To Guide or Not to Guide: Sparse Transductive Guidance in Program Synthesis

Janis Zenkner, Tobias Sesterhenn, Christian Bartelt

Keywords: program synthesis, inductive reasoning, transductive reasoning, programming-by-example

TL;DR: Program synthesis benefits from combining transductive guidance (direct output prediction) and inductive synthesis (program generation). We find that the key isn't combining them, but when to apply guidance. Exploiting this, TIIPS yields SOTA results

Abstract:

Program synthesis faces the dual challenge of achieving high success rates while maintaining interpretability and generalization, motivating hybrid approaches that combine complementary learning paradigms.
Integrating transductive methods, which provide strong predictive power by directly mapping inputs to outputs but lack interpretability, with inductive methods, which excel at producing explicit and interpretable programs, creates a new opportunity for programming-by-example.
While recent work has explored this integration through transductive guidance, we show, that permanent transductive guidance can, and in practice does, mislead search by overriding inductive reasoning strategies that would otherwise succeed.
To address this limitation, we introduce TIIPS, a novel framework that, for the first time, applies transductive assistance sparsely and selectively to inductive synthesis.
TIIPS adopts a teacher-student paradigm, where guidance is provided selectively, activated only when inductive synthesis fails, thereby preserving the natural problem-solving capabilities of inductive approaches.
Experiments on two standard programming-by-example domains (string and list manipulation) demonstrate that TIIPS outperforms related work, solving more tasks and producing more robust solutions, particularly under distribution shifts.
These results show that the timing and extent of transductive guidance matter more than its mere presence, establishing them as key factors for robust, interpretable, and effective program synthesis.

Primary Area: neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)

Submission Number: 3107