I am a coding monkey, and I am proud of it. I have done lots of work in machine learning area, especially recommendation system and AutoML. This blog summarize my journey to become an expert monkey in distributed system and LLM.
Welcome back to the sixth installment of A Random Walk Down Recsys. This batch is smaller than the last — four papers — but two of them line up neatly enough that they deserve a side-by-side read: Industrial distillation pipelines — ByteDance’s Rec-Distill and Meta’s LoopFM are both attacking the same problem (transferring knowledge from an expensive teacher / foundation model into the prod...
Welcome back to the fifth installment of A Random Walk Down Recsys. This batch of papers reflects how quickly the generative recommendation playbook is being adapted to new verticals and tightened for production. Six papers, four themes: Generative Recommendation for Ads — two concurrent works, one from Kuaishou and one from Tencent, both attacking the same gap: today’s GR systems are heavi...
Welcome back to the fourth installment of A Random Walk Down Recsys. This time, the three papers span a range of practical challenges in generative recommendation: efficiently compressing long user sequences through recurrent memory, accelerating constrained decoding on hardware accelerators via trie vectorization, and rethinking how semantic IDs are trained and maintained with a dynamic, end-t...
Welcome back to the third installment of A Random Walk Down Recsys. This time, all five papers revolve around a single theme: Semantic IDs (SIDs) — how to generate them, how to improve their quality, and how to leverage them effectively in generative recommender (GR) models. The papers span a wide range of ideas: compressing long user sequences through SID hierarchies, injecting reasoning capab...
Welcome back to the second installment of A Random Walk Down Recsys. In this post, I continue surveying interesting papers from the Arxiv IR section, covering five recent works: HyFormer, Token-level Collaborative Alignment, OneMall, a Sparse Attention approach for long-term user behaviors, and Farewell to Item IDs. HyFormer: Hybrid Cross-Attention for Sequential and Non-Sequential Features ...