CALMRec Splits LLM Recommender Memory Into Short-Term, Long-Term, and Exposure Stores to Break Feedback Loops

arXiv 2607.23647·low signalXBlueskyLinkedInCopy link

Current LLM recommenders collapse enduring preference, transient intent, and exposure-induced behavior into a single profile, which makes repeated exposure look like preference and lets fluent explanations drift from the actual ranking decision. CALMRec (arXiv 2607.23647) uses a frozen multimodal LM to convert item content and feedback into evidence-grounded 'semantic atoms', maintains three separate memories, applies propensity-weighted updates to counter policy-induced exposure bias, and reranks with a conservative offline critic under a behavior-support constraint. Across ten seeds it improves discounted long-term value over the strongest alternative by 6.1%, 7.6%, and 6.7% in e-commerce-like, news-like, and short-video-like environments; twenty-seed ablations show significant drops when propensity correction or conservative regularization is removed.