tech
Fu Cong Team and Xiamen University Propose ManCAR: Manifold-Constrained Adaptive Reasoning Boosts Recommendation by 46%
Want to read in a language you're more familiar with?
TL;DR
- ManCAR (Manifold-Constrained Adaptive Reasoning) is a new framework for generative recommendation, developed by Fu Cong's team and Xiamen University.
- It addresses the question of whether generative recommendation models truly reason about user interests or perform deeper feature transformations.
- ManCAR redefines multi-step reasoning as navigating a manifold of user collaborative behavior, rather than unconstrained latent state iteration.
- The framework incorporates a manifold constraint to guide reasoning along a continuous path from user behavior triggers to target interest regions.
- It learns a local manifold to smooth discrete path spaces, allowing the reasoning process to find continuous paths and solve the latent drift problem.
- ManCAR achieves up to 46.88% improvement in NDCG@10 and is the first theoretically grounded framework for meaningful reasoning process guidance in recommendations.
- The framework has been open-sourced to encourage broader research and industrial adoption.