CAST: Contrastive Adjudication via Selective Triggering for Efficient LLM Reasoning

Large language models (LLMs) often follow conflicting reasoning paths on difficult tasks, creating uncertainty that standard self-consistency (SC) addresses by voting over multiple chains without resolving the contradictions that produced it. Uniformly scaling inference-time compute by sampling more paths is expensive and, as we show, less effective than targeted allocation. We propose Contrastive Adjudication via Selective Triggering (CAST), a lightweight inference-time method that detects disagreement across chain-of-thought samples and triggers contrastive adjudication only when model confidence falls below a threshold. Instead of re-asking the same question, CAST extracts the strongest argument for each competing reasoning path, presents them side by side, and asks the model to resolve the conflict definitively, spending extra compute only where it is most needed, across three benchmarks spanning multiple-choice medical exams, evidence-based biomedical QA, and free-response clinical dialogue. CAST outperforms standard SC, compute-matched SC, SC Max (with the maximum possible number of calls, usually 13), and the pooled Contrastive Adjudication via Self-Consistency (CASC) variant across all datasets. CAST triggers on only 8\%--20\% of questions on multiple-choice benchmarks and averages roughly 1.2 additional generation calls per multiple-choice question, while gains on uncertain questions reach up to 29 percentage points over SC. These results show that selective compute allocation outperforms uniform sampling, making CAST a practical algorithm for efficient, uncertainty-aware LLM deployment.