A new method tries to stop AI systems from supervising themselves with bad guesses.
Abductive learning pairs neural perception - the pattern-matching part of a model - with symbolic reasoning, which is just logic rules applied to what the perception model sees. When the symbolic side needs to explain an outcome, it uses abduction, a process of working backward to generate candidate explanations. Those explanations become pseudo-labels that train the perception model. The snag: several explanations can be equally valid but assign different labels to the same input, and current methods either commit to one candidate (risking that a wrong guess gets reinforced) or spread weight across all of them, which dilutes the signal. The paper's fix, Abductive Candidate Retention (ACR), keeps a curated subset instead, adding a candidate only when the information it recovers outweighs the uncertainty it introduces.
This is really a data-quality problem wearing a neurosymbolic costume, the same kind of noisy-label challenge that has dogged weak-supervision techniques like distant supervision in NLP for years. The authors report ACR beats both single-candidate baselines and a prior method called A3BL on most of their "aggregated mod-addition" tests, with ablations showing the gain comes from combining uncertainty and posterior mass rather than either one alone.
The catch is the benchmark: mod-addition tasks are a clean, synthetic arena for testing label-selection theory, not evidence this scales to messier real-world symbolic systems.