Jiangnan Li

Chengyang Fang, Jiangnan Li, Liang Li et al.

Text-based Visual Question Answering (TextVQA) aims at answering questions about the text in images. Most works in this field focus on designing network structures or pre-training tasks. All these methods list the OCR texts in reading order (from left to right and top to bottom) to form a sequence, which is treated as a natural language ``sentence''. However, they ignore the fact that most OCR words in the TextVQA task do not have a semantical contextual relationship. In addition, these approaches use 1-D position embedding to construct the spatial relation between OCR tokens sequentially, which is not reasonable. The 1-D position embedding can only represent the left-right sequence relationship between words in a sentence, but not the complex spatial position relationship. To tackle these problems, we propose a novel method named Separate and Locate (SaL) that explores text contextual cues and designs spatial position embedding to construct spatial relations between OCR texts. Specifically, we propose a Text Semantic Separate (TSS) module that helps the model recognize whether words have semantic contextual relations. Then, we introduce a Spatial Circle Position (SCP) module that helps the model better construct and reason the spatial position relationships between OCR texts. Our SaL model outperforms the baseline model by 4.44% and 3.96% accuracy on TextVQA and ST-VQA datasets. Compared with the pre-training state-of-the-art method pre-trained on 64 million pre-training samples, our method, without any pre-training tasks, still achieves 2.68% and 2.52% accuracy improvement on TextVQA and ST-VQA. Our code and models will be released at https://github.com/fangbufang/SaL.

Junjie Wu, Jiangnan Li, Yuqing Li et al.

Retrieval-augmented generation (RAG) over long documents typically involves splitting the text into smaller chunks, which serve as the basic units for retrieval. However, due to dependencies across the original document, contextual information is often essential for accurately interpreting each chunk. To address this, prior work has explored encoding longer context windows to produce embeddings for longer chunks. Despite these efforts, gains in retrieval and downstream tasks remain limited. This is because (1) longer chunks strain the capacity of embedding models due to the increased amount of information they must encode, and (2) many real-world applications still require returning localized evidence due to constraints on model or human bandwidth. We propose an alternative approach to this challenge by representing short chunks in a way that is conditioned on a broader context window to enhance retrieval performance -- i.e., situating a chunk's meaning within its context. We further show that existing embedding models are not well-equipped to encode such situated context effectively, and thus introduce a new training paradigm and develop the situated embedding models (SitEmb). To evaluate our method, we curate a book-plot retrieval dataset specifically designed to assess situated retrieval capabilities. On this benchmark, our SitEmb-v1 model based on BGE-M3 substantially outperforms state-of-the-art embedding models, including several with up to 7-8B parameters, with only 1B parameters. Our 8B SitEmb-v1.5 model further improves performance by over 10% and shows strong results across different languages and several downstream applications.

Jinfu Fan, Xiaohui Zhong, Kangrui Ren et al.

Learning from ambiguous labels is a long-standing problem in practical machine learning applications. The purpose of \emph{partial label learning} (PLL) is to identify the ground-truth label from a set of candidate labels associated with a given instance. Inspired by the remarkable performance of diffusion models in various generation tasks, this paper explores their potential to denoise ambiguous labels through the reverse denoising process. Therefore, this paper reformulates the label disambiguation problem from the perspective of generative models, where labels are generated by iteratively refining initial random guesses. This perspective enables the diffusion model to learn how label information is generated stochastically. By modeling the generation uncertainty, we can use the maximum likelihood estimate of the label for classification inference. However, such ambiguous labels lead to a mismatch between instance and label, which reduces the quality of generated data. To address this issue, this paper proposes a \emph{diffusion disambiguation model for PLL} (DDMP), which first uses the potential complementary information between instances and labels to construct pseudo-clean labels for initial diffusion training. Furthermore, a transition-aware matrix is introduced to estimate the potential ground-truth labels, which are dynamically updated during the diffusion generation. During training, the ground-truth label is progressively refined, improving the classifier. Experiments show the advantage of the DDMP and its suitability for PLL.

Jiangnan Li, Thuy-Trang Vu, Christian Herold et al.

Naive joint training of large language models (LLMs) for multilingual preference alignment can suffer from negative interference. This is a known issue in multilingual training, where conflicting objectives degrade overall performance. However, the impact of this phenomenon in the context of multilingual preference alignment remains largely underexplored. To address this issue, we propose CONGRAD, a scalable and effective filtering method that selects high-quality preference samples with minimal gradient conflicts across languages. Our method leverages gradient surgery to retain samples aligned with an aggregated multilingual update direction. Additionally, we incorporate a sublinear gradient compression strategy that reduces memory overhead during gradient accumulation. We integrate CONGRAD into self-rewarding framework and evaluate on LLaMA3-8B and Gemma2-2B across 10 languages. Results show that CONGRAD consistently outperforms strong baselines in both seen and unseen languages, with minimal alignment tax.

Yingyuan Yang, Xueli Huang, Jiangnan Li et al.

Leveraging users' behavioral data sampled by various sensors during the identification process, implicit authentication (IA) relieves users from explicit actions such as remembering and entering passwords. Various IA schemes have been proposed based on different behavioral and contextual features such as gait, touch, and GPS. However, existing IA schemes suffer from false positives, i.e., falsely accepting an adversary, and false negatives, i.e., falsely rejecting the legitimate user due to users' behavior change and noise. To deal with this problem, we propose BubbleMap (BMap), a framework that can be seamlessly incorporated into any existing IA system to balance between security (reducing false positives) and usability (reducing false negatives) as well as reducing the equal error rate (EER). To evaluate the proposed framework, we implemented BMap on five state-of-the-art IA systems. We also conducted an experiment in a real-world environment from 2016 to 2020. Most of the experimental results show that BMap can greatly enhance the IA schemes' performances in terms of the EER, security, and usability, with a small amount of penalty on energy consumption.