Kuaishou has quietly put a large language model in charge of explaining your video preferences to its own recommendation algorithm.
A new paper describes SARA (Scaling Articulated Rationales), an industrial framework built around Kuaishou Live, the short-video platform's live-streaming arm. Researchers first mined natural-language explanations from 240 million Kuaishou Live users to build SARA-HQ, a curated rationale dataset. Since real, human-written rationales existed for only 86,564 authors, the team fine-tuned a general-purpose multimodal model into SARA-7B, using supervised fine-tuning and a technique called Quality-Refining DPO, to generate plausible rationales for all 10 million authors on the platform. A separate component, SARA-Ranker, then feeds both the positive and negative generated rationales into production ranking.
Most recommendation systems only see behavior, clicks, watch time, skips, never the reasoning behind it, which is why this sparse-signal problem has lingered for years. Kuaishou reports the generated rationales beat other multimodal-model baselines on specificity and consistency in human evaluation, and that SARA-Ranker increased engagement and reduced negative feedback across a 30-day production A/B test.
Worth noting: almost all of those 10 million rationale sets are synthetic, an AI's guess about why you feel things, generated from a training set covering less than 1 percent of the platform's creators.