Tri-ABL: Enhancing ABL Using Three Learners

Abductive learning (ABL) unifies machine learning and logical reasoning in a mutually beneficial loop. During training, a central challenge is ambiguity: multiple hypotheses can satisfy the background knowledge. Existing ABL variants either select the single candidate closest to the model’s prediction, which cannot guarantee correctness, or consider all candidates via expectation maximization, but degrade when reasoning cannot get complete candidate sets within realistic time budgets. Tri-ABL addresses this by training three learners to maintain minimal, incomplete, yet distinct abduction sets from complementary perspectives, thereby avoiding exhaustive abduction. We aggregate these sets via a tri-learner ensemble to construct supervision that amplifies consistently supported hypotheses and suppresses spurious ones, stabilizing optimization and improving training efficiency under uncertainty. Across multiple benchmarks, Tri-ABL consistently outperforms prior ABL methods and achieves state-of-the-art performance.