Debiasing LLM Recommendations via Decision-Aware Counterfactual Alignment

Large language models (LLMs) increasingly recommend options for consequential decisions, but information irrelevant to a decision can change which options they present and how they rank them. Mitigating this bias requires removing that influence without suppressing legitimate personalization. We develop Decision-Aware Counterfactual Alignment (DACA), which makes the recommendation delivered to the user the object of mitigation. DACA compares versions of decision cases that differ only in an irrelevant cue and uses the same candidate options for each. It trains a reward model on task-relevant evidence rather than the irrelevant cue, directly penalizes cue-induced differences in the final recommendation, and accepts an update only if it meets an independently measured utility requirement. We derive an end-to-end bound linking cue effects in generation, scoring, and selection to differences in final recommendations. We show that cue invariance at all three stages guarantees exact alignment, while invariance at any two does not. In a controlled college-major experiment with five open-weight LLMs, changing only a race or gender description replaces about 2.5 of ten recommended majors. DACA provides otherwise identical academic profiles the same majors in the sameorder, raises average rubric-based academic fit from 63.5 to 86.4 out of 100, and retains responsiveness to academic differences. Omitting the cue from the prompt also equalizes lists, but compromises fit and reduces responsiveness. Ablations show that reward learning supplies most of the fit gain, while a common candidate pool and final-output alignment remove changes in named majors and ranks that persist after the STEM-exposure gap closes.