Modern recommender systems mainly infer user preferences from implicit signals such as clicks, watch time and negative feedback. These signals reveal what users do but not why they like or dislike an item. This work introduces articulated user rationales (AUR) – natural‑language explanations provided by users – as a polarity‑aware, reason‑level textual signal. AURs are inherently sparse, often low‑quality, and cover only a tiny fraction of items, making them hard to adopt in production.
The SARA (Scaling Articulated Rationales) framework first builds a data engine that solicits and curates AURs from 240 M Kuaishou Live users, yielding SARA‑HQ, a quality‑controlled, author‑centric rationale dataset. It then aligns a general‑purpose multi‑modal LLM to SARA‑7B via large‑scale supervised fine‑tuning (SFT) and quality‑refining DPO, expanding rationale generation from 86,564 covered authors to the full 10 M‑author space. Finally, SARA‑Ranker incorporates the generated positive and negative rationales into production ranking using rationale‑aware interaction modeling and rejection‑memory modeling.
Extensive offline evaluation, human calibration, and online A/B tests demonstrate that SARA‑7B produces more specific, polarity‑consistent, and grounded rationales than strong baselines, while SARA‑Ranker improves engagement and reduces negative feedback in production. Deployed with daily refresh for over 30 days, SARA confirms articulated rationales as a practical first‑class textual signal for industrial recommendation systems.
Review