Recommendation World Models for Future-State Control
Abstract Overview
This paper studies how a trained sequential recommender can be wrapped with a decision layer that reasons about the future consequences of its displayed slates. The proposed interface, UA-TWM, keeps the original recommender frozen as a utility anchor, generates nearby alternative slates, predicts their target-relevant future effects, and selects an alternative only when utility and risk constraints are satisfied; otherwise it falls back to the anchor slate. The framework is instantiated in two settings: logged replay with calibrated utility, target-gain, and failure-risk estimates, and closed-loop control with one-step state-action prediction. Experiments cover transfer across twelve sequential recommendation backbones on MovieLens-25M and KuaiRand-Pure, plus repeated target-directed interaction in KuaiSim.
Novelty
The distinctive contribution is a utility-anchored world-model interface for recommendation that performs local slate-action control around an already trained ranker rather than replacing the ranker itself. It also combines constrained selection, fallback to the anchor, and action-conditioned consequence prediction to support future-state objectives under utility-preservation requirements.
Results
In logged evaluations, attaching UA-TWM improved Recall@20, NDCG@20, and future-state alignment for every matched backbone on both MovieLens-25M and KuaiRand-Pure. Reported gains on MovieLens-25M were Recall@20 increases of 0.0119-0.0144, NDCG@20 increases of 0.0028-0.0102, and future-state L1 reductions of 0.037-0.046; on KuaiRand-Pure, Recall@20 increased by 0.0054-0.0068, NDCG@20 by 0.0018-0.0037, and future-state loss dropped by 0.064-0.082. In KuaiSim closed-loop control, UA-TWM achieved click 42.936 and target L1 2.581, made four valid target-control claims, used fallback on three scenarios, and had zero invalid claims.
Key Points
- UA-TWM treats the trained recommender as a frozen anchor and searches among nearby slate alternatives using predicted utility, target gain, and failure risk.
- The method was tested broadly across twelve sequential recommendation backbones and showed consistent matched improvements in both ranking metrics and future-state alignment on two public datasets.
- Closed-loop diagnostics in KuaiSim indicate that action-conditioned prediction matters: the full state-action interface could make valid target-control interventions while a state-only ablation made no claims.