reasoningfactbullishSymbol-equivariant RRMs exhibit improved extrapolation to larger problem sizes compared to standard recurrent reasoning modelsArtificial Intelligence27 Jul 2026http://arxiv.org/abs/2607.02491v1