New SELFDOUBT method enables single-pass uncertainty detection for reasoning LLMs without requiring model internals or multiple sampling passes

Addresses practical challenge of uncertainty estimation for reasoning language models, especially proprietary APIs that don't expose logits or token probabilities

Introduces Hedge-to-Verify Ratio (HVR) signal that analyzes uncertainty markers and self-checking behavior directly from a single reasoning trace

Eliminates computational expense of sampling-based methods while providing more reliable uncertainty signals than existing single-pass proxies like verbalized confidence or trace length

Works on a single observed reasoning trajectory, making it deployable across different models without requiring access to model internals