GENERATIVE · 2026-11-17

Generative — Paper Daily, 2026-11-17

Today's generative track has a single paper, and it asks whether a long history can be compressed into a fixed-size memory without dropping the signals needed for prediction. Once a generative recommender feeds very long interactions straight into the model, compute and memory grow with length. This paper writes history into a fixed number of memory slots segment by segment, and uses a fixed-length context at decode time. On a long-sequence e-commerce benchmark, the compressed model still approaches a counterpart that processes the long history directly, and distant behavior still contributes signal. For teams that want to serve long-sequence generative recommenders, the point is not to stretch the attention window without limit, but to keep the memory stable and usable as the history keeps growing.

4.30
LEAD STORY · Alibaba / Shanghai Jiao Tong University

1. RecMem: Recurrent Memory Compression for Long-Sequence Generative Recommendation

NeurIPS 2026AlibabaShanghai Jiao Tong University📖 Deep dive →

Once a generative recommender feeds a user's long history straight into the model, compute and memory grow with history length, and long interaction sequences quickly become slow and expensive. The authors compress history segment by segment into a fixed number of memory slots, and the decoder uses a fixed-length context that no longer expands as the history grows.