MMIE: Massive Multimodal Interleaved Comprehension Benchmark for Large Vision-Language Models

Interleaved multimodal comprehension and generation, enabling models to produce and interpret both images and text in arbitrary sequences, have become a pivotal area in multimodal learning. Despite significant advancements, the evaluation of this capability remains insufficient. Existing benchmarks suffer from limitations in data scale, scope, and evaluation depth, while current evaluation metrics are often costly or biased, lacking in reliability for practical applications. To address these challenges, we introduce MMIE, a large-scale knowledge-intensive benchmark for evaluating interleaved multimodal comprehension and generation in Large Vision-Language Models (LVLMs). MMIE comprises 20K meticulously curated multimodal queries, spanning 3 categories, 12 fields, and 102 subfields, including mathematics, coding, physics, literature, health, and arts. It supports both interleaved inputs and outputs, offering a mix of multiple-choice and open-ended question formats to evaluate diverse competencies. Moreover, we propose a reliable automated evaluation metric, leveraging a scoring model fine-tuned with human-annotated data and systematic evaluation criteria, aimed at reducing bias and improving evaluation accuracy. Extensive experiments demonstrate the effectiveness of our benchmark and metrics in providing a comprehensive evaluation of interleaved LVLMs. Specifically, we evaluate eight LVLMs, revealing that even the best models show significant room for improvement, with most achieving only moderate results. We believe MMIE will drive further advancements in the development of interleaved LVLMs. We publicly release our benchmark and code in https://mmie-bench.github.io/.

VHELM: A Holistic Evaluation of Vision Language Models

Current benchmarks for assessing vision-language models (VLMs) often focus on their perception or problem-solving capabilities and neglect other critical aspects such as fairness, multilinguality, or toxicity. Furthermore, they differ in their evaluation procedures and the scope of the evaluation, making it difficult to compare models. To address these issues, we extend the HELM framework to VLMs to present the Holistic Evaluation of Vision Language Models (VHELM). VHELM aggregates various datasets to cover one or more of the 9 aspects: visual perception, knowledge, reasoning, bias, fairness, multilinguality, robustness, toxicity, and safety. In doing so, we produce a comprehensive, multi-dimensional view of the capabilities of the VLMs across these important factors. In addition, we standardize the standard inference parameters, methods of prompting, and evaluation metrics to enable fair comparisons across models. Our framework is designed to be lightweight and automatic so that evaluation runs are cheap and fast. Our initial run evaluates 22 VLMs on 21 existing datasets to provide a holistic snapshot of the models. We uncover new key findings, such as the fact that efficiency-focused models (e.g., Claude 3 Haiku or Gemini 1.5 Flash) perform significantly worse than their full models (e.g., Claude 3 Opus or Gemini 1.5 Pro) on the bias benchmark but not when evaluated on the other aspects. For transparency, we release the raw model generations and complete results on our website (https://crfm.stanford.edu/helm/vhelm/v2.0.1). VHELM is intended to be a living benchmark, and we hope to continue adding new datasets and models over time.

CARES: A Comprehensive Benchmark of Trustworthiness in Medical Vision Language Models

Artificial intelligence has significantly impacted medical applications, particularly with the advent of Medical Large Vision Language Models (Med-LVLMs), sparking optimism for the future of automated and personalized healthcare. However, the trustworthiness of Med-LVLMs remains unverified, posing significant risks for future model deployment. In this paper, we introduce CARES and aim to comprehensively evaluate the Trustworthiness of Med-LVLMs across the medical domain. We assess the trustworthiness of Med-LVLMs across five dimensions, including trustfulness, fairness, safety, privacy, and robustness. CARES comprises about 41K question-answer pairs in both closed and open-ended formats, covering 16 medical image modalities and 27 anatomical regions. Our analysis reveals that the models consistently exhibit concerns regarding trustworthiness, often displaying factual inaccuracies and failing to maintain fairness across different demographic groups. Furthermore, they are vulnerable to attacks and demonstrate a lack of privacy awareness. We publicly release our benchmark and code in https://github.com/richard-peng-xia/CARES.

Multimodal Clinical Trial Outcome Prediction with Large Language Models

The clinical trial is a pivotal and costly process, often spanning multiple years and requiring substantial financial resources. Therefore, the development of clinical trial outcome prediction models aims to exclude drugs likely to fail and holds the potential for significant cost savings. Recent data-driven attempts leverage deep learning methods to integrate multimodal data for predicting clinical trial outcomes. However, these approaches rely on manually designed modal-specific encoders, which limits both the extensibility to adapt new modalities and the ability to discern similar information patterns across different modalities. To address these issues, we propose a multimodal mixture-of-experts (LIFTED) approach for clinical trial outcome prediction. Specifically, LIFTED unifies different modality data by transforming them into natural language descriptions. Then, LIFTED constructs unified noise-resilient encoders to extract information from modal-specific language descriptions. Subsequently, a sparse Mixture-of-Experts framework is employed to further refine the representations, enabling LIFTED to identify similar information patterns across different modalities and extract more consistent representations from those patterns using the same expert model. Finally, a mixture-of-experts module is further employed to dynamically integrate different modality representations for prediction, which gives LIFTED the ability to automatically weigh different modalities and pay more attention to critical information. The experiments demonstrate that LIFTED significantly enhances performance in predicting clinical trial outcomes across all three phases compared to the best baseline, showcasing the effectiveness of our proposed key components.

STAN: Stage-Adaptive Network for Multi-Task Recommendation by Learning User Lifecycle-Based Representation

Recommendation systems play a vital role in many online platforms, with their primary objective being to satisfy and retain users. As directly optimizing user retention is challenging, multiple evaluation metrics are often employed. Existing methods generally formulate the optimization of these evaluation metrics as a multitask learning problem, but often overlook the fact that user preferences for different tasks are personalized and change over time. Identifying and tracking the evolution of user preferences can lead to better user retention. To address this issue, we introduce the concept of "user lifecycle", consisting of multiple stages characterized by users' varying preferences for different tasks. We propose a novel Stage-Adaptive Network (STAN) framework for modeling user lifecycle stages. STAN first identifies latent user lifecycle stages based on learned user preferences, and then employs the stage representation to enhance multi-task learning performance. Our experimental results using both public and industrial datasets demonstrate that the proposed model significantly improves multi-task prediction performance compared to state-of-the-art methods, highlighting the importance of considering user lifecycle stages in recommendation systems. Furthermore, online A/B testing reveals that our model outperforms the existing model, achieving a significant improvement of 3.05% in staytime per user and 0.88% in CVR. These results indicate that our approach effectively improves the overall efficiency of the multi-task recommendation system.

Knowledge Soft Integration for Multimodal Recommendation

One of the main challenges in modern recommendation systems is how to effectively utilize multimodal content to achieve more personalized recommendations. Despite various proposed solutions, most of them overlook the mismatch between the knowledge gained from independent feature extraction processes and downstream recommendation tasks. Specifically, multimodal feature extraction processes do not incorporate prior knowledge relevant to recommendation tasks, while recommendation tasks often directly use these multimodal features as side information. This mismatch can lead to model fitting biases and performance degradation, which this paper refers to as the \textit{curse of knowledge} problem. To address this issue, we propose using knowledge soft integration to balance the utilization of multimodal features and the curse of knowledge problem it brings about. To achieve this, we put forward a Knowledge Soft Integration framework for the multimodal recommendation, abbreviated as KSI, which is composed of the Structure Efficiently Injection (SEI) module and the Semantic Soft Integration (SSI) module. In the SEI module, we model the modality correlation between items using Refined Graph Neural Network (RGNN), and introduce a regularization term to reduce the redundancy of user/item representations. In the SSI module, we design a self-supervised retrieval task to further indirectly integrate the semantic knowledge of multimodal features, and enhance the semantic discrimination of item representations. Extensive experiments on three benchmark datasets demonstrate the superiority of KSI and validate the effectiveness of its two modules.

Robust Image Ordinal Regression with Controllable Image Generation

Image ordinal regression has been mainly studied along the line of exploiting the order of categories. However, the issues of class imbalance and category overlap that are very common in ordinal regression were largely overlooked. As a result, the performance on minority categories is often unsatisfactory. In this paper, we propose a novel framework called CIG based on controllable image generation to directly tackle these two issues. Our main idea is to generate extra training samples with specific labels near category boundaries, and the sample generation is biased toward the less-represented categories. To achieve controllable image generation, we seek to separate structural and categorical information of images based on structural similarity, categorical similarity, and reconstruction constraints. We evaluate the effectiveness of our new CIG approach in three different image ordinal regression scenarios. The results demonstrate that CIG can be flexibly integrated with off-the-shelf image encoders or ordinal regression models to achieve improvement, and further, the improvement is more significant for minority categories.

Click-aware Structure Transfer with Sample Weight Assignment for Post-Click Conversion Rate Estimation

Post-click Conversion Rate (CVR) prediction task plays an essential role in industrial applications, such as recommendation and advertising. Conventional CVR methods typically suffer from the data sparsity problem as they rely only on samples where the user has clicked. To address this problem, researchers have introduced the method of multi-task learning, which utilizes non-clicked samples and shares feature representations of the Click-Through Rate (CTR) task with the CVR task. However, it should be noted that the CVR and CTR tasks are fundamentally different and may even be contradictory. Therefore, introducing a large amount of CTR information without distinction may drown out valuable information related to CVR. This phenomenon is called the curse of knowledge problem in this paper. To tackle this issue, we argue that a trade-off should be achieved between the introduction of large amounts of auxiliary information and the protection of valuable information related to CVR. Hence, we propose a Click-aware Structure Transfer model with sample Weight Assignment, abbreviated as CSTWA. It pays more attention to the latent structure information, which can filter the input information that is related to CVR, instead of directly sharing feature representations. Meanwhile, to capture the representation conflict between CTR and CVR, we calibrate the representation layer and reweight the discriminant layer to excavate the click bias information from the CTR tower. Moreover, it incorporates a sample weight assignment algorithm biased towards CVR modeling, to make the knowledge from CTR would not mislead the CVR. Extensive experiments on industrial and public datasets have demonstrated that CSTWA significantly outperforms widely used and competitive models.

CTT-Net: A Multi-view Cross-token Transformer for Cataract Postoperative Visual Acuity Prediction

Surgery is the only viable treatment for cataract patients with visual acuity (VA) impairment. Clinically, to assess the necessity of cataract surgery, accurately predicting postoperative VA before surgery by analyzing multi-view optical coherence tomography (OCT) images is crucially needed. Unfortunately, due to complicated fundus conditions, determining postoperative VA remains difficult for medical experts. Deep learning methods for this problem were developed in recent years. Although effective, these methods still face several issues, such as not efficiently exploring potential relations between multi-view OCT images, neglecting the key role of clinical prior knowledge (e.g., preoperative VA value), and using only regression-based metrics which are lacking reference. In this paper, we propose a novel Cross-token Transformer Network (CTT-Net) for postoperative VA prediction by analyzing both the multi-view OCT images and preoperative VA. To effectively fuse multi-view features of OCT images, we develop cross-token attention that could restrict redundant/unnecessary attention flow. Further, we utilize the preoperative VA value to provide more information for postoperative VA prediction and facilitate fusion between views. Moreover, we design an auxiliary classification loss to improve model performance and assess VA recovery more sufficiently, avoiding the limitation by only using the regression metrics. To evaluate CTT-Net, we build a multi-view OCT image dataset collected from our collaborative hospital. A set of extensive experiments validate the effectiveness of our model compared to existing methods in various metrics. Code is available at: https://github.com/wjh892521292/Cataract OCT.

PNM: Pixel Null Model for General Image Segmentation

A major challenge in image segmentation is classifying object boundaries. Recent efforts propose to refine the segmentation result with boundary masks. However, models are still prone to misclassifying boundary pixels even when they correctly capture the object contours. In such cases, even a perfect boundary map is unhelpful for segmentation refinement. In this paper, we argue that assigning proper prior weights to error-prone pixels such as object boundaries can significantly improve the segmentation quality. Specifically, we present the \textit{pixel null model} (PNM), a prior model that weights each pixel according to its probability of being correctly classified by a random segmenter. Empirical analysis shows that PNM captures the misclassification distribution of different state-of-the-art (SOTA) segmenters. Extensive experiments on semantic, instance, and panoptic segmentation tasks over three datasets (Cityscapes, ADE20K, MS COCO) confirm that PNM consistently improves the segmentation quality of most SOTA methods (including the vision transformers) and outperforms boundary-based methods by a large margin. We also observe that the widely-used mean IoU (mIoU) metric is insensitive to boundaries of different sharpness. As a byproduct, we propose a new metric, \textit{PNM IoU}, which perceives the boundary sharpness and better reflects the model segmentation performance in error-prone regions.