Counterfactual training teaches LLMs to flag influential inputs

A new preprint tests whether language models actually tell you when the information you hand them shaped the answer they gave. The result, from Zikai Zhou, Jiacheng Sang, Yilin Xu, Yufei Jin and Monica S. Lam, is that they mostly do not.

The paper, posted to arXiv, reports that "most of the eleven frontier models rarely attribute influential information in their answers, and the rest attribute non-influential information." Pushing the acknowledgment into the model's chain of thought is not a fix: "the acknowledgment does not reach the final answer."

The authors' proposal is a training method they call CREDIT, which "directly supervises final-answer self-attribution with counterfactual labels, without the need for human annotation." They report it "substantially improves attribution in four open models, including two distilled reasoning models, and outperforms reasoning-level fine-tuning and prompting baselines." Among answers that happen to agree with a user's suggestion, CREDIT is also claimed to separate those changed by the suggestion from those the model would have given anyway, with gains that "transfer partly to unseen information types."

The abstract names no per-model attribution figures, no accuracy delta against the baselines, and does not say which eleven frontier models were in the evaluation set.