Did the Model Use That? Counterfactual Training Boosts Self-Attribution in Language Models
Agreeing with a user’s suggestion does not mean it changed the model’s answer—and CREDIT trains models to distinguish the two.
Across eleven frontier models, most rarely acknowledged information that influenced their answers; the others credited information that had no influence. Teaching acknowledgment during reasoning did not carry it into final answers. CREDIT directly trains final-answer attribution using counterfactual labels, without human annotation, and substantially improved four open models against prompting and reasoning-level training baselines. The gains transferred partly to unseen information types and prompts with conflicting sources.
HF Daily Papers' note
Across eleven frontier models, most rarely acknowledged information that influenced their answers; the others credited information that had no influence. Teaching acknowledgment during reasoning did not carry it into final answers. CREDIT directly trains final-answer attribution using counterfactual labels, without human annotation, and substantially improved four open models against prompting and reasoning-level training baselines. The gains transferred partly to unseen information types and prompts with conflicting sources.
HF Daily Papers' note
score 5