Yingyi Zhang

Lei Zhao, Yingyi Zhang, Chengzhi Zhang

[Purpose] To understand the meaning of a sentence, humans can focus on important words in the sentence, which reflects our eyes staying on each word in different gaze time or times. Thus, some studies utilize eye-tracking values to optimize the attention mechanism in deep learning models. But these studies lack to explain the rationality of this approach. Whether the attention mechanism possesses this feature of human reading needs to be explored. [Design/methodology/approach] We conducted experiments on a sentiment classification task. Firstly, we obtained eye-tracking values from two open-source eye-tracking corpora to describe the feature of human reading. Then, the machine attention values of each sentence were learned from a sentiment classification model. Finally, a comparison was conducted to analyze machine attention values and eye-tracking values. [Findings] Through experiments, we found the attention mechanism can focus on important words, such as adjectives, adverbs, and sentiment words, which are valuable for judging the sentiment of sentences on the sentiment classification task. It possesses the feature of human reading, focusing on important words in sentences when reading. Due to the insufficient learning of the attention mechanism, some words are wrongly focused. The eye-tracking values can help the attention mechanism correct this error and improve the model performance. [Originality/value] Our research not only provides a reasonable explanation for the study of using eye-tracking values to optimize the attention mechanism, but also provides new inspiration for the interpretability of attention mechanism.

Pengyue Jia, Yingyi Zhang, Xiangyu Zhao et al.

Image geolocalization aims to predict the geographic location of images captured anywhere on Earth, but its global nature presents significant challenges. Current evaluation methodologies suffer from two major limitations. First, data leakage: advanced approaches often rely on large vision-language models (LVLMs) to predict image locations, yet these models are frequently pretrained on the test datasets, compromising the accuracy of evaluating a model's actual geolocalization capability. Second, existing metrics primarily rely on exact geographic coordinates to assess predictions, which not only neglects the reasoning process but also raises privacy concerns when user-level location data is required. To address these issues, we propose GeoArena, a first open platform for evaluating LVLMs on worldwide image geolocalization tasks, offering true in-the-wild and human-centered benchmarking. GeoArena enables users to upload in-the-wild images for a more diverse evaluation corpus, and it leverages pairwise human judgments to determine which model output better aligns with human expectations. Our platform has been deployed online for two months, during which we collected over thousands voting records. Based on this data, we conduct a detailed analysis and establish a leaderboard of different LVLMs on the image geolocalization task. GeoArena has been open-sourced to support future research.

Qihang Ma, Shengyu Li, Jie Tang et al.

Multi-modal keyphrase prediction (MMKP) aims to advance beyond text-only methods by incorporating multiple modalities of input information to produce a set of conclusive phrases. Traditional multi-modal approaches have been proven to have significant limitations in handling the challenging absence and unseen scenarios. Additionally, we identify shortcomings in existing benchmarks that overestimate model capability due to significant overlap in training tests. In this work, we propose leveraging vision-language models (VLMs) for the MMKP task. Firstly, we use two widely-used strategies, e.g., zero-shot and supervised fine-tuning (SFT) to assess the lower bound performance of VLMs. Next, to improve the complex reasoning capabilities of VLMs, we adopt Fine-tune-CoT, which leverages high-quality CoT reasoning data generated by a teacher model to finetune smaller models. Finally, to address the "overthinking" phenomenon, we propose a dynamic CoT strategy which adaptively injects CoT data during training, allowing the model to flexibly leverage its reasoning capabilities during the inference stage. We evaluate the proposed strategies on various datasets and the experimental results demonstrate the effectiveness of the proposed approaches. The code is available at https://github.com/bytedance/DynamicCoT.

Derong Xu, Pengyue Jia, Xiaopeng Li et al.

Large language models (LLMs) have demonstrated remarkable capabilities, but still struggle with issues like hallucinations and outdated information. Retrieval-augmented generation (RAG) addresses these issues by grounding LLM outputs in external knowledge with an Information Retrieval (IR) system. Building on this foundation, graph-based RAG systems go a step further by retrieving subgraphs, which preserve the relationships between knowledge entities and provide more comprehensive context. However, graph RAG faces two challenges: (1) Retrieving relevant information introduces irrelevant nodes (especially in dense graph databases, where retrieval usually extends to adjacent nodes), and leads to overly lengthy inputs that hinder efficiency; (2) The representation gap between graph and language during generation with LLMs limits the ability to fully leverage graph structures for enhanced understanding. To address these limitations, we propose Align-GRAG, a novel reasoning-guided dual alignment framework in post-retrieval phrase. It first formulates a subgraph by retrieving nodes and edges. Then an Aligner is proposed to jointly optimize a graph encoder with an LLM-summarized reasoning chain. It achieves dual alignment of graph node and representation by leveraging KL divergence loss and contrastive loss, facilitating efficient pruning of irrelevant knowledge and establishing a unified semantic space. The Generator integrates the aligned graph data with LLM to produce coherent and accurate answers. Experiments on the GraphQA benchmark across three tasks (including common sense reasoning, scene graph understanding, and knowledge graph reasoning) validate the effectiveness of our method. The codes are available in this repository\footnote{https://anonymous.4open.science/r/Align-GRAG-F3D8/}.

Fatemeh Akbarian, Anahita Baninajjar, Yingyi Zhang et al.

Multi-modal foundation models align images, text, and other modalities in a shared embedding space but remain vulnerable to adversarial illusions (Zhang et al., 2025), where imperceptible perturbations disrupt cross-modal alignment and mislead downstream tasks. To counteract the effects of adversarial illusions, we propose a task-agnostic mitigation mechanism that reconstructs the input from the attacker's perturbed input through generative models, e.g., Variational Autoencoders (VAEs), to maintain natural alignment. To further enhance our proposed defense mechanism, we adopt a generative sampling strategy combined with a consensus-based aggregation scheme over the outcomes of the generated samples. Our experiments on the state-of-the-art multi-modal encoders show that our approach substantially reduces the illusion attack success rates to near-zero and improves cross-modal alignment by 4% (42 to 46) and 11% (32 to 43) in unperturbed and perturbed input settings respectively, providing an effective and model-agnostic defense against adversarial illusions.

Chengzhi Zhang, Lei Zhao, Mengyuan Zhao et al.

With the development of Internet technology, the phenomenon of information overload is becoming more and more obvious. It takes a lot of time for users to obtain the information they need. However, keyphrases that summarize document information highly are helpful for users to quickly obtain and understand documents. For academic resources, most existing studies extract keyphrases through the title and abstract of papers. We find that title information in references also contains author-assigned keyphrases. Therefore, this article uses reference information and applies two typical methods of unsupervised extraction methods (TF*IDF and TextRank), two representative traditional supervised learning algorithms (Naïve Bayes and Conditional Random Field) and a supervised deep learning model (BiLSTM-CRF), to analyze the specific performance of reference information on keyphrase extraction. It is expected to improve the quality of keyphrase recognition from the perspective of expanding the source text. The experimental results show that reference information can increase precision, recall, and F1 of automatic keyphrase extraction to a certain extent. This indicates the usefulness of reference information on keyphrase extraction of academic papers and provides a new idea for the following research on automatic keyphrase extraction.