ChainUQ Framework for Reasoning Consistency Uncertainty

ChainUQ introduces a framework to quantify LLM uncertainty by evaluating logical consistency across the entire reasoning chain rather than relying on token-level probabilities. It uses an alignment-aware lightweight module to estimate intrinsic confidence based on frozen features aligned to final conclusions.