Recommending creators that a user does not yet follow is a cold-start problem: from hundreds of thousands of popular creators and only a short trail of prior engagements, the system must predict the next creator worth surfacing. The prevailing generative recommendation recipe tokenizes each creator into a Semantic ID (SID) by quantizing a frozen embedding and trains a language model to generate SIDs. We show that this SID layer, not the language model, is the binding constraint. The creator-to-SID map is many-to-one, so colliding creators must be resolved by a popularity tie-break that caps creator-level recall and favors already-popular creators, and the frozen quantizer stops the training gradient from reaching the creator representation. We propose Atomic Creator Token (ACT), which gives every creator its own vocabulary token trained end-to-end. The map is bijective, so collisions are impossible and inference is a direct top-k over creator logits. In a controlled comparison on production data in which the token layer is the only variable, ACT improves creator-level Recall@15 by 19.8 points (1.84×) over a matched SID baseline, still leads when the SID codebook is enlarged 64×, and serves at ∼2.7× lower per-user latency. The advantage carries over to the public KuaiRand dataset, where ACT is the only method that beats a global-popularity baseline, provided each creator token sees enough training examples. In a live A/B test, ACT is directionally positive on every reported metric against both the SID-enhanced and the classic production rankers.
Figures & tables
Figure 1: Overview of ACT. (01) Each candidate creator gets one dedicated vocabulary token through a bijection ϕ , instead of a many-to-one Semantic ID. (02) The creator embeddings are warmed up with the backbone frozen and then fine-tuned end-to-end with Low-Rank Adaptation (LoRA), so the gradient reaches them. (03) At inference, a single decode step per creator yields a top- k over creator tokens, decoded by an O(1) lookup without a codebook join or popularity rerank.
Figure 2: Creator-level Recall@ K of ACT and SID on production data at each vocabulary scale and SID codebook size. ACT leads at every K in every setting.
Vocabulary
SID codebook
νˉ
ρ
Tax
SID
ACT
Δ
Y (top)
X3
2.35
0.61
25.9
40.81
59.80
+18.99
10Y (mid)
X3
3.79
0.47
26.8
23.58
43.41
+19.83
10Y (mid)
(2X)3
1.75
0.64
16.7
29.21
43.41
+14.20
30Y (broad)
(4X)3
1.37
0.74
10.4
30.15
38.87
+8.72
Table 1: Recall@15 and collision statistics as the vocabulary and the SID codebook grow. νˉ : mean creators per SID; ρ=Rcreator/Rcluster for SID; tax =Rcluster−Rcreator (points); Δ : gain of ACT over SID. ACT has νˉ=ρ=1 and no tax at every scale.
Figure 3: Creator-level Recall@15 of SID and ACT per setting, with the fraction of creators that share their SID. A larger codebook reduces collisions and narrows the gap, but the gap never closes.
Figure 4: ACT across vocabulary tiers: conditional Recall@15 falls as the vocabulary grows, but engagement coverage c rises faster, so effective Recall@15 ( c× conditional) rises monotonically.
Figure 5: Creator-level Recall@ K .
Figure 8: Inference cost of ACT and SID across vocabulary tiers ( 32× A100-80GB, beam width 15, ∼1 M-user test). Throughput is concurrent requests at fixed memory, relative to SID.
Variant
R@15
R@50
R@200
Random init, trainable
43.91
60.96
80.01
GNN warm-start, trainable
44.66
61.62
80.44
+ contrastive loss
43.39
59.54
77.22
Table 2: Ablations of ACT design choices (SFT, 10Y ; creator-level Recall@ K with beam width 200; best per column in bold ).
Metric
vs. prod+SID
vs. prod
Overall consumption metric 1
+0.11
+0.05
Overall consumption metric 2
+0.01
+0.01
Product consumption metric
+0.02
+0.03
Product engagement metric
+0.29
+0.66
Table 3: Online relative change (%) of ACT against the SID-enhanced production arm (prod+SID) and the classic production ranker (prod). The product-engagement change against prod approaches statistical significance ( p≈0.05 ).
Appendix figures & tables5 assets
Supplementary material from the paper’s appendix.
Appendix
Collisions against codebook size
Codebook
Reachable codes
Used / reachable
νˉ
Colliding creators
SID X3
X3
88.7%
8.06
98.4%
SID (2X)3
(2X)3
43.5%
2.06
73.2%
SID (4X)3
(4X)3
8.2%
1.37
43.3%
ACT (bijective)
–
–
1.00
0%
Creator-level Recall@ K (%): largest SID codebook against ACT
Appendix
Table 4: Codebook stress test ( 30Y catalog, L=3 ). Top: collisions against codebook size. Bottom: creator-level Recall@ K of the largest SID codebook and ACT (beam width 15, ∼1 M test examples).
Method
Y (top)
10Y (mid)
30Y (broad)
C2C source A
6.70
5.58
5.04
C2C source B
4.12
5.30
5.68
C2C ensemble C
9.45
9.53
9.41
SID (creator-level)
40.81
23.58
30.15
ACT
59.80
43.41
38.87
Appendix
Table 5: ACT against non-generative retrieval baselines: creator-level Recall@15 (%) on the in-vocabulary population, ∼1 M test users; best per column in bold .
Method
R@1
R@5
R@10
R@15
Δ @15
SFT (reference)
17.04
30.57
38.29
43.41
–
+ DPO (best)
17.07
30.49
38.20
43.26
−0.15
RL-flow reference
13.88
23.83
29.56
33.47
–
+ GRPO (best)
14.44
23.37
28.61
32.20
−1.27
+ SDPO (best)
11.19
18.98
23.59
26.71
−6.77
Appendix
Table 6: Post-training the ACT SFT model with DPO, GRPO, and SDPO ( 10Y ; creator-level Recall@ K in %, beam width 15; best configuration per method). Δ @15 is relative to the applicable reference.
Figure 9: Independent t-SNE of the ACT creator embeddings (3072-d) and the production GNN prior (128-d) for the same 10,000 creators. The similar look is a t-SNE artifact; Table 7 shows the two spaces are largely decoupled.
Measure
Value
Reading
Linear CKA
0.10
near-zero shared global structure (1 = identical)
kNN@10 overlap
0.07
∼70× the random baseline of 0.001; 93% of neighbors change
Procrustes disparity (t-SNE)
0.887
layouts nearly uncorrelated (1 = uncorrelated)
Appendix
Table 7: Structural agreement between the ACT creator embeddings and the frozen GNN prior ( n=10,000 creators).