G-RRM guides symbolic solvers with recurrent reasoning models for constraint satisfaction

The authors propose G-RRM, a neuro-symbolic approach that integrates symbol-equivariant recurrent reasoning models (SE-RRMs) with classical symbolic solvers to improve search efficiency for constraint satisfaction problems. SE-RRMs act as neural solvers generating solution proposals to guide backtracking or SAT-based methods like Glucose 4.1 and CaDiCaL 3.0.0.

• G-RRM efficacy requires expansive combinatorial search spaces and solver architectures capable of dynamically overwriting branching choices.
• On $9\times9$ Sudoku, where SE-RRMs solve 91.1% of instances, backtracking accelerates by 33.3x and Glucose 4.1 by 1.70x.
• Glucose 4.1 retains a 1.17x speedup on perfect-hint $25\times25$ grids.
• CaDiCaL 3.0.0 shows no significant speedup because it always respects injected hints rather than overwriting them.

These results delineate the specific regimes where neural guidance translates into practical wall-clock speedups for symbolic solvers.