Alibaba International AI Team Lands Multiple Papers at SIGIR, WWW, and WSDM 2026

1. Sparse Scaling Recommendation (SSR) – SIGIR 2026

Pain point: Dense MLP backbones hit a capacity ceiling; weight visualisation shows 92% of connections near zero and 80% of weight energy concentrated in only 4% of input dimensions, meaning most parameters learn to be unused.

Solution: The SSR (Scalable Sparse Recommendation) framework adopts a "Filter‑then‑Fuse" two‑stage paradigm. Stage 1 applies explicit sparsity selectors: SSR‑S (static random sparsity, 99.99% structural sparsity, zero FLOP) and SSR‑D (dynamic differentiable sparsity, ~91% activation sparsity). Stage 2 performs dense non‑linear fusion in the filtered low‑dimensional subspace. The core idea is to make sparsity an architectural principle rather than a learned by‑product.

Results: On a billion‑scale AliExpress dataset, SSR‑S with roughly half the parameters and FLOPs surpasses the current SOTA RankMixer. Scaling experiments show dense MLP saturates quickly, while SSR‑D continues to improve up to ~900 M parameters, breaking the dense scaling ceiling. Online A/B tests report significant lifts in CTR, order volume, and GMV with almost no inference latency increase.

2. Counterfactual Multi‑task Learning for Delayed Conversion – SIGIR 2026

Pain point: Conventional CVR models focus on the promotion day and ignore the pre‑promotion stage, where users browse and add‑to‑cart but postpone purchase, causing a distribution shift (OOD) and a delayed‑conversion problem. ATC (add‑to‑cart) and conversion are strongly correlated but confounded, and existing continuous‑delay models assume uniform delay, which is invalid.

Solution: CM‑DCM introduces three innovations: (1) a multi‑task architecture jointly modelling direct conversion and delayed conversion using pre‑promotion data; (2) personalized gating that transfers pretrained daily CVR/ATC representations based on real‑time pre‑promotion behavior to alleviate data sparsity; (3) a counterfactual causal regularizer built on a doubly‑robust estimator to align predictions with the causal effect of ATC on delayed conversion. A freeze/stop‑gradient strategy protects the pretrained tower from degrading the main task.

Results: Offline experiments show CM‑DCM markedly outperforms mainstream delayed‑feedback baselines and naive historical‑data reuse methods. Ablation studies confirm each component adds measurable gain. Online A/B testing during Double 11/12 promotions yields higher advertising revenue, delayed‑conversion GMV, and overall GMV, demonstrating real‑world effectiveness.

3. SIGMA – Semantic‑Grounded Instruction‑Driven Generative Multi‑Task Recommender – SIGIR 2026

Pain point: Collaborative‑filtering recommender systems lack semantic understanding and logical reasoning, leading to "filter‑bubble" effects, slow response to new trends, and inability to follow natural‑language business instructions.

Solution: SIGMA injects item and collaborative signals into a large language model (LLM) semantic space and employs an instruction‑following fine‑tuning regime. It builds a hybrid "semantic ID + item ID" encoding via contrastive learning and knowledge distillation, then applies multi‑task instruction fine‑tuning to endow the model with the ability to execute diverse recommendation tasks. Generation follows a three‑step pipeline: prefix generation → subset retrieval → probability fusion.

Results: Experiments demonstrate SIGMA significantly outperforms existing methods across various recommendation tasks. Online A/B tests show simultaneous improvements in efficiency and diversity metrics, confirming its practicality as a unified multi‑task recommendation backbone.

4. RCLRec – Reverse Curriculum Learning for Sparse Conversions – SIGIR 2026

Pain point: Generative recommendation (GR) models still struggle with extremely sparse conversion signals; behavior‑aware attention alone cannot surface the few decisive actions that drive purchase.

Solution: RCLRec constructs a conversion‑conditioned curriculum prefix (RCPM) by selecting the top‑k most relevant historical interactions for each conversion target, using a pay‑conditioned query. The curriculum quality‑aware loss enforces that the prefix improves the likelihood of the conversion token via a hinge‑style NLL comparison between "with curriculum" and "without curriculum". Training proceeds in two stages: (1) pre‑training on mixed‑behavior data to learn generic semantic tokens; (2) SFT on conversion samples with the curriculum prefix and quality loss.

Results: Offline experiments on two real datasets show RCLRec surpasses SASRec, BERT4Rec, TIGER, MBGen, and GEAR by 10‑14% in Recall@5/10 and NDCG@5/10. Ablations reveal the conversion‑aware prefix contributes the bulk of the gain, while the quality‑aware loss ensures the prefix truly helps conversion. Online A/B testing in an advertising scenario yields significant lifts in advertising revenue and order count.

5. LWGR – Lagrangian‑Constrained Personalized World Knowledge for Generative Recommendation – SIGIR 2026

Pain point: Fixed prompts cannot capture the multi‑dimensional heterogeneity of user interests, and LLM‑generated world knowledge injected without constraints can destabilize GR performance.

Solution: LWGR quantizes user context into personalized soft prompts, enabling LLMs to generate world knowledge tailored to individual behavior. Knowledge integration is formulated as a constrained optimization problem solved via a primal‑dual Lagrangian method that adaptively penalizes performance degradation, retaining only helpful knowledge for GR.

Results: On public benchmarks and large‑scale industrial data, LWGR achieves consistent improvements over strong baselines across multiple retrieval metrics. In real‑world advertising experiments, LWGR delivers noticeable gains in revenue, CVR, and ROI.

6. Scalable MoE Framework for Heterogeneous Experts – WWW 2026

Pain point: Cross‑border e‑commerce search must handle multilingual, short‑query, and semantic‑blur challenges. Scaling a single dense model yields diminishing returns and language‑specific performance imbalance.

Solution: The proposed coarse‑grained request‑level MoE coordinates multiple frozen heterogeneous LLM experts (e.g., Qwen, Gemma, Sailor). Dynamic sparse routing activates the top‑k experts per query, with hard end‑to‑end routing performing best. Instead of weighted averaging, each expert’s representation is projected to a unified dimension and concatenated, followed by a lightweight MLP that learns a nonlinear decision boundary to select the most appropriate expert.

Results: On Lazada’s six Southeast Asian markets, the MoE improves AUC by 0.72 pp to 92.49 and raises QPS to 13.72 (+9%) compared with dense baselines of similar parameter count. After online distillation to a ColBERT student, Bad Ratio drops from 9.00% to 8.55% and Online AUC rises from 90.39% to 91.28%, confirming the efficiency of routing heterogeneous experts.

7. Symmetric Masked Generative Paradigm for CTR – WWW 2026

Pain point: Generative CTR models suffer from training‑inference asymmetry; the generative capability is discarded during inference, preventing the model from leveraging learned distributional knowledge to handle noisy inputs.

Solution: SGCTR makes generation the core of inference. Training uses a discrete diffusion process to learn feature dependencies. In inference, an iterative refinement strategy treats raw features as noisy observations, repeatedly predicting and reconstructing them to denoise and improve representation. A cache‑based key/value computation reduces latency.

Results: SGCTR outperforms the SOTA DGenCTR across several benchmarks. The iterative refinement mitigates representation collapse, boosting robustness and accuracy. Online A/B tests show significant lifts in CTR and order volume while keeping inference latency within industrial production constraints.

8. MPCoT‑LRKD – Multi‑Perspective Chain‑of‑Thought Distillation for E‑Commerce Relevance – WWW 2026

Pain point: Deploying LLMs for relevance is prohibitive due to inference cost; single‑view reasoning cannot capture the complex user intent and business rules in e‑commerce, and traditional knowledge distillation discards the chain‑of‑thought reasoning.

Solution: MPCoT‑LRKD builds a multi‑view expert teacher that integrates user intent, structured analysis, and business rules via multi‑perspective chain‑of‑thought (MPCoT). Implicit reasoning distillation (LRKD) compresses the teacher’s reasoning into a lightweight student model, preserving the reasoning logic without explicit text generation.

Results: The approach significantly outperforms single‑view baselines and exceeds BERT baselines. Online A/B testing in a large‑scale e‑commerce search advertising system confirms superior performance and practical value.

9. MMQ – Multimodal Mixture‑of‑Quantization for Semantic ID Generation – WSDM 2026

Pain point: Semantic ID methods excel in storage efficiency but struggle with multimodal fusion and downstream alignment, limiting their representational power and task synergy.

Solution: MMQ introduces multimodal shared‑specific mixture experts that learn common and modality‑specific patterns, enforcing orthogonal expert parameters to keep codebooks low‑redundancy. A user‑behavior fine‑tuning stage uses Soft Indices to connect the tokenizer with recommendation tasks, narrowing the semantic‑behavior gap.

Results: MMQ outperforms prior semantic‑ID quantization methods in recall and ranking. Online experiments report significant gains in REV, CVR, GMV, and ROI. The work has been accepted at WSDM 2026.