For large-scale personalization scenarios, we study how to balance long-tail alleviation, explainability, and multi-objective optimization. Our systematic research spans graph-structure modeling, cross-topic/cross-modal augmentation, personalized contrastive learning, and causal debiasing, targeting key problems such as cold start, cross-domain generalization, and controllable exposure, validated against industrial data distributions and online metrics (CTR/CVR/GMV).

• Cross-subtopic graphs / multi-path retrieval and reranking for long-tail interest coverage
• Personalized contrastive learning with mutual-information constraints for robust user representations and interest disentanglement
• Causal debiasing (position/popularity/exposure bias) for explainable and controllable recommendation

Fusing text-image-audio-temporal signals, we study alignment, fusion, and robust representations. Combined with pretraining/instruction alignment and parameter-efficient fine-tuning (LoRA/QLoRA), we deliver end-to-end transferable multimodal capabilities for retrieval, QA, understanding, and generation.

• Modality alignment with mutual-information constraints for semantic consistency and fine-grained alignment
• Cross-modal retrieval/annotation/summarization with multi-task, multi-scenario transfer
• Compression, distillation, and robust training for long-tail and noisy scenarios

Toward collaborative decision-making and task decomposition, we study role assignment, communication mechanisms, and adaptive coordination of multiple agents. Combining LLM agents, hierarchical planning, and self-play/self-evolution, we build reliable and scalable collective intelligence for recommendation, retrieval, and autonomous systems.

• Expert-agent collaboration (plan-execute-evaluate-self-reflect)
• Communication protocols and credit assignment for stable cooperation and fair resource utilization
• Cross-task transfer and open-environment adaptation with safety and controllability mechanisms