New technique fixes logical inconsistencies in AI systems answering true/false/unknown questions by checking negations and using targeted probes

CGD-PD addresses two key failure modes in three-way logical QA: negation inconsistency (where answers to H and ¬H contradict each other) and epistemic Unknown predictions caused by model uncertainty

The lightweight test-time method queries a classifier on both a hypothesis and its mechanically negated form, then projects answers onto logically consistent decisions

Uses proof-driven disambiguation with binary entailment probes to selectively resolve Unknown outcomes, requiring only 4-5 average model calls

Demonstrates improvements on FOLIO benchmark's first-order-logic tasks, showing potential for more reliable logical reasoning in large language models