🎁 Recommender Systems

🧪 ICML2026 · 1 paper notes

📌 Same area in other venues: 💬 ACL2026 (12) · 🔬 ICLR2026 (10) · 🤖 AAAI2026 (26) · 🧠 NeurIPS2025 (24)

Can Recommender Systems Teach Themselves? A Recursive Self-Improving Framework with Fidelity Control

RSIR enables sequential recommendation models to generate new synthetic user interaction sequences using their own predictive capabilities, retrain a new model, and filter out samples deviating from the user preference manifold via a ranking-based "fidelity check," thus preventing self-consuming model collapse. On 4 datasets × 3 mainstream backbones, it consistently improves NDCG/Recall by 4–11%, and theoretically proves this process is equivalent to implicit regularization along the tangent space of the user preference manifold.