SELFDOUBT: Uncertainty Quantification for Reasoning LLMs via the Hedge-to-Verify Ratio

Uncertainty estimation for reasoning language models remains difficult to deploy: sampling-based methods are expensive, while common single-pass proxies such as verbalized confidence or trace length are inconsistent across models. The problem is sharper for proprietary reasoning APIs, which expose neither logits nor reasoning-token probabilities, leaving practitioners with few reliable online signals for failure detection. We propose SelfDoubt, a single-pass trace diagnostic that extracts behavioral uncertainty signals directly from the reasoning trajectory. Its key signal, the Hedge-to-Verify Ratio (HVR), measures whether expressed doubt is offset by explicit self-checking. Unlike methods that require multiple sampled traces or model internals, SelfDoubt operates on one observed trace, making it suitable for latency- and cost-constrained deployment. Across seven models and three multi-step reasoning benchmarks (BBH, GPQA-Diamond, and MMLU-Pro), traces with no hedging markers are correct 96% of the time, yielding a high-precision zero-cost accept gate. On the remaining cases, the full SelfDoubt score significantly outperforms sampling-based Semantic Entropy at 10x lower inference cost. A deployment cascade combining both stages attains 90% accuracy at 71% coverage without task-specific labels. These results position SelfDoubt as a lightweight trace-level failure diagnostic for reasoning models and a natural building block for future agentic evaluation.