Hallucinate Less by Thinking More: Aspect-Based Causal Abstention for Large Language Models

People do this too: when a question feels familiar, we often jump to the first interpretation that fits instead of considering every plausible reading.

Language models behave similarly. They tend to follow the most convenient path, usually the association they have seen most often, and leave alternative interpretations unexplored.

Take the question: “Who rings the bells at Notre Dame?”

One reading is fictional: in The Hunchback of Notre-Dame, the bell-ringer is Quasimodo. Another reading is literal: at Notre-Dame Cathedral in Paris, there is no single person who can simply be named as “the bell-ringer”.

A model may latch onto the familiar fictional association and answer Quasimodo with confidence. The problem is not that the model knows nothing else. It may already contain the alternative reading that should make it hesitate. It simply does not bring that reading into the reasoning before committing.

A check on the finished answer comes too late. By then, the convenient path has already won.

I developed Aspect-Based Causal Abstention, or ABCA, to bring those overlooked readings back into the reasoning process before the model commits.

The model considers different aspects of the question and asks whether they support the same conclusion. If the interpretations conflict, or if none provides enough support for a definite answer, the model abstains.

Said plainly, do not trust the first plausible thought just because it came easily.