MANTIS: Semantics-driven Inductive Program Synthesis
summary
MANTIS applies semantic biases by augmenting user examples with relational perturbation properties (e.g., permutation invariance, value preservation) and then hands the expanded example set to an existing synthesizer as a black box. An implementation called SketchAX uses a perturbation engine plus property inference via MaxSAT/SMT, three UIs (selection, validation, inference), and benchmarks over scalars, arrays and matrices. Results show up to a 60% lift in success rate (191% on hard subsets), 22% overall improvement, and an early SysLambda prototype yields ~26% gain over Lambda2.
AI-written summary · about
chapters (25)
- 0:00Introduction
- 0:27Classical Program Synthesis
- 1:48Inductive Program Synthesis
- 5:23Searchbased Program Synthesis
- 7:31Similarities with Inductive Learning
- 8:46Challenges of Inductive Learning
- 11:36Inductive Bias
- 13:32Next Big Idea
- 13:46Semantic Bias
- 14:37SemanticsDriven Inductive Program Synthesis
- 17:00Application of Semantic Bias
- 18:20Relational Permutation Properties
- 19:36Synthesis Strategy
- 20:08User Interfaces
- 22:35Property Selection
- 24:52Data
- 25:53Success Rate
- 26:50Performance
- 27:38Success Rates
- 28:37lambdasquare
- 29:12example
- 29:47cool problems
- 30:43future work
- 31:51student collaborators
- 32:19Q A
related talks
38mProgram Synthesis: A Dream Realized?
47mBig Ideas in Program Synthesis
29mSyntia: Synthesizing the Semantics of Obfuscated Code
40mType-Driven Program Synthesis
17mSynthesizing MILP Constraints for Efficient and Robust Optimization
16mFeedback-Driven Semi-supervised Synthesis of Program Transformations
01h 10mNeurosymbolic Program Synthesis: Bridging Perception and Reasoning in Real-World(…)
25mInductive Synthesis of Structurally Recursive Functional Programs from Non-recur...