SAPO: Step-Aligned Policy Optimization for Reasoning-Based Generative Recommendation
The episode discusses SAPO, a method for improving reasoning-based generative recommenders. It solves the 'action-granularity mismatch' where traditional training methods fail unless the entire final output is perfect. SAPO fixes this by breaking the model's output into steps and assigning partial credit for each step, allowing the model to learn from near misses, resulting in faster and more stable training.