Selected work

Representative Papers

Factorized Transport Alignment for Multimodal and Multiview E-commerce Representation Learning

Dec 19, 2025

Existing vision-language models (VLMs) align only image titles with primary images, neglecting the critical semantic information conveyed by non-primary images and auxiliary textual modalities (e.g., descriptions, tags) in e-commerce scenarios—thereby limiting multimodal, multi-view representation learning. To address this, we propose Factorized Transport: a lightweight, factorized optimal transport approximation method designed for open-platform e-commerce. It enables scalable multi-view alignment—spanning primary/auxiliary images and titles/descriptions/tags—while supporting zero-overhead online inference fusion. Our approach integrates stochastic view sampling, dual-tower embedding caching, and multi-view contrastive learning. Evaluated on a million-scale industrial product dataset, it achieves a +7.9% improvement in Recall@500 over strong multimodal baselines, demonstrating both effectiveness and deployability for large-scale, real-time e-commerce search.

0 citationsRead paper

OptAgent: Optimizing Query Rewriting for E-commerce via Multi-Agent Simulation

Oct 04, 2025

E-commerce query rewriting (QR) suffers from subjective intent evaluation and the absence of reliable automatic evaluation metrics. To address this, we propose a multi-LLM-agent-driven dynamic evolutionary optimization framework: a multi-agent system simulating realistic user shopping behavior generates interactive, fine-grained feedback—replacing static scoring models as reward signals—and integrates genetic algorithms to close the loop between iterative query generation and evaluation. Evaluated on 1,000 real-world e-commerce queries, our method improves over the original queries by an average of 21.98%, significantly outperforming the Best-of-N baseline by 3.36%. Our core contribution is the first integration of collaborative multi-agent feedback mechanisms with evolutionary search, enabling end-to-end, intent-driven, and fully learnable QR optimization.

0 citationsRead paper

LLM-Enhanced Self-Evolving Reinforcement Learning for Multi-Step E-Commerce Payment Fraud Risk Detection

Sep 23, 2025

In multi-step e-commerce payment fraud detection, conventional reinforcement learning approaches rely on manually designed reward functions, resulting in poor generalization across scenarios. Method: This paper proposes an LLM-driven multi-stage risk detection framework. By modeling the transaction process as a Markov Decision Process (MDP), it leverages large language models to autonomously generate and iteratively refine high-order logical reward functions—enabling fully automated, human-in-the-loop-free reward evolution. Contribution/Results: The framework exhibits zero-shot transfer capability, significantly enhancing cross-scenario adaptability and robustness. Long-term evaluation on real-world industrial data demonstrates substantial improvements in fraud detection accuracy while maintaining a low false positive rate and strong resilience to adversarial perturbations, confirming its practical deployability.

0 citationsRead paper

Recent publications

Latest Papers

Factorized Transport Alignment for Multimodal and Multiview E-commerce Representation Learning

Dec 19, 2025

Existing vision-language models (VLMs) align only image titles with primary images, neglecting the critical semantic information conveyed by non-primary images and auxiliary textual modalities (e.g., descriptions, tags) in e-commerce scenarios—thereby limiting multimodal, multi-view representation learning. To address this, we propose Factorized Transport: a lightweight, factorized optimal transport approximation method designed for open-platform e-commerce. It enables scalable multi-view alignment—spanning primary/auxiliary images and titles/descriptions/tags—while supporting zero-overhead online inference fusion. Our approach integrates stochastic view sampling, dual-tower embedding caching, and multi-view contrastive learning. Evaluated on a million-scale industrial product dataset, it achieves a +7.9% improvement in Recall@500 over strong multimodal baselines, demonstrating both effectiveness and deployability for large-scale, real-time e-commerce search.

0 citationsRead paper

OptAgent: Optimizing Query Rewriting for E-commerce via Multi-Agent Simulation

Oct 04, 2025

E-commerce query rewriting (QR) suffers from subjective intent evaluation and the absence of reliable automatic evaluation metrics. To address this, we propose a multi-LLM-agent-driven dynamic evolutionary optimization framework: a multi-agent system simulating realistic user shopping behavior generates interactive, fine-grained feedback—replacing static scoring models as reward signals—and integrates genetic algorithms to close the loop between iterative query generation and evaluation. Evaluated on 1,000 real-world e-commerce queries, our method improves over the original queries by an average of 21.98%, significantly outperforming the Best-of-N baseline by 3.36%. Our core contribution is the first integration of collaborative multi-agent feedback mechanisms with evolutionary search, enabling end-to-end, intent-driven, and fully learnable QR optimization.

0 citationsRead paper

LLM-Enhanced Self-Evolving Reinforcement Learning for Multi-Step E-Commerce Payment Fraud Risk Detection

Sep 23, 2025

In multi-step e-commerce payment fraud detection, conventional reinforcement learning approaches rely on manually designed reward functions, resulting in poor generalization across scenarios. Method: This paper proposes an LLM-driven multi-stage risk detection framework. By modeling the transaction process as a Markov Decision Process (MDP), it leverages large language models to autonomously generate and iteratively refine high-order logical reward functions—enabling fully automated, human-in-the-loop-free reward evolution. Contribution/Results: The framework exhibits zero-shot transfer capability, significantly enhancing cross-scenario adaptability and robustness. Long-term evaluation on real-world industrial data demonstrates substantial improvements in fraud detection accuracy while maintaining a low false positive rate and strong resilience to adversarial perturbations, confirming its practical deployability.

0 citationsRead paper