Why Thinking Hurts? Diagnosing and Rectifying the Reasoning Shift in Foundation Recommender Models
Chain-of-Thought reasoning was breaking recommendation models. A new training-free method fixes it.
Researchers from multiple universities diagnosed why adding Chain-of-Thought (CoT) reasoning to foundation recommender models like OpenOneRec paradoxically hurts performance. They found verbose reasoning causes 'textual inertia,' making models ignore critical Semantic IDs. Their solution, Inference-Time Subspace Alignment, compresses reasoning chains and uses bias-subtracted contrastive decoding. This training-free framework allows models to leverage reasoning without sacrificing the accuracy of their ID-based recommendations.
Why It Matters
Enables more explainable AI recommendations without costly retraining, improving trust and performance in systems like streaming and e-commerce.