Arpit Shah

How Faithful Is the Reasoning of LLM Recommenders? A Counterfactual Audit

This work introduces a counterfactual audit that edits a model’s own reasoning trace using four operators that flip the stated preference, swap in another user’s reasoning, erase it, or paraphrase it as a control, and measures whether the top-1 recommendation changes.