Test‑time scaling methods usually assume that extra computation from the same model supplies independent evidence. This assumption breaks down inside a single reasoning model: errors are correlated across repeated samples, across reasoning depth, and in the model’s own self‑evaluation, so more compute from the same model does not provide more evidence.
To address this, the authors introduce DIA (Disagreement‑triggered Independent Arbitration). The key idea is to treat a disagreement between an intermediate‑stage prediction and the final‑stage prediction as a signal of reasoning instability for that question. When such a disagreement is detected, a small panel of heterogeneous models is asked to solve the question independently, without seeing any candidate answers. The final answer is obtained by directly aggregating these independent solutions.
Experiments on seven benchmarks and two base models show that DIA outperforms full chain‑of‑thought in 12 out of 14 settings, including a 7.00% gain on MuSiQue. DIA activates on only about 17% of examples and uses up to 47.2% fewer tokens than a multi‑agent baseline. Ablation studies confirm that both independent solving and solver heterogeneity are necessary for the observed improvements.
Review