This article recounts a JD engineer’s journey from theory to practice, detailing the development of the PODM‑MI re‑ranking framework, its three‑layer distribution‑based design, the discovery of a novel SID bottleneck, and the resulting multi‑million‑order impact validated at SIGIR 2024.
This study investigates the “sandwich” bottleneck observed in residual‑quantized semantic identifiers (RQ‑SID) used in generative search and recommendation systems, revealing that token concentration in intermediate codebooks caused by path sparsity and long‑tail distributions degrades performance, and proposes two effective mitigation strategies that improve efficiency and generalization in e‑commerce applications.
This paper investigates the "sandglass" phenomenon in residual‑quantized semantic identifiers for generative search and recommendation, analyzes its causes of path sparsity and long‑tail token distribution, and proposes heuristic and adaptive token‑removal methods that substantially improve model performance in e‑commerce scenarios.
This paper investigates the “hourglass” phenomenon in residual‑quantized semantic identifiers for generative search and recommendation, revealing that token concentration in intermediate codebooks causes path sparsity and long‑tail distributions, and proposes heuristic layer removal and adaptive token‑pruning strategies that markedly improve model performance.
