Institution profile

Alibaba Group

Industry researchasia · cn
Official website
Research library2,267linked papers
Opportunities1,460open roles
Selected work

Representative Papers

Stick to the Facts: Learning towards a Fidelity-oriented E-Commerce Product Description Generation

Nov 01, 2019Conference on Empirical Methods in Natural Language Processing

To address factual inconsistency and attribute deviation in product description generation for e-commerce, this paper proposes an attribute-faithful controllable generation method. The core innovation lies in an entity-label-guided ELSTM decoder unit integrated with a keyword memory mechanism: fine-grained word-level entity labels serve as explicit control signals, jointly leveraged via label-aware attention and a dynamic keyword memory bank to enable attribute-aware generation. The method synergistically combines entity label embeddings, RNN-based architecture, and structured memory to significantly enhance the fidelity of generated text with respect to product attributes. Evaluated on a large-scale real-world e-commerce dataset, the approach achieves a 25% improvement in attribute fidelity over prior methods, while also attaining state-of-the-art performance on standard automatic metrics (BLEU, ROUGE) and human evaluations.

23 citations2 influentialRead paper

Qwen3-VL-Embedding and Qwen3-VL-Reranker: A Unified Framework for State-of-the-Art Multimodal Retrieval and Ranking

Jan 08, 2026arXiv.org

This work proposes an end-to-end multimodal retrieval system to address key challenges in cross-modal semantic alignment, limited embedding flexibility, and insufficient multilingual support. The system unifies text, images, document images, and videos into a shared representation space through a multi-stage training paradigm comprising contrastive pretraining followed by reranking distillation. It integrates a cross-attention reranking mechanism with Matryoshka representation learning to enable flexible-dimension embeddings, supports context lengths up to 32k tokens, and offers proficiency in over 30 languages. The resulting model, Qwen3-VL-Embedding-8B, achieves state-of-the-art performance across multiple tasks—including image-text retrieval, visual question answering, and video-text matching—attaining a leading score of 77.8 on the MMEB-V2 benchmark, significantly outperforming existing approaches.

9 citations3 influentialRead paper

Probe before You Talk: Towards Black-box Defense against Backdoor Unalignment for Large Language Models

Jun 19, 2025International Conference on Learning Representations

In black-box LLM-as-a-Service (LLMaaS) settings, stealthy backdoor alignment attacks—where models violate safety alignment upon inputs containing hidden triggers—are notoriously difficult to detect. Method: We propose BEAT, the first sample-agnostic, black-box detectable defense leveraging distortions in refusal signals. Its core insight is the “probe concatenation effect”: a stable, significant drop in refusal rate upon backdoor activation. Instead of analyzing output semantics, BEAT monitors the stability of safety signals via multi-sample output distribution estimation, probe concatenation perturbations, and KL-divergence–based distortion quantification—requiring neither gradients nor internal model access. Results: Evaluated on closed- and open-source models including GPT-3.5-turbo, BEAT achieves AUC > 0.96 in detecting diverse backdoor attacks and generalizes effectively against mainstream jailbreak techniques.

9 citationsRead paper

A Comprehensive Survey of Direct Preference Optimization: Datasets, Theories, Variants, and Applications

Oct 21, 2024

To address the high computational cost and reliance on reinforcement learning in RLHF-based alignment of large language models (LLMs), this paper presents a systematic survey of Direct Preference Optimization (DPO)—a reinforcement-learning-free alignment paradigm grounded solely in preference data. We introduce the first multidimensional taxonomy of DPO, unifying its theoretical foundations, algorithmic variants, benchmark datasets, and application domains. Through rigorous analysis grounded in Bradley–Terry modeling, loss function characterization, and data quality assessment, we empirically synthesize over 120 works to identify DPO’s convergence conditions, data sensitivity patterns, and scenario-specific adaptation strategies. Crucially, we uncover its fundamental theoretical limitations, training biases, and generalization bottlenecks for the first time. Finally, we propose three key future directions: scalability enhancement, robustness improvement, and multimodal extension—providing a principled methodological foundation for efficient, stable human preference alignment.

6 citationsRead paper
Recent publications

Latest Papers