Scaling Articulated Rationales for MLLM-based Recommendation

1 Introduction

• Behavioral signals and multimodal content describe actions or attributes, but often do not specify why a user likes or dislikes an item.Source: §1 · p. 3
• AUR scaling is constrained by below-1% rationale coverage, low-quality self-reports, and general-purpose MLLMs’ risk of ungrounded post-hoc rationalization.Source: §1 · p. 3
• SARA collects AURs from 240M users and curates 187,532 refined rationales into SARA-HQ.Source: §1 · p. 4
• SARA-HQ supervises two-stage SFT and Quality-Refining DPO alignment, while generated positive and negative rationales become ranking features for 10M authors.Source: §1 · p. 4
• SARA is presented as the first industrial-scale framework for articulated rationales, with daily-refresh deployment and measurable online gains.Source: §1 · p. 4

2.1 Overview

• The data engine has online collection and offline curation stages, forming the quantity–quality pipeline summarized in Figure 2.The overview precedes the detailed descriptions of both stages.Source: §2.1 · p. 4, Figure 2 · p. 5

2.2 Online Collection

• Questionnaires collect a positive or negative preference choice followed by an open-ended rationale, with delayed, sparse delivery balancing response quality and survey fatigue.The trigger follows 10 seconds of continuous viewing; baseline exposure is 5% and adjusts to recent response rates.Source: §2.2 · pp. 5–6

2.3 Post Curation

• Offline curation uses specialist assessments and a non-compensatory quality gate, then aggregates verified responses by author and polarity into semantic-tag profiles.Rationales are retained when bGagent > 2, but critical defects in relevance, safety, polarity consistency, or grounding cannot be offset by other scores.Source: §2.3 · p. 6
• SARA-HQ pairs multimodal livestream context with polarity-conditioned, normalized rationale targets distilled from retained questionnaire responses.The context includes sampled frames, metadata, ASR, and comments; the instruction conditions generation on positive or negative polarity.Source: §2.3 · pp. 6–7

2.4 Dataset Statistics and Analysis

• 1,915,718 candidate AURs yielded 187,532 high-quality rationales (9.8%) under the offline criterion bGagent > 2.The retained supervision pool is named SARA-HQ; Table 1 reports the latest data-engine statistics.Source: §2.4 · p. 7, Table 1 · p. 7
• 82% of 6,134 segmented rationales contain no more than ten Chinese characters, while late-window daily candidate volume and HQ yield differ by polarity.By the observation window’s end, positive candidates reach roughly 10K per day with yield near 5%, versus 2.5K per day and 21% yield for negative candidates.Source: §2.4 · pp. 7–8, Figure 4 · p. 8
• Unlike public livestream datasets using implicit behavior, categorical feedback, or content-understanding targets, SARA-HQ supplies real-user, polarity-conditioned, reason-level supervision.The comparison distinguishes preference-bearing explanations from datasets whose free-form responses are not used as benchmark targets.Source: §2.4 · pp. 8–9
• Quality-filtered rationales span 18 live-streaming domains and 141 categories with a substantial long tail; positive tags emphasize interaction, while negative tags include language and scripting.The semantic-tag word clouds also associate positive feedback with atmosphere, performance, and authenticity, and negative feedback with authenticity and transactional behavior.Source: §2.4 · p. 9, Figure 5 · p. 9, Figure 6 · p. 10

3.1 Overview

• SARA-7B aligns a general MLLM with SARA-HQ through two stages, extending rationale generation beyond the dataset’s 86,564 covered authors.The stages are Large-Scale SFT and Quality-Refining DPO.Source: §3.1 · p. 10

3.2 Large-Scale SFT

• SFT trains the model to generate a positive or negative rationale from multimodal livestream context and a polarity instruction.The task is single-turn conditional generation, optimized with full-parameter autoregressive negative log-likelihood.Source: §3.2 · p. 10

3.3 Quality-Refining DPO

• QR-DPO iteratively constructs preference pairs from the SFT model’s own responses, using the highest- and lowest-ranked rationales to shift probability toward higher-quality candidates.Figure 7 overviews the procedure; Figure 8 illustrates a chosen rationale linking stream details to perceived enjoyment and sincerity more concretely than the rejected response.Source: §3.3 · pp. 10, 12, Figure 7 · p. 11, Figure 8 · p. 11
• QR-DPO samples candidate rationales for 7K popular authors and uses a six-dimension Agent Judge and Senior Reviewer to select preferred and rejected responses.The resulting preference pairs train the policy from the SFT checkpoint using standard DPO with that checkpoint as the fixed reference.Source: §3.3 · pp. 10–11

4.1 overview

• The human-grounded protocol scores rationales on six dimensions and a holistic usability scale, with critical defects non-compensatory.The six dimensions are coherence, relevance, specificity, safety, polarity consistency, and grounding; both dimension and holistic scores use four levels.Source: §4.1 · p. 12

4.2 Online LLM Judge and Offline Agent Judge

• A single-pass online LLM supports latency-sensitive assessment, while offline Agent Judge evaluation uses dimension-specialized inspection followed by Senior Review.Both regimes are trained with the same human annotations; the online judge supports incentive allocation, and offline instances support curation, QR-DPO, and final evaluation.Source: §4.1 · p. 12, §4.2 · pp. 12–13, Figure 9 · p. 13

4.3 Judge Corpus and Evaluator Training

• The judge corpus combines native rationales, model-generated responses, and expert-verified controlled failures to train the online and agentic evaluators against human annotations.Controlled failures target context mismatch, polarity reversal, generic content description, and unsupported-detail injection; author and livestream context are split before augmentation.Source: §4.3 · pp. 13–14

4.4 Evaluator Validation

• 0.58 to 0.35 to 0.29 to 0.20 holistic-level MAE marks successive gains from judge-corpus training, specialist evaluation, and Senior Review; the inter-expert reference is 0.12.Table 3 reports MAE on a held-out evaluator test set, with lower values indicating closer agreement with experts.Source: §4.4 · pp. 14–15, Table 3 · p. 15

5.1 Overview

• Rationale embeddings guide fine-grained user–author interaction modeling, while rationale SIDs enrich target-aware negative-feedback histories.The two routes are embedding integration and hierarchical semantic-ID integration.Source: §5.1 · p. 15, Figure 10 · p. 15

5.2 Rationale Embedding Integration

• Positive questionnaire rationales provide reason-level supervision, and generated author-centric rationales extend it beyond sparse questionnaire coverage.The generated positive rationale is author-shared rather than conditioned on a specific user.Source: §5.2 · p. 16
• A symmetric soft-contrastive objective aligns user–author and rationale representations using targets that preserve similarities among semantically related rationales.Unlike one-hot targets, semantically related rationales receive nonzero target weights rather than being treated as negatives.Source: §5.2 · pp. 16–17

5.3 Rationale SID Integration

• Three-level rationale SIDs add shared reason-level semantics to negative-feedback histories, addressing the limits of author IDs for transferring rejection signals.The SID combines author-specific memorization, cross-author parameter sharing, and target-aware aggregation.Source: §5.3 · pp. 17–18
• Target-aware attention aggregates rejected-author history according to each candidate, allowing different candidates to use different parts of the same history.The resulting representation is incorporated into the negative-feedback task tower.Source: §5.3 · p. 18

6.1 SARA-7B Experiments

• 0.82 BGE-Sim on unseen authors exceeds Gemini-3.1-Pro (0.70) and Qwen3-VL-8B (0.69).SARA-7B also shows advantages in relevance, holistic quality, and BGE-Sim over general-purpose MLLMs.Source: §6.1 · p. 19, Table 4 · p. 19
• SARA-7B’s qualitative examples connect observable livestream evidence to specific positive and negative viewing reasons rather than merely describing content.QR-DPO examples likewise move beyond generic explanations toward more specific, verifiable rationales.Source: §6.1 · pp. 19–21, Figure 14 · p. 22
• As SFT supervision scales, relevance, specificity, grounding, and holistic quality continue improving, while coherence, safety, and polarity consistency saturate earlier.Marginal gains diminish at larger scales, but performance has not fully saturated.Source: §6.1 · p. 19, Figure 12 · p. 21
• QR-DPO raises specificity from 3.07 to 3.32, grounding from 3.63 to 3.75, and holistic quality from 3.15 to 3.33 over SFT.It preserves the other quality dimensions while improving rationale quality.Source: §6.1 · p. 19, Table 4 · p. 19
• Under the evaluated quantization setting, SARA’s SID-level and author-level collision rates are 1.5% and 3.3%, versus 7.9% and 22.5% for TagNex.SARA also raises two-level prefix utilization from 5.2% to 7.3% versus TagNex.Source: Table 5 · p. 21, §6.1 · p. 22

6.2 SARA-Ranker Experiment

• 8.16% online reduction in Hate feedback accompanies negative-rationale integration, while positive integration consistently improves online engagement and retention.The negative-integration test also reports a 0.44% absolute drop in Report and a positive LT-7 retention trend (+0.03%).Source: Table 7 · p. 23, §6.2 · p. 23, Table 6 · p. 23

7 Conclusion

• SARA-HQ supplies curated questionnaire supervision for SARA-7B alignment through SFT and Quality-Refining DPO, extending rationale generation from 86,564 covered authors to 10M.SARA-Ranker uses positive and negative rationales in user–author interaction and negative-feedback history modeling.Source: §7 · p. 23
• On unseen authors, SARA-7B generates more specific, relevant, and grounded rationales than the evaluated general-purpose MLLMs.Separate online A/B tests report a 0.99% watch-time increase with positive rationales and an 8.16% Hate-feedback reduction with negative rationales.Source: §7 · p. 23

Appendix

• Candidates are assessed against multimodal context and queried polarity, not writing quality alone.Among candidates failing hard constraints, defect severity separates poor from deficient; among those meeting them, remaining limitations separate acceptable from excellent.Source: §8 · p. 24
• The four-level rubric defines scores 4, 3, 2, and 1 as excellent, acceptable, deficient, and poor.Its score anchors are shared by human annotators, the online LLM, and the agentic evaluator.Source: Table 8 · p. 24, §8 · p. 24
• The holistic score is non-compensatory: any hard-dimension score at most 2 caps the holistic level at 2.Relevance, safety, polarity consistency, and multimodal grounding are treated as hard dimensions.Source: §8 · p. 24

• Haoke Xiao, Xiang Chen, and Jia Xu are credited with the Data Engine.Source: §9 · p. 25
• Haoke Xiao, Xiang Chen, and Yalong Guan are credited with the SARA-7B part.Source: §9 · p. 25
• Yueyang Liu, Yuhui Zhang, Xiaolan Zhu, Xiaoyu Zhang, Shijun Wang, and Shuang Yang are credited with the SARA-Ranker part.Source: §9 · p. 25
• Writing is credited to Haoke Xiao, Yuhui Zhang, Yueyang Liu, Xiang Chen, Yufei Liu, Zijie Meng, Zejian Zhang, and Ruochen Yang.Source: §9 · p. 25
• Xiang Chen is credited as Project Lead, while Xiangyu Wu, Tingting Gao, Han Li, Lantao Hu, Cheng Luo, and Kun Gai are credited as Advisors.Source: §9 · p. 25