Alibaba's 10 CIKM 2026 Papers: Multimodal Large Models Transform E‑Commerce Search & Recommendation

SORT: Systematically Optimized Ranking Transformer for Industrial‑Scale Recommenders

Problem: Transformers excel in LLMs but face extreme feature sparsity, low label density, and massive vocabularies in industrial ranking; traditional DLRMs rely on low‑cost vector interactions and cannot exploit modern accelerator parallelism. Solution: SORT redesigns sample organization around a request (sharing user history across candidates), introduces sparse local attention with window = 256 and layer‑wise query pruning, employs generative next‑item pre‑training to densify supervision, and incorporates LLM best practices (BOS/SEP tokens, QKNorm, gated attention, sparse MoE). System‑level optimizations boost training MFU (22 %–45 %) and inference throughput while cutting latency. Result: Offline, SORT outperforms standard Transformer, HSTU and OneTrans across three scale settings with only 58 % of FLOPs; scaling experiments confirm data size as the primary driver of gains. Online A/B on AliExpress (four core scenes) shows +7.47 % orders, +6.67 % buyers, +8.65 % GMV, -61 % latency and +589 % throughput, leading to full deployment.