Researchers trained small AI models to provide evidence for their decisions without human help

A new research paper explores how we can teach AI models to explain their reasoning even when we lack perfect training data. Often, businesses and organizations keep records of final decisions but never bother to note down the specific evidence used to reach them.

Because labeling this data manually is expensive, models usually learn to mimic the final result without understanding the context. The researchers tested two methods to see if models could recover this missing evidence using only the recorded verdicts.

First, they used a simple label-only training approach. This pushed the model to align its outputs with the final decision. Second, they used a technique called rejection sampling, where the system keeps only the generated explanations that match the recorded outcome.

Both methods showed improvement. The model was better at providing actual citations from documents rather than just guessing based on the final answer. This is important because, in professional reviews, a correct decision is often useless if nobody can verify how the AI reached it.

Imagine you ask a legal AI if a contract is valid. It might say yes, but without a citation, you have to read the whole document to check its work. Being able to trace that 'yes' back to a specific clause makes the AI far more useful.

Accuracy metrics and citation quality do not always align. A system might get the right answer for the wrong reasons, making it unreliable for professional review. This research highlights that we need to prioritize transparent evidence.

While these results are promising, the study was limited to specific corpora like ContractNLI. The researchers noted that they cannot yet definitively say which method is superior, as the effectiveness of these techniques can vary based on the data used.