations
29mSyntia: Synthesizing the Semantics of Obfuscated Code
40mType-Driven Program Synthesis
15mGauss: Program Synthesis by Reasoning Over Graphs
18mActive Learning for Neurosymbolic Program Synthesis