Did the model really adopt it? Counterfactual training improves language models' self-attribution ability

Researchers proposed the CREDIT training method to teach language models to accurately self-report whether their answers actually adopted user suggestions, all without human annotation. The approach has the model label at the end of its answer "adopted the suggestion" or "none," then tests the self-report by asking the same question four times with the hint removed; the authors found that frontier models' self-report performance was close to random. The training data comes from the model's own answers, modifying only the final attribution line based on how often the model gave the suggested answer when no hint was present. After fine-tuning, Qwen3-8B's balanced accuracy rose from 0.49 to 0.86; the authors say that on the same questions, all four open-source models tested exceeded all frontier models tested, and the code has been released.