Recommender System as Slow and Fast Thinkers
Zichen Yuan, Xiaoxuan Dong, Linkun Dai, Jinwei Yang, Jining Luan, Dexu Yu, Chunxiao Li, Joemon M. Jose, Youhua Li, Hanwen Du, Junchen Fu
cs.IR
2026-09-02
DS-Frame adds budgeted fast/slow routing to sequential recommenders: SASRec NDCG@10 rises 7.5% on average and 8.4% on long-history users, beating uniform slow inference.
Sequential recommenders run the same graph, with the same compute, for every user sequence. Short, mainstream histories often need one pass. Long, mixed, or long-tail histories may not. The issue is not only that some users are harder. Static inference spends the same budget whether an extra pass would help or not.
The paper splits users by interaction length and by average item popularity in history. SASRec and BERT4Rec are clearly stronger in the “common” bucket than in the “challenging” one. Under the length split, NDCG@10 drops 22.0% and 20.6% relative, and the hard group is 13.8% of users. Under a top/bottom 20% popularity split the drops are 14.7% and 13.7%.
DS-Frame is a plug-in fast/slow wrapper. A shared Transformer encoder produces the user representation. Fast is the backbone’s one-pass head. Slow refines the full hidden sequence for K steps: each step appends a reasoning token with a step embedding, a shared reasoning block updates the sequence, and the last token becomes the new state. Training supervises every step for next-item prediction, adds a KL continuity term between adjacent steps, anneals temperature across steps, and averages logits at the end.
The selector is a two-layer MLP that emits the probability of Slow. At train time both paths run. The oracle label is whichever path has lower loss plus a normalized cost: Fast costs 1, Slow costs 1+γK with γ=1 by default. A budget term pulls the mini-batch mean activation toward a target b. At inference the gate score and a threshold decide; labels are not available. K is tuned in {1,2,3}. The Beauty tables use K=2 and about 40% Slow.
Five datasets: Yelp plus Amazon 2023 Video Games, Beauty, Sports, and Toys. Evaluation is full ranking, not sampled negatives. Sequences are truncated at 50.
DS-Frame raises SASRec by 7.5% / 6.6% / 6.6% / 6.2% relative on NDCG@10 / NDCG@20 / HR@10 / HR@20, and BERT4Rec by 6.7% / 6.3% / 6.0% / 5.9%. On Beauty, SASRec NDCG@10 goes from 0.0418 to 0.0453 (+8.4%). Under a shared SASRec backbone, average NDCG vs the vanilla model is STREAM +5.2%, ReaRec-PRL +7.2%, DS-Frame +7.5%, LARES +7.3%, ManCAR +7.6%. The method is even with the best single-path refiners; the claim is allocation, not a new high-water mark.
Gains follow difficulty. Length split, SASRec: +3.3% on common users (24.8% Slow) vs +8.4% on hard users (58.6% Slow). BERT4Rec: +3.2% / +8.5%. Beauty routing table:
| Policy | NDCG@10 | Slow act. | Rel. cost (K=2) |
| Fast only | 0.0418 | 0% | 1.00× |
| Slow only | 0.0431 | 100% | 3.00× |
| Random 40% | 0.0438 | 40% | 1.80× |
| Learned selector | 0.0453 | 40.2% | 1.80× |
| Oracle | 0.0461 | 40% | 1.80× |
Learned routing at 40%–60% budget already beats always-Slow. The selector, without labels at test time, sits close to the oracle.
For teams already on SASRec or BERT4Rec, this is conditional compute on the inference side, not a new backbone. Extra FLOPs should go to samples with a larger Slow-minus-Fast gap: about 4× larger on the hard length group, and the activation rates match that gap. Always thinking longer is both costlier and worse, so Slow is not a strictly stronger model. It is a refinement that has to be rationed.
ReaRec, LARES, and ManCAR decide when to stop inside one reasoning module. DS-Frame keeps the backbone one-pass path and learns a single sample-level yes/no for a fixed K-step refinement.
There is no dedicated limitations section. Sequences are capped at 50, so the “long history” bucket is truncated and is not real unbounded behavior. Only two backbones are wrapped; graph and contrastive sequential models are untested. All logs are academic click data; there is no online A/B under a latency SLA. The training oracle uses labels, so the deployed gate has a distribution shift, and the learned selector still trails the oracle (0.0453 vs 0.0461). The popularity split is not a debiasing method. The hard length group is 13.8% of users, so most of the headline 7.5% still comes from small gains on common users. ManCAR’s average is slightly higher; adaptive routing is not the quality ceiling.