LinkedIn's TransX cuts serving compute 80%, lifts CTR 6%

LinkedIn's AI team (Da Xu and colleagues) introduced TransX, a production-oriented encoder-decoder model that reframes recommendation as sequence-to-sequence action transduction. Traditional transformer-based recommenders collapse long-term user behaviors and real-time serving events into a single monolithic token stream, obscuring their distinct causal roles and raising training and serving costs. TransX explicitly decouples behavior-stream modeling from serving-event modeling, then uses scalable cross-attention between nearline behavior encodings and real-time serving representations to condition next-action decoding. To make this viable at scale, the system uses an amortized serving strategy combining incremental behavior encoding with per-request key-value caching, making serving latency insensitive to behavior sequence length.

Extensive offline experiments and large-scale online A/B tests on LinkedIn's recommender systems show TransX consistently outperforms state-of-the-art deep learning recommenders (DLRMs) and sequential baselines. It delivers a substantial +6.0% lift in click-through rate and +4.4% conversion gain, while maintaining serving costs comparable to existing production models. The co-designed serving strategy reduces online computation by approximately 80%, enabling low-latency, high-QPS deployment. This work demonstrates that explicitly modeling the distinct roles of behavioral and serving streams—rather than merging them—can unlock both quality and efficiency gains in industrial recommendation systems, pointing toward a new design pattern for large-scale transformer-based recommenders.