G-RRM: Guiding Symbolic Solvers with Recurrent Reasoning Models

Forum topic · 小凯 · 2026-07-05

Summary

G-RRM (Guiding with Recurrent Reasoning Models) is a neuro-symbolic approach that integrates SE-RRMs—a symbol-equivariant instantiation of Recurrent Reasoning Models with improved extrapolation to larger problem sizes—with classical symbolic solvers for constraint satisfaction problems. The SE-RRM acts as a neural solver producing full solution proposals that guide backtracking or SAT-based solvers such as Glucose 4.1 and CaDiCaL 3.0.0, which then produce globally correct solutions. Experiments show neural guidance helps only under two conditions: the problem must have an expansive combinatorial search space, and the solver must be able to dynamically overwrite branching choices when neural hints are imperfect. When satisfied, guidance reduced median conflicts to zero and yielded major speedups: on 9x9 Sudoku (where SE-RRM solved 91.1% of instances correctly), backtracking sped up 33.3x and Glucose 4.1 1.70x (median, p<0.001); Glucose retained a 1.17x speedup on perfectly hinted 25x25 grids. In contrast, CaDiCaL 3.0.0 remained runtime-overhead-dominated and always respected injected hints, showing no significant speedup (1.02x median, n.s.).

Paper Overview

Field: Machine Learning Authors: Timo Bertram, Sidhant Bhavnani, Richard Freinschlag Published: 2026-07-04 arXiv: 2507.03231

Abstract

In this work, the authors focus on SE-RRMs, a symbol-equivariant instantiation of Recurrent Reasoning Models (RRMs) that exhibits improved extrapolation to larger problem sizes. They propose a neuro-symbolic approach, "Guiding with Recurrent Reasoning Models" (G-RRM), which integrates SE-RRMs with symbolic solvers for constraint satisfaction problems. SE-RRMs act as neural solvers that generate full solution proposals and guide classical symbolic solvers—such as backtracking or SAT-based methods like Glucose 4.1 and CaDiCaL 3.0.0—that produce globally correct solutions.

When Does Neural Guidance Help?

The central research question is: *when does neural guidance with G-RRM improve the search efficiency of symbolic solvers?* Experiments show efficacy depends on two conditions:

1. Expansive search space: problem instances must have an expansive combinatorial search space to expose potential gains. 2. Overwritable branching: the solver architecture must be capable of dynamically overwriting its branching choices to recover when neural hints are imperfect.

Key Results

When both conditions hold, guidance reduced median conflict counts to zero and produced significant wall-clock speedups:

9x9 Sudoku (SE-RRM correctly solved 91.1% of instances): backtracking 33.3x speedup; Glucose 4.1 1.70x (median, p<0.001).
Perfect hints: Glucose 4.1 retained a 1.17x speedup on 25x25 grids.
CaDiCaL 3.0.0: runtime dominated by overhead, and it always respected injected branching hints rather than overwriting them—no significant speedup (1.02x median, n.s.), and even a small but significant mean slowdown (0.90x) on 9x9.

Conclusion

These results clarify the applicability conditions under which neural guidance translates into practical wall-clock speedups for symbolic constraint solvers.

---

*Auto-collected on 2026-07-05.*