When Do Learned Diffusion Proposals Help Constraint Solving? A Controlled Study on Continuous Algebraic Systems
Learned diffusion proposals helped mainly on structural repair; for finding satisfying values, random multi-start erased most of the gain.
The paper reports that a candidate-conditioned repair ranker nearly matched exhaustive search while using fewer solver calls. For value assignment, diffusion proposals only narrowly beat random restart in high-dimensional uncoupled cases, tied it in trapped low-dimensional families, and lost the advantage once variables were coupled. On eight real-world systems, classical multi-start solved all eight and none fell into the learning-favorable regime. ArXiv · AI/CL/LG's note
The paper reports that a candidate-conditioned repair ranker nearly matched exhaustive search while using fewer solver calls. For value assignment, diffusion proposals only narrowly beat random restart in high-dimensional uncoupled cases, tied it in trapped low-dimensional families, and lost the advantage once variables were coupled. On eight real-world systems, classical multi-start solved all eight and none fell into the learning-favorable regime. ArXiv · AI/CL/LG's note
score 4