Large Vision-Language Models (LVLMs) have shown impressive performances on multimodal tasks, yet remain fundamentally constrained by object hallucination, i.e., generating response inconsistent with visual input. To investigate its underlying causes, we conduct an empirical analysis of internal attention distribution in various LVLMs and reveal that the bias in attention distribution imbalance varies across different architectural models. Specifically, we introduce the attention look-back rate to design a contextual entropy metric, which quantifies the balance of attention distributions across different input components (including vision, instruction and prior-generated response). This metric helps distinguish between hallucinated and non-hallucinated tokens, offering a model-agnostic method to detect over-reliance on either visual or text modalities. In light of this insight, we propose Contextual Entropy Calibration (CEC), a training-free decoding strategy that dynamically incorporates contextual entropy during beam search to steer token selection toward more balanced attention across input components. Unlike existing approaches, CEC divides inputs at each generation step into three parts and synthesizes their effects, requiring no extra training or external models, with a maximum socre improvement of 1.5% on AMBER and 79 points on MME perception Benchmark compared to previous methods. Extensive experiments on multiple hallucination benchmarks demonstrate the effectiveness of our method in mitigating object hallucinations, highlighting its potential to improve the reliability of LVLMs generation.

Recently, neural module networks (NMNs) have yielded ongoing success in answering compositional visual questions, especially those involving multi-hop visual and logical reasoning. NMNs decompose the complex question into several sub-tasks using instance-modules from the reasoning paths of that question and then exploit intermediate supervisions to guide answer prediction, thereby improving inference interpretability. However, their performance may be hindered due to sketchy modeling of intermediate supervisions. For instance, (1) a prior assumption that each instance-module refers to only one grounded object yet overlooks other potentially associated grounded objects, impeding full cross-modal alignment learning; (2) IoU-based intermediate supervisions may introduce noise signals as the bounding box overlap issue might guide the model's focus towards irrelevant objects. To address these issues, a novel method, Detection-based Intermediate Supervision (DIS), is proposed, which adopts a generative detection framework to facilitate multiple grounding supervisions via sequence generation. As such, DIS offers more comprehensive and accurate intermediate supervisions, thereby boosting answer prediction performance. Furthermore, by considering intermediate results, DIS enhances the consistency in answering compositional questions and their sub-questions. Extensive experiments demonstrate the superiority of our proposed DIS, showcasing both improved accuracy and state-of-the-art reasoning consistency compared to prior approaches.

The Visual Question Answering (VQA) task has recently become notorious because models are prone to predicting well-educated "guesses" as answers rather than deriving them through visual understanding. The main culprit for this is that VQA models memorize the shortcut biases in the dataset during the training process. While a variety of solutions have been proposed, they solely focus on the shortcuts in the language modality, leaving other kinds of shortcut biases untouched. In this paper, we shift our lens to all kinds of shortcuts and resort to causal inference to circumvent these issues. Causal inference methods can discover the causal effect (P(Y|do(X))) [27] rather than statistic-based spurious correlations (P(Y|X)) in the dataset, making them naturally suitable for debiasing learning. To deconfound these shortcut biases, we propose a causality-aware method, coined as Dual Causal Intervention (DCI), to endow VQA models with better generalization by combining two components: linguistic backdoor intervention and visual front-door intervention. To be specific, we harness backdoor intervention to cut off the effects of confounders in the language modality and employ front-door intervention to eliminate the impact of confounders in the visual modality. We conducted extensive experiments on two challenging Out-of-Distribution (OOD) benchmarks, including VQA-VS and VQA-CE, which are designed to assess the robustness of VQA models under different shortcut biases. The experimental results show the effectiveness of our method. Specifically, our approach outperforms the current state-of-the-art debiasing methods on the IID metric and all nine OOD metrics of the VQA-VS dataset, and also surpasses the performance of the best-performing methods on all metrics of the VQA-CE dataset.

Previous studies have pointed out that visual question answering (VQA) models are prone to relying on language priors for answer predictions. In this context, predictions often depend on linguistic shortcuts rather than a comprehensive grasp of multimodal knowledge, which diminishes their generalization ability. In this paper, we propose a novel method, namely, KDAR, leveraging knowledge distillation to address the prior-dependency dilemmas within the VQA task. Specifically, the regularization effect facilitated by soft labels from a well-trained teacher is employed to penalize overfitting to the most common answers. The soft labels, which serve a regularization role, also provide semantic guidance that narrows the range of candidate answers. Additionally, we design an adaptive samplewise reweighting learning strategy to further mitigate bias by dynamically adjusting the importance of each sample. Experimental results demonstrate that our method enhances performance in both OOD and IID settings. Our method achieves state-of-the-art performance on the VQA-CPv2 out-of-distribution (OOD) benchmark, significantly outperforming previous state-of-the-art approaches.

Generalization beyond in-domain experience to out-of-distribution data is of paramount significance in the AI domain. Of late, state-of-the-art Visual Question Answering (VQA) models have shown impressive performance on in-domain data, partially due to the language prior bias which, however, hinders the generalization ability in practice. This paper attempts to provide new insights into the influence of language modality on VQA performance from an empirical study perspective. To achieve this, we conducted a series of experiments on six models. The results of these experiments revealed that, 1) apart from prior bias caused by question types, there is a notable influence of postfix-related bias in inducing biases, and 2) training VQA models with word-sequence-related variant questions demonstrated improved performance on the out-of-distribution benchmark, and the LXMERT even achieved a 10-point gain without adopting any debiasing methods. We delved into the underlying reasons behind these experimental results and put forward some simple proposals to reduce the models' dependency on language priors. The experimental results demonstrated the effectiveness of our proposed method in improving performance on the out-of-distribution benchmark, VQA-CPv2. We hope this study can inspire novel insights for future research on designing bias-reduction approaches.

Visual relationship understanding plays an indispensable role in grounded language tasks like visual question answering (VQA), which often requires precisely reasoning about relations among objects depicted in the given question. However, prior works generally suffer from the deficiencies as follows, (1)spatial-relation inference ambiguity, it is challenging to accurately estimate the distance of a pair of visual objects in 2D space if there is a visual-overlap between their 2D bounding-boxes, and (2) language-visual relational alignment missing, it is insufficient to generate a high-quality answer to the question if there is a lack of alignment in the language-visual relations of objects during fusion, even using a powerful fusion model like Transformer. To this end, we first model the spatial relation of a pair of objects in 3D space by augmenting the original 2D bounding-box with 1D depth information, and then propose a novel model namedDepth-awareSemanticGuided RelationalAttention Network (DSGANet), to explicitly exploit the formed 3D spatial relations of objects in an intra-/inter-modality manner for precise relational alignment. Extensive experiments conducted on the benchmarks (VQA v2.0 and GQA) demonstrate DSGANet achieves competitive performance compared to pretrained and non-pretrained models, such as 72.7% vs. 74.6% based on the learned grid features on VQA v2.0.

In recent years, the pre-training-then-fine-tuning paradigm has yielded immense success on a wide spectrum of cross-modal tasks, such as visual question answering (VQA), in which a visual-language (VL) model is first optimized via self-supervised task objectives, e.g., masked language modeling (MLM) and image-text matching (ITM), and then fine-tuned to adapt to downstream task (e.g., VQA) via a brand-new objective function, e.g., answer prediction. However, the inconsistency of the objective forms not only severely limits the generalization of pre-trained VL models to downstream tasks, but also requires a large amount of labeled data for fine-tuning. To alleviate the problem, we propose an innovative VL fine-tuning paradigm (named Declaration-based Prompt Tuning, abbreviated as DPT), which fine-tunes the model for downstream VQA using the pre-training objectives, boosting the effective adaptation of pre-trained models to the downstream task. Specifically, DPT reformulates the VQA task via (1) textual adaptation, which converts the given questions into declarative sentence form for prompt-tuning, and (2) task adaptation, which optimizes the objective function of VQA problem in the manner of pre-training phase. Experimental results on GQA dataset show that DPT outperforms the fine-tuned counterpart by a large margin regarding accuracy in both fully-supervised (2.68%) and zero-shot/fewshot (over 31%) settings. All the data and codes will be available to facilitate future research.

This paper presents a novel accelerated exact k-means called as "Ball k-means" by using the ball to describe each cluster, which focus on reducing the point-centroid distance computation. The "Ball k-means" can exactly find its neighbor clusters for each cluster, resulting distance computations only between a point and its neighbor clusters' centroids instead of all centroids. What's more, each cluster can be divided into "stable area" and "active area", and the latter one is further divided into some exact "annular area". The assignment of the points in the "stable area" is not changed while the points in each "annular area" will be adjusted within a few neighbor clusters. There are no upper or lower bounds in the whole process. Moreover, ball k-means uses ball clusters and neighbor searching along with multiple novel stratagems for reducing centroid distance computations. In comparison with the current state-of-the art accelerated exact bounded methods, the Yinyang algorithm and the Exponion algorithm, as well as other top-of-the-line tree-based and bounded methods, the ball k-means attains both higher performance and performs fewer distance calculations, especially for large-k problems. The faster speed, no extra parameters and simpler design of "Ball k-means" make it an all-around replacement of the naive k-means.