Neutralizing Popularity Bias in LLM-based Recommendation via Counterfactual Reasoning Guidelines

September 9, 2026 | Source: arXiv AI

Tags: recommendation-systems, LLMs, popularity-bias, counterfactual-reasoning, personalization

NPRec fixes popularity bias in LLM-based recommender systems through counterfactual reasoning at inference time — no parameter updates required — generating debiased user preference guidelines that shift the LLM from trend-following toward genuine preference matching.

Details

LLMs trained on massive web corpora absorb popularity signals that contaminate their recommendations. A model asked to suggest movies tends to recommend popular films even when the user's actual preferences point elsewhere, because global trend signals dominate the parametric knowledge.\n\nNPRec (Neutralizing Popularity Bias via Counterfactual Reasoning Guidelines) addresses this without retraining the base model. Instead, it performs counterfactual refinement at inference time: it causally separates intrinsic user interests from popularity-driven conformity and generates debiased textual guidelines reflecting actual user preferences. These guidelines are injected as explicit premises at inference time.\n\nThe framework is model-agnostic — it works by prompting the LLM with structured guidelines rather than modifying parameters. This makes it practical to layer on top of existing LLM-based recommendation pipelines. Evaluation across three real-world datasets shows improvements in recommendation accuracy, explanation quality, and debiasing capability.\n\nFor teams building LLM-powered recommendation features (product discovery, content feeds, search ranking), NPRec offers a training-free debiasing approach that can be applied to any LLM.