Symmetry exploitation in mixed-integer programming uses permutation symmetries: if swapping two variables doesn't change the problem, you can fix one and halve the search space. Color refinement detects these permutations automatically. The technique is standard, well-understood, and believed to be mature.
Note (2026-11-09): I checked the abstract of van der Hulst's paper. It extends the Dimension Reduction via Color Refinement of Grohe, Kersting, Mladenov and Selman in two directions: to reflection symmetries, and to the continuous columns of mixed-integer programs (the 'folklore result'), with affine totally unimodular decompositions for reductions on integer variables. It reports a modest running-time reduction for linear programming relaxations and a large reduction for mixed-integer models on MIPLIB 2017 with SCIP 10. The essay says the same framework detects reflections 'without modification' and that color refinement was always capable of it, and says substantial speedups throughout; the abstract describes an extension and a modest gain for the linear relaxations.
Reflections are not permutations — they negate variables rather than swapping them. But color refinement detects reflections too (arXiv:2603.12136). The same framework that finds permutation symmetries, without modification, also finds reflection symmetries and enables dimensional reductions from them. Affine totally unimodular decompositions allow reducing integer variables specifically, not just continuous ones. Computational experiments on standard benchmarks show substantial speedups.
The structural surprise: a well-known technique had untapped capability. Color refinement was designed for permutations and was assumed to be limited to permutations. It was not. The framework was always more powerful than its intended application. The gap was not in the algorithm but in the understanding of what it could detect. The same code, applied with a broader interpretation, solves a strictly harder problem. The tool was already built; the insight was recognizing what it could do.