Scaling Articulated Rationales for MLLM-based Recommendation
Haoke Xiao, Yueyang Liu, Yuhui Zhang, Xiang Chen, Yufei Liu, Jia Xu, Yalong Guan, Xiaolan Zhu, Xiaoyu Zhang, Shijun Wang, Shuang Yang, Zijie Meng, Zejian Zhang, Ruochen Yang, Xiangyu Wu, Tingting Gao, Han Li, Lantao Hu, Cheng Luo, Kun Gai
cs.IR, cs.AI
2026-09-15
SARA turns sparse Kuaishou Live like/dislike reasons into ranking features for 10M authors. Online click +0.342%, watch time +0.986%, Hate -8.164%, daily-refreshed for 30 days.
Recommenders infer what users did from clicks, watch time, and dislikes. They rarely know why. The same long view can mean topical fit or companionship; the same dislike can mean a hard sell or an unsafe stunt. Tags and comments describe content. They are not bound to a preference decision. A natural-language reason for like or dislike is a polarity-aware, reason-level signal, and industry almost never treats it as a first-class feature: coverage is under 1%, text is short and colloquial, and general MLLMs invent post-hoc excuses.
SARA's job is to turn those sparse utterances into a daily-refreshed ranking signal.
A questionnaire sits in Kuaishou Live. After 10 continuous seconds of watching, it fires at a 5% base rate, adjusted between 1% and 20% from the last five minutes' response rate. Users pick positive or negative, then write a reason. An online BitCPM4-0.5B judge scores six dimensions; a small coin reward is paid only if the holistic score exceeds 2. From 12 Feb to 30 Aug 2026 the pipeline collected 1.92M candidates. Offline, an Agent Judge filters with non-compensatory rules: a fatal miss on relevance, safety, polarity, or grounding cannot be rescued by fluent prose. 187,532 items remain, a 9.8% yield. Negative yield is 24.2% versus 6.8% for positives, because complaints name defects while praise is often generic. 82% of raw replies are at most ten Chinese characters. They are grouped by author and polarity into 86,564 author profiles, then packed as multimodal examples with frames, metadata, ASR, comments, and a polarity instruction. Questionnaire text is the target only; it never enters the context.
SARA-7B is fully fine-tuned on SARA-HQ, then refined with Quality-Refining DPO: 7K popular authors, self-samples from the SFT model, the same Agent Judge ranking first versus last. The generator sees author context and polarity, not a specific user, so outputs are author-level shared rationales, extended from 86k covered authors to about 10M.
Ranking has two branches. Positive: user profile, author description, and positive rationale are looked up, encoded, fused with bidirectional cross-attention, and aligned to the rationale via mutual-information maximization before the production MMoE. Negative: negative rationales become three-level residual semantic IDs attached to the user's Hate history; the candidate is the query, so "dislikes hard-sell" can transfer to unseen authors with similar reasons.
On a 5K author-disjoint set, a frozen Agent Judge scores 1 to 4:
| Model | Specificity | Holistic | BGE-Sim |
| Gemini 3.1 Pro | 2.92 | 2.51 | 0.70 |
| Qwen3-VL-8B | 2.67 | 2.37 | 0.69 |
| SARA-7B-SFT | 3.07 | 3.15 | 0.82 |
| SARA-7B-DPO | 3.32 | 3.33 | 0.82 |
Coherence, safety, and polarity are high for everyone. The gap is relevance, specificity, and grounding. DPO lifts specificity 3.07 to 3.32 and grounding 3.63 to 3.75. Scaling SFT data keeps those axes moving after fluency saturates.
Under an (8192)^3 codebook, rationale SIDs collide at 1.5%, versus 7.9% for in-house TagNex and 3.6% for OneLive MLLM embeddings.
On the production MMoE plus SARM stack, independent 1% A/B tests stay inside the latency budget, refresh daily, and have run more than 30 days. Positive branch: Long-view AUC +0.28 pp offline; online click +0.342%, watch time +0.986%, effective view +0.365%, follow +0.616%. Negative branch: Hate AUC +0.35 pp, Report +0.46 pp; online Hate -8.164%, Report -0.4396%, 7-day retention +0.027%.
Industrial ranking is not short on multimodal encoders. It is short on reasons users actually stated. SARA turns elicitation, QA, generation, and ranking into a daily pipeline, and still moves online metrics on top of a loaded multimodal baseline. Negative SIDs turn a dislike from "remember this author" into "remember this kind of reason", which helps the rare Report head.
The fit is specific: live or short-video products that can field questionnaires and accept author-level shared reasons. Without an elicitation engine, SARA-7B has no supervision.
Rationales are shared per author, not "why this user likes this author". MIM can only pull the interaction toward author-side language. The 10-second delay undersamples early exits, so negative coverage is biased. Curation, DPO, and generation eval share one Agent Judge, so score gains can self-reinforce; BGE-Sim of 0.82 also measures fidelity to that same questionnaire distribution. Online lifts are 1% traffic on a strong baseline; Hate dropping 8% needs a check that the survey and rewards did not themselves change feedback. Failure cases and transfer outside livestream ranking are barely discussed.