Retrieval models for recommendation need to account for outcomes such as watch time while searching large item corpora efficiently. We present the Controllable Retrieval Model (CRM), an outcome-conditioned retrieval framework that adds watch time as an input to the user tower. During training, CRM learns to predict the next item given its observed watch time; at inference, we supply a desired watch time to steer retrieval without retraining. The item tower is independent of the condition, so item embeddings can still be precomputed and indexed for approximate nearest neighbor (ANN) search. We implement CRM as both a DNN-based two-tower model and a Decision Transformer. The Transformer interleaves item tokens with watch-time-to-go condition tokens, using the target watch time as a numerical prompt for next-item prediction. For each online request, we choose between the average and maximum watch times in the user's recent history as the condition. CRM is deployed in Kuaishou's short-video recommendation system, which serves over 400 million users. Online A/B tests show video watch time improvements of 0.372% in Kuaishou and 0.457% in Kuaishou Lite.
Figures & tables
Figure 1. Retrieval and ranking in a recommendation system: (a) a standard two-tower retrieval model trained with a classification objective; (b) a ranking model with classification and regression objectives; (c) CRM, which adds an outcome condition to the user tower. Three architecture diagrams compare standard two-tower retrieval, multi-task ranking, and CRM. CRM adds a condition to the user tower and retains an independent item tower and ANN search.
Figure 2. Input sequences for (a) the original Decision Transformer and (b) CRM. The original model uses return-to-go, state, and action tokens. CRM alternates item tokens with watch-time-to-go condition tokens. Both models use causal attention masks. The original Decision Transformer interleaves return-to-go, state, and action tokens. CRM alternates watch-time-to-go and item tokens to predict the next item.
Applications
Video Watch Time
Total App Usage Time
Usage Time per User
Avg. Time per Video View
Video Views
Likes
Follows
Comments
Kuaishou Single Page
+0.372%
+0.175%
+0.144%
+0.323%
+0.048%
+0.029%
+0.447%
+0.330%
Kuaishou Lite Single Page
+0.457%
+0.196%
+0.122%
+0.133%
+0.324%
+0.137%
+0.828%
+0.697%
Table 1. Online A/B test results for short-video recommendation at Kuaishou.
Figure 3. Average and maximum watch-time conditions over a day. Two curves show the average and maximum watch-time conditions over 24 hours. They follow similar daily trends, with the maximum condition higher than the average.
Figure 4. Average watch time per video view for different retrieval models. Average watch time in seconds: Swing 23, TDM 27, LTR 29, ComiRec 31, DimeRec 33, KuaiFormer 36, and CRM 42.