Recent work leverages large-scale pretrained foundation models (FMs) to transfer generalizable representations to downstream advertising tasks. Most existing approaches transfer user representations (URs) and item representations (IRs), while underutilizing the user-item cross representations (CRs) generated by deeper interaction layers. Although a recent work transfers target-aware embeddings (TAEs), it requires real-time upstream FM inference alongside the downstream, introducing additional engineering complexity and overhead. To address these limitations, we propose Large Foundation Model for Ads Recommendation (LFM4Ads), an efficient pretrain-then-transfer paradigm for cross-domain ad recommendation. LFM4Ads jointly exploits URs, IRs, and CRs from the upstream FM. It extracts sample-level CRs from pretrained interaction layers and aggregates them into user-level CRs using a time-aware exponential moving average, thereby decoupling upstream inference from downstream online serving. To use CRs efficiently, we propose three downstream usages at different levels of granularity: feature-level gating to fuse CRs into downstream embeddings, module-level fine-tuning to yield downstream CRs, and model-level retrieval by matching CRs with IRs. LFM4Ads has been deployed on Tencent's advertising platform since Q4 2024. It has supported 10+ launches across major scenarios, yielding a platform-wide gross merchandise value (GMV) lift of 2.45% and annual revenue gains in the hundreds of millions of dollars. Offline experiments on KuaiRand-1K further demonstrate our effectiveness. Our code is available at https://github.com/ZhuangQu/LFM4Ads.
Figures & tables
Figure 1. Overview. (a) LFM4Ads is pretrained on content and ad samples across scenarios and tasks, and representations are transferred to downstream models. (b) Cross representations are fused by gating for feature-level usage. (c) Modules that yield cross representations are fine-tuned for module-level usage. (d) Cross and item representations are matched for model-level usage.
Figure 2. Offline results on two typical scenarios of KuaiRand-1K.
Scenario
1
0
4
2
6
3
8
5
Average
# of samples
7,717,601
2,407,352
895,385
402,293
183,403
37,418
17,430
16,424
–
% of samples
65.9%
20.6%
7.6%
3.4%
1.6%
0.3%
0.1%
0.1%
–
Feature
SUM
0.7310
0.7123
0.7101
0.7975
0.7183
0.7017
0.7622
0.6562
0.7275
concat CR
0.7325
0.7123
0.7142
0.7993
0.7277
0.7114
0.7668
0.6622
0.7291
gating CR
0.7334
0.7132
0.7150
0.8000
0.7249
0.7302
0.7657
0.6699
0.7300
Module
Vanilla
0.7309
0.7120
0.7086
0.7971
0.7173
0.7038
0.7627
0.6606
0.7273
Table 1. Offline results on all available scenarios of KuaiRand-1K. Highest is bolded and lowest is italicized. Average is weighted by the number of samples. CR is extracted from Cross3 for feature-level usage, and DNN2 for model-level usage. All layers are fine-tuned to yield CR′ for module-level usage.