ByteDance’s SequenceO1 paper proposes “sketches” for recommendation histories of up to 100,000 events

A post describing ByteDance’s SequenceO1 paper says the model summarizes up to 100,000 user events in 1,024 learned representations, while a separate branch uses the most recent 10,000 events. The post says the authors observed 2.33% more video finishes and 6.98% fewer dislikes on Douyin. They estimated the ultra-long-history branch would require 50 times fewer training FLOPs and 64 times fewer inference FLOPs than processing the full 100,000-event sequence directly with STCA.