Solver-Aware Decomposition Improves Programming-by-Example Synthesis

A recent paper on arXiv (2608.03461) critiques the conventional method of decomposition in Programming-by-Example (PBE). In this framework, a task is divided into smaller subtasks by a decomposer, while a synthesizer creates programs for each subtask. Typically, decomposers are trained to replicate ground-truth (GT) subgoals, which often reflect the preferences of annotators. The authors contend that these GT decompositions may not correspond with the search behaviors of bounded solvers that have fixed inductive biases, resulting in subgoals that, although logically sound, are impractical for the solver. To remedy this, they introduce Solver-Aware Decomposition (SAD), a training method that combines supervised learning on GT subgoals with direct feedback from a static synthesizer. This paper, recently submitted, contributes to advancements in program synthesis and machine learning.