Route-Verify-Vote Improves Mixed-Domain Reasoning Without Model Updates

Summary

Compositional generalization remains difficult when language models must combine familiar reasoning operations in unfamiliar ways. The SCoRE 2026 evaluation tests this problem across three mixed domains that are absent from training and requires models to identify the complete set of correct options for each question. The paper introduces Route-Verify-Vote (RVV), a training-free framework that uses a supplied domain label to choose a reasoning procedure, checks each option against the resulting constraints, and aggregates complete answer sets. Its voting stage allocates extra samples to questions where the two most common answer sets have a small margin, while keeping samples for a question on the same domain-specific procedure. On the official test set, voting across 16 sampled answer sets achieved 74.6% exact-set accuracy. Adaptive RVV raised this to 77.3%, and combining models on selected domain routes reached 79.4%. The final system ranked second among participating systems. The results suggest that domain-specific procedures and disagreement among answer sets can guide inference-time computation for mixed-domain reasoning.