Yangyang Guo

Harry Cheng, Yangyang Guo, Liqiang Nie et al.

Training an effective video action recognition model poses significant computational challenges, particularly under limited resource budgets. Current methods primarily aim to either reduce model size or utilize pre-trained models, limiting their adaptability to various backbone architectures. This paper investigates the issue of over-sampled frames, a prevalent problem in many approaches yet it has received relatively little attention. Despite the use of fewer frames being a potential solution, this approach often results in a substantial decline in performance. To address this issue, we propose a novel method to restore the intermediate features for two sparsely sampled and adjacent video frames. This feature restoration technique brings a negligible increase in computational requirements compared to resource-intensive image encoders, such as ViT. To evaluate the effectiveness of our method, we conduct extensive experiments on four public datasets, including Kinetics-400, ActivityNet, UCF-101, and HMDB-51. With the integration of our method, the efficiency of three commonly used baselines has been improved by over 50%, with a mere 0.5% reduction in recognition accuracy. In addition, our method also surprisingly helps improve the generalization ability of the models under zero-shot settings.

Yangyang Guo, Guangzhi Wang, Mohan Kankanhalli

Applying a pre-trained large model to downstream tasks is prohibitive under resource-constrained conditions. Recent dominant approaches for addressing efficiency issues involve adding a few learnable parameters to the fixed backbone model. This strategy, however, leads to more challenges in loading large models for downstream fine-tuning with limited resources. In this paper, we propose a novel method for increasing the parameter efficiency of pre-trained models by introducing an intermediate pre-training stage. To this end, we first employ low-rank approximation to compress the original large model and then devise a feature distillation module and a weight perturbation regularization module. These modules are specifically designed to enhance the low-rank model. In particular, we update only the low-rank model while freezing the backbone parameters during pre-training. This allows for direct and efficient utilization of the low-rank model for downstream fine-tuning tasks. The proposed method achieves both efficiencies in terms of required parameters and computation time while maintaining comparable results with minimal modifications to the backbone architecture. Specifically, when applied to three vision-only and one vision-language Transformer models, our approach often demonstrates a merely $\sim$0.6 point decrease in performance while reducing the original parameter size by 1/3 to 2/3.

Harry Cheng, Yangyang Guo, Qingpei Guo et al.

Multi-modal Large Language Models (MLLMs) have dramatically advanced the research field and delivered powerful vision-language understanding capabilities. However, these models often inherit deep-rooted social biases from their training data, leading to uncomfortable responses with respect to attributes such as race and gender. This paper addresses the issue of social biases in MLLMs by i) introducing a comprehensive counterfactual dataset with multiple social concepts (CMSC), which complements existing datasets by providing 18 diverse and balanced social concepts; and ii) proposing a counter-stereotype debiasing (CSD) strategy that mitigates social biases in MLLMs by leveraging the opposites of prevalent stereotypes. CSD incorporates both a novel bias-aware data sampling method and a loss rescaling method, enabling the model to effectively reduce biases. We conduct extensive experiments with four prevalent MLLM architectures. The results demonstrate the advantage of the CMSC dataset and the edge of CSD strategy in reducing social biases compared to existing competing methods, without compromising the overall performance on general multi-modal reasoning benchmarks.

Yiming Yang, Yangyang Guo, Hui Lu et al.

Recently, Large Vision-Language Models (LVLMs) have made significant strides across diverse multimodal tasks and benchmarks. This paper reveals a largely under-explored problem from existing video-involved LVLMs - language bias, where models tend to prioritize language over video and thus result in incorrect responses. To address this research gap, we first collect a Video Language Bias Evaluation Benchmark, which is specifically designed to assess the language bias in video-involved LVLMs through two key tasks: ambiguous video contrast and interrogative question probing. Accordingly, we design accompanied evaluation metrics that aim to penalize LVLMs being biased by language. In addition, we also propose Multi-branch Contrastive Decoding (MCD), introducing two expert branches to simultaneously counteract language bias potentially generated by the amateur text-only branch. Our experiments demonstrate that i) existing video-involved LVLMs, including both proprietary and open-sourced, are largely limited by the language bias problem; ii) our MCD can effectively mitigate this issue and maintain general-purpose capabilities in various video-involved LVLMs without any additional retraining or alteration to model architectures.

Wey Yeh Choong, Yangyang Guo, Mohan Kankanhalli

Vision Large Language Models (VLLMs) are widely acknowledged to be prone to hallucinations. Existing research addressing this problem has primarily been confined to image inputs, with limited exploration of video-based hallucinations. Furthermore, current evaluation methods fail to capture nuanced errors in generated responses, which are often exacerbated by the rich spatiotemporal dynamics of videos. To address this, we introduce VidHal, a benchmark specially designed to evaluate video-based hallucinations in VLLMs. VidHal is constructed by bootstrapping video instances across a wide range of common temporal aspects. A defining feature of our benchmark lies in the careful creation of captions which represent varying levels of hallucination associated with each video. To enable fine-grained evaluation, we propose a novel caption ordering task requiring VLLMs to rank captions by hallucinatory extent. We conduct extensive experiments on VidHal and comprehensively evaluate a broad selection of models. Our results uncover significant limitations in existing VLLMs regarding hallucination generation. Through our benchmark, we aim to inspire further research on 1) holistic understanding of VLLM capabilities, particularly regarding hallucination, and 2) extensive development of advanced VLLMs to alleviate this problem.

Haoyu Zhang, Yangyang Guo, Mohan Kankanhalli

Vision-Language (V-L) pre-trained models such as CLIP show prominent capabilities in various downstream tasks. Despite this promise, V-L models are notoriously limited by their inherent social biases. A typical demonstration is that V-L models often produce biased predictions against specific groups of people, significantly undermining their real-world applicability. Existing approaches endeavor to mitigate the social bias problem in V-L models by removing biased attribute information from model embeddings. However, after our revisiting of these methods, we find that their bias removal is frequently accompanied by greatly compromised V-L alignment capabilities. We then reveal that this performance degradation stems from the unbalanced debiasing in image and text embeddings. To address this issue, we propose a novel V-L debiasing framework to align image and text biases followed by removing them from both modalities. By doing so, our method achieves multi-modal bias mitigation while maintaining the V-L alignment in the debiased embeddings. Additionally, we advocate a new evaluation protocol that can 1) holistically quantify the model debiasing and V-L alignment ability, and 2) evaluate the generalization of social bias removal models. We believe this work will offer new insights and guidance for future studies addressing the social bias problem in CLIP.

Sizhe Dang, Yangyang Guo, Yanjun Zhao et al.

Fine-tuning large language models (LLMs) often faces GPU memory bottlenecks: the backward pass of first-order optimizers like Adam increases memory usage to more than 10 times the inference level (e.g., 633 GB for OPT-30B). Zeroth-order (ZO) optimizers avoid this cost by estimating gradients only from forward passes, yet existing methods like MeZO usually require many more steps to converge. Can this trade-off between speed and memory in ZO be fundamentally improved? Normalized-SGD demonstrates strong empirical performance with greater memory efficiency than Adam. In light of this, we introduce FZOO, a Fast Zeroth-Order Optimizer toward Adam-Scale Speed. FZOO reduces the total forward passes needed for convergence by employing batched one-sided estimates that adapt step sizes based on the standard deviation of batch losses. It also accelerates per-batch computation through the use of Rademacher random vector perturbations coupled with CUDA's parallel processing. Extensive experiments on diverse models, including RoBERTa-large, OPT (350M-66B), Phi-2, and Llama3, across 11 tasks validate FZOO's effectiveness. On average, FZOO outperforms MeZO by 3 percent in accuracy while requiring 3 times fewer forward passes. For RoBERTa-large, FZOO achieves average improvements of 5.6 percent in accuracy and an 18 times reduction in forward passes compared to MeZO, achieving convergence speeds comparable to Adam. We also provide theoretical analysis proving FZOO's formal equivalence to a normalized-SGD update rule and its convergence guarantees. FZOO integrates smoothly into PEFT techniques, enabling even larger memory savings. Overall, our results make single-GPU, high-speed, full-parameter fine-tuning practical and point toward future work on memory-efficient pre-training.

Yangyang Guo, Mohan Kankanhalli

While contrastive pre-training is widely employed, its data efficiency problem has remained relatively under-explored thus far. Existing methods often rely on static coreset selection algorithms to pre-identify important data for training. However, this static nature renders them unable to dynamically track the data usefulness throughout pre-training, leading to subpar pre-trained models. To address this challenge, our paper introduces a novel dynamic bootstrapping dataset pruning method. It involves pruning data preparation followed by dataset mutation operations, both of which undergo iterative and dynamic updates. We apply this method to two prevalent contrastive pre-training frameworks: \textbf{CLIP} and \textbf{MoCo}, representing vision-language and vision-centric domains, respectively. In particular, we individually pre-train seven CLIP models on two large-scale image-text pair datasets, and two MoCo models on the ImageNet dataset, resulting in a total of 16 pre-trained models. With a data pruning rate of 30-35\% across all 16 models, our method exhibits only marginal performance degradation (less than \textbf{1\%} on average) compared to corresponding models trained on the full dataset counterparts across various downstream datasets, and also surpasses several baselines with a large performance margin. Additionally, the byproduct from our method, \ie coresets derived from the original datasets after pre-training, also demonstrates significant superiority in terms of downstream performance over other static coreset selection approaches.

Yangyang Guo, Yangyan Li, Mohan Kankanhalli

In this study, we disclose a worrying new vulnerability in Large Language Models (LLMs), which we term \textbf{involuntary jailbreak}. Unlike existing jailbreak attacks, this weakness is distinct in that it does not involve a specific attack objective, such as generating instructions for \textit{building a bomb}. Prior attack methods predominantly target localized components of the LLM guardrail. In contrast, involuntary jailbreaks may potentially compromise the entire guardrail structure, which our method reveals to be surprisingly fragile. We merely employ a single universal prompt to achieve this goal. In particular, we instruct LLMs to generate several questions that would typically be rejected, along with their corresponding in-depth responses (rather than a refusal). Remarkably, this simple prompt strategy consistently jailbreaks the majority of leading LLMs, including Claude Opus 4.1, Grok 4, Gemini 2.5 Pro, and GPT 4.1. We hope this problem can motivate researchers and practitioners to re-evaluate the robustness of LLM guardrails and contribute to stronger safety alignment in future.

Yangyang Guo, Ziwei Xu, Xilie Xu et al.

This technical report introduces our top-ranked solution that employs two approaches, \ie suffix injection and projected gradient descent (PGD) , to address the TiFA workshop MLLM attack challenge. Specifically, we first append the text from an incorrectly labeled option (pseudo-labeled) to the original query as a suffix. Using this modified query, our second approach applies the PGD method to add imperceptible perturbations to the image. Combining these two techniques enables successful attacks on the LLaVA 1.5 model.

Xiangkun Yin, Yangyang Guo, Liqiang Nie et al.

Personalization lies at the core of boosting the product search system performance. Prior studies mainly resorted to the semantic matching between textual queries and user/product related documents, leaving the user collaborative behaviors untapped. In fact, the collaborative filtering signals between users intuitively offer a complementary information for the semantic matching. To close the gap between collaborative filtering and product search, we propose a Hierarchical Heterogeneous Graph Neural Network (HHGNN) approach in this paper. Specifically, we organize HHGNN with a hierarchical graph structure according to the three edge types. The sequence edge accounts for the syntax formulation from word nodes to sentence nodes; the composition edge aggregates the semantic features to the user and product nodes; and the interaction edge on the top performs graph convolutional operation between user and product nodes. At last, we integrate the higher-order neighboring collaborative features and the semantic features for better representation learning. We conduct extensive experiments on six Amazon review datasets. The results show that our proposed method can outperform the state-of-the-art baselines with a large margin. In addition, we empirically prove that collaborative filtering and semantic matching are complementary to each other in product search performance enhancement.

Fangkai Jiao, Yangyang Guo, Yilin Niu et al.

Pre-trained Language Models (PLMs) have achieved great success on Machine Reading Comprehension (MRC) over the past few years. Although the general language representation learned from large-scale corpora does benefit MRC, the poor support in evidence extraction which requires reasoning across multiple sentences hinders PLMs from further advancing MRC. To bridge the gap between general PLMs and MRC, we present REPT, a REtrieval-based Pre-Training approach. In particular, we introduce two self-supervised tasks to strengthen evidence extraction during pre-training, which is further inherited by downstream MRC tasks through the consistent retrieval operation and model architecture. To evaluate our proposed method, we conduct extensive experiments on five MRC datasets that require collecting evidence from and reasoning across multiple sentences. Experimental results demonstrate the effectiveness of our pre-training approach. Moreover, further analysis shows that our approach is able to enhance the capacity of evidence extraction without explicit supervision.

Yangyang Guo, Liqiang Nie, Zhiyong Cheng et al.

A number of studies point out that current Visual Question Answering (VQA) models are severely affected by the language prior problem, which refers to blindly making predictions based on the language shortcut. Some efforts have been devoted to overcoming this issue with delicate models. However, there is no research to address it from the angle of the answer feature space learning, despite of the fact that existing VQA methods all cast VQA as a classification task. Inspired by this, in this work, we attempt to tackle the language prior problem from the viewpoint of the feature space learning. To this end, an adapted margin cosine loss is designed to discriminate the frequent and the sparse answer feature space under each question type properly. As a result, the limited patterns within the language modality are largely reduced, thereby less language priors would be introduced by our method. We apply this loss function to several baseline models and evaluate its effectiveness on two VQA-CP benchmarks. Experimental results demonstrate that our adapted margin cosine loss can greatly enhance the baseline models with an absolute performance gain of 15\% on average, strongly verifying the potential of tackling the language prior problem in VQA from the angle of the answer feature space learning.

Shaobo Cui, Xintong Bao, Xinxing Zu et al.

Large-scale question-answer (QA) pairs are critical for advancing research areas like machine reading comprehension and question answering. To construct QA pairs from documents requires determining how to ask a question and what is the corresponding answer. Existing methods for QA pair generation usually follow a pipeline approach. Namely, they first choose the most likely candidate answer span and then generate the answer-specific question. This pipeline approach, however, is undesired in mining the most appropriate QA pairs from documents since it ignores the connection between question generation and answer extraction, which may lead to incompatible QA pair generation, i.e., the selected answer span is inappropriate for question generation. However, for human annotators, we take the whole QA pair into account and consider the compatibility between question and answer. Inspired by such motivation, instead of the conventional pipeline approach, we propose a model named OneStop generate QA pairs from documents in a one-stop approach. Specifically, questions and their corresponding answer span is extracted simultaneously and the process of question generation and answer extraction mutually affect each other. Additionally, OneStop is much more efficient to be trained and deployed in industrial scenarios since it involves only one model to solve the complex QA generation task. We conduct comprehensive experiments on three large-scale machine reading comprehension datasets: SQuAD, NewsQA, and DuReader. The experimental results demonstrate that our OneStop model outperforms the baselines significantly regarding the quality of generated questions, quality of generated question-answer pairs, and model efficiency.

Yangyang Guo, Liqiang Nie, Zhiyong Cheng et al.

Recent studies have pointed out that many well-developed Visual Question Answering (VQA) models are heavily affected by the language prior problem, which refers to making predictions based on the co-occurrence pattern between textual questions and answers instead of reasoning visual contents. To tackle it, most existing methods focus on enhancing visual feature learning to reduce this superficial textual shortcut influence on VQA model decisions. However, limited effort has been devoted to providing an explicit interpretation for its inherent cause. It thus lacks a good guidance for the research community to move forward in a purposeful way, resulting in model construction perplexity in overcoming this non-trivial problem. In this paper, we propose to interpret the language prior problem in VQA from a class-imbalance view. Concretely, we design a novel interpretation scheme whereby the loss of mis-predicted frequent and sparse answers of the same question type is distinctly exhibited during the late training phase. It explicitly reveals why the VQA model tends to produce a frequent yet obviously wrong answer, to a given question whose right answer is sparse in the training set. Based upon this observation, we further develop a novel loss re-scaling approach to assign different weights to each answer based on the training data statistics for computing the final loss. We apply our approach into three baselines and the experimental results on two VQA-CP benchmark datasets evidently demonstrate its effectiveness. In addition, we also justify the validity of the class imbalance interpretation scheme on other computer vision tasks, such as face recognition and image classification.

Yangyang Guo, Zhiyong Cheng, Jiazheng Jing et al.

Factorization Machines (FMs) are effective in incorporating side information to overcome the cold-start and data sparsity problems in recommender systems. Traditional FMs adopt the inner product to model the second-order interactions between different attributes, which are represented via feature vectors. The problem is that the inner product violates the triangle inequality property of feature vectors. As a result, it cannot well capture fine-grained attribute interactions, resulting in sub-optimal performance. Recently, the Euclidean distance is exploited in FMs to replace the inner product and has delivered better performance. However, previous FM methods including the ones equipped with the Euclidean distance all focus on the attribute-level interaction modeling, ignoring the critical intrinsic feature correlations inside attributes. Thereby, they fail to model the complex and rich interactions exhibited in the real-world data. To tackle this problem, in this paper, we propose a FM framework equipped with generalized metric learning techniques to better capture these feature correlations. In particular, based on this framework, we present a Mahalanobis distance and a deep neural network (DNN) methods, which can effectively model the linear and non-linear correlations between features, respectively. Besides, we design an efficient approach for simplifying the model functions. Experiments on several benchmark datasets demonstrate that our proposed framework outperforms several state-of-the-art baselines by a large margin. Moreover, we collect a new large-scale dataset on second-hand trading to justify the effectiveness of our method over cold-start and data sparsity problems in recommender systems.

Yangyang Guo, Zhiyong Cheng, Liqiang Nie et al.

Benefiting from the advancement of computer vision, natural language processing and information retrieval techniques, visual question answering (VQA), which aims to answer questions about an image or a video, has received lots of attentions over the past few years. Although some progress has been achieved so far, several studies have pointed out that current VQA models are heavily affected by the \emph{language prior problem}, which means they tend to answer questions based on the co-occurrence patterns of question keywords (e.g., how many) and answers (e.g., 2) instead of understanding images and questions. Existing methods attempt to solve this problem by either balancing the biased datasets or forcing models to better understand images. However, only marginal effects and even performance deterioration are observed for the first and second solution, respectively. In addition, another important issue is the lack of measurement to quantitatively measure the extent of the language prior effect, which severely hinders the advancement of related techniques. In this paper, we make contributions to solve the above problems from two perspectives. Firstly, we design a metric to quantitatively measure the language prior effect of VQA models. The proposed metric has been demonstrated to be effective in our empirical studies. Secondly, we propose a regularization method (i.e., score regularization module) to enhance current VQA models by alleviating the language prior problem as well as boosting the backbone model performance. The proposed score regularization module adopts a pair-wise learning strategy, which makes the VQA models answer the question based on the reasoning of the image (upon this question) instead of basing on question-answer patterns observed in the biased training set. The score regularization module is flexible to be integrated into various VQA models.

Yangyang Guo, Zhiyong Cheng, Liqiang Nie et al.

E-commerce users may expect different products even for the same query, due to their diverse personal preferences. It is well-known that there are two types of preferences: long-term ones and short-term ones. The former refers to user' inherent purchasing bias and evolves slowly. By contrast, the latter reflects users' purchasing inclination in a relatively short period. They both affect users' current purchasing intentions. However, few research efforts have been dedicated to jointly model them for the personalized product search. To this end, we propose a novel Attentive Long Short-Term Preference model, dubbed as ALSTP, for personalized product search. Our model adopts the neural networks approach to learn and integrate the long- and short-term user preferences with the current query for the personalized product search. In particular, two attention networks are designed to distinguish which factors in the short-term as well as long-term user preferences are more relevant to the current query. This unique design enables our model to capture users' current search intentions more accurately. Our work is the first to apply attention mechanisms to integrate both long- and short-term user preferences with the given query for the personalized search. Extensive experiments over four Amazon product datasets show that our model significantly outperforms several state-of-the-art product search methods in terms of different evaluation metrics.