Predictable Compression Failures: Order Sensitivity and Information Budgeting for Evidence-Grounded Binary Adjudication
Researchers have developed a new method called Quantified Martingale Violation (QMV) to address order sensitivity in transformer models used for evidence-based decision-making. This approach aims to reduce unreliable answers by formalizing an expectation-realization gap, where training minimizes expected description length across evidence permutations while a fixed ordering remains position-sensitive. The method introduces metrics like Bits-to-Trust (B2T) and Risk-of-Hallucination (RoH) to help determine when a model should provide an answer or abstain, showing promising results on several datasets. AI
IMPACT Introduces a framework to improve reliability in AI systems that make decisions based on evidence, potentially reducing hallucinations.