AMap's IntBMoE lifts UVCTR 2.4% in production A/B test

A new mixture-of-experts design called IntBMoE is running live inside AMap's generative recommendation system, serving hundreds of millions of users under a 60-millisecond latency budget and posting a 2.4% relative UVCTR gain in online A/B testing, according to the paper's abstract by Ran Cheng and coauthors.

The framing the authors put on the architecture side is that today's MoE designs cannot set three quantities independently. In their words, 'participation is how many experts contribute knowledge to its output, execution is how many are actually computed (compute cost), and materialization is how many expert-sized parameter sets must be built and stored (memory cost).' Sparse routing keeps execution and materialization low but shrinks participation. Dense output-mixing restores participation but grows execution. Parameter-merging keeps execution at one expert but grows materialization. IntBMoE, the paper claims, 'decouples all three by pairing dense expert composition with sparse block execution,' with a small learned codebook fixing how many blocks exist and a lightweight hypernetwork merging expert bases at each internal layer.

The design also adds Dual-Path Residual Gating, which the authors describe as coupling 'two independently composed paths through multiplicative gating.' Image-classification experiments show gains over representative sparse and dense MoE baselines; further tests on language modeling and sequential recommendation, the authors write, 'validate its generalization beyond vision.'

The abstract does not name the incumbent model the A/B test was measured against, nor publish per-block cost figures.