Order Matters: Exploring Order Sensitivity in Multimodal Large Language Models

Multimodal Large Language Models (MLLMs) utilize multimodal contexts consisting of text, images, or videos to solve various multimodal tasks. However, we find that changing the order of multimodal input can cause the model's performance to fluctuate between advanced performance and random guessing. This phenomenon exists in both single-modality (text-only or image-only) and mixed-modality (image-text-pair) contexts. Furthermore, we demonstrate that popular MLLMs pay special attention to certain multimodal context positions, particularly the beginning and end. Leveraging this special attention, we place key video frames and important image/text content in special positions within the context and submit them to the MLLM for inference. This method results in average performance gains of 14.7% for video-caption matching and 17.8% for visual question answering tasks. Additionally, we propose a new metric, Position-Invariant Accuracy (PIA), to address order bias in MLLM evaluation. Our research findings contribute to a better understanding of Multi-Modal In-Context Learning (MMICL) and provide practical strategies for enhancing MLLM performance without increasing computational costs.

Towards Defining an Efficient and Expandable File Format for AI-Generated Contents

Recently, AI-generated content (AIGC) has gained significant traction due to its powerful creation capability. However, the storage and transmission of large amounts of high-quality AIGC images inevitably pose new challenges for recent file formats. To overcome this, we define a new file format for AIGC images, named AIGIF, enabling ultra-low bitrate coding of AIGC images. Unlike compressing AIGC images intuitively with pixel-wise space as existing file formats, AIGIF instead compresses the generation syntax. This raises a crucial question: Which generation syntax elements, e.g., text prompt, device configuration, etc, are necessary for compression/transmission? To answer this question, we systematically investigate the effects of three essential factors: platform, generative model, and data configuration. We experimentally find that a well-designed composable bitstream structure incorporating the above three factors can achieve an impressive compression ratio of even up to 1/10,000 while still ensuring high fidelity. We also introduce an expandable syntax in AIGIF to support the extension of the most advanced generation models to be developed in the future.

Supportiveness-based Knowledge Rewriting for Retrieval-augmented Language Modeling

Retrieval-augmented language models (RALMs) have recently shown great potential in mitigating the limitations of implicit knowledge in LLMs, such as untimely updating of the latest expertise and unreliable retention of long-tail knowledge. However, since the external knowledge base, as well as the retriever, can not guarantee reliability, potentially leading to the knowledge retrieved not being helpful or even misleading for LLM generation. In this paper, we introduce Supportiveness-based Knowledge Rewriting (SKR), a robust and pluggable knowledge rewriter inherently optimized for LLM generation. Specifically, we introduce the novel concept of "supportiveness"--which represents how effectively a knowledge piece facilitates downstream tasks--by considering the perplexity impact of augmented knowledge on the response text of a white-box LLM. Based on knowledge supportiveness, we first design a training data curation strategy for our rewriter model, effectively identifying and filtering out poor or irrelevant rewrites (e.g., with low supportiveness scores) to improve data efficacy. We then introduce the direct preference optimization (DPO) algorithm to align the generated rewrites to optimal supportiveness, guiding the rewriter model to summarize augmented content that better improves the final response. Comprehensive evaluations across six popular knowledge-intensive tasks and four LLMs have demonstrated the effectiveness and superiority of SKR. With only 7B parameters, SKR has shown better knowledge rewriting capability over GPT-4, the current state-of-the-art general-purpose LLM.

Joint Event Extraction via Structural Semantic Matching

Event Extraction (EE) is one of the essential tasks in information extraction, which aims to detect event mentions from text and find the corresponding argument roles. The EE task can be abstracted as a process of matching the semantic definitions and argument structures of event types with the target text. This paper encodes the semantic features of event types and makes structural matching with target text. Specifically, Semantic Type Embedding (STE) and Dynamic Structure Encoder (DSE) modules are proposed. Also, the Joint Structural Semantic Matching (JSSM) model is built to jointly perform event detection and argument extraction tasks through a bidirectional attention layer. The experimental results on the ACE2005 dataset indicate that our model achieves a significant performance improvement

NVTC: Nonlinear Vector Transform Coding

In theory, vector quantization (VQ) is always better than scalar quantization (SQ) in terms of rate-distortion (R-D) performance. Recent state-of-the-art methods for neural image compression are mainly based on nonlinear transform coding (NTC) with uniform scalar quantization, overlooking the benefits of VQ due to its exponentially increased complexity. In this paper, we first investigate on some toy sources, demonstrating that even if modern neural networks considerably enhance the compression performance of SQ with nonlinear transform, there is still an insurmountable chasm between SQ and VQ. Therefore, revolving around VQ, we propose a novel framework for neural image compression named Nonlinear Vector Transform Coding (NVTC). NVTC solves the critical complexity issue of VQ through (1) a multi-stage quantization strategy and (2) nonlinear vector transforms. In addition, we apply entropy-constrained VQ in latent space to adaptively determine the quantization boundaries for joint rate-distortion optimization, which improves the performance both theoretically and experimentally. Compared to previous NTC approaches, NVTC demonstrates superior rate-distortion performance, faster decoding speed, and smaller model size. Our code is available at https://github.com/USTC-IMCL/NVTC

Exploiting Hybrid Semantics of Relation Paths for Multi-hop Question Answering Over Knowledge Graphs

Answering natural language questions on knowledge graphs (KGQA) remains a great challenge in terms of understanding complex questions via multi-hop reasoning. Previous efforts usually exploit large-scale entity-related text corpora or knowledge graph (KG) embeddings as auxiliary information to facilitate answer selection. However, the rich semantics implied in off-the-shelf relation paths between entities is far from well explored. This paper proposes improving multi-hop KGQA by exploiting relation paths' hybrid semantics. Specifically, we integrate explicit textual information and implicit KG structural features of relation paths based on a novel rotate-and-scale entity link prediction framework. Extensive experiments on three existing KGQA datasets demonstrate the superiority of our method, especially in multi-hop scenarios. Further investigation confirms our method's systematical coordination between questions and relation paths to identify answer entities.

Eliciting Knowledge from Pretrained Language Models for Prototypical Prompt Verbalizer

Recent advances on prompt-tuning cast few-shot classification tasks as a masked language modeling problem. By wrapping input into a template and using a verbalizer which constructs a mapping between label space and label word space, prompt-tuning can achieve excellent results in zero-shot and few-shot scenarios. However, typical prompt-tuning needs a manually designed verbalizer which requires domain expertise and human efforts. And the insufficient label space may introduce considerable bias into the results. In this paper, we focus on eliciting knowledge from pretrained language models and propose a prototypical prompt verbalizer for prompt-tuning. Labels are represented by prototypical embeddings in the feature space rather than by discrete words. The distances between the embedding at the masked position of input and prototypical embeddings are used as classification criterion. For zero-shot settings, knowledge is elicited from pretrained language models by a manually designed template to form initial prototypical embeddings. For few-shot settings, models are tuned to learn meaningful and interpretable prototypical embeddings. Our method optimizes models by contrastive learning. Extensive experimental results on several many-class text classification datasets with low-resource settings demonstrate the effectiveness of our approach compared with other verbalizer construction methods. Our implementation is available at https://github.com/Ydongd/prototypical-prompt-verbalizer.

Blind Quality Assessment for Image Superresolution Using Deep Two-Stream Convolutional Networks

Numerous image superresolution (SR) algorithms have been proposed for reconstructing high-resolution (HR) images from input images with lower spatial resolutions. However, effectively evaluating the perceptual quality of SR images remains a challenging research problem. In this paper, we propose a no-reference/blind deep neural network-based SR image quality assessor (DeepSRQ). To learn more discriminative feature representations of various distorted SR images, the proposed DeepSRQ is a two-stream convolutional network including two subcomponents for distorted structure and texture SR images. Different from traditional image distortions, the artifacts of SR images cause both image structure and texture quality degradation. Therefore, we choose the two-stream scheme that captures different properties of SR inputs instead of directly learning features from one image stream. Considering the human visual system (HVS) characteristics, the structure stream focuses on extracting features in structural degradations, while the texture stream focuses on the change in textural distributions. In addition, to augment the training data and ensure the category balance, we propose a stride-based adaptive cropping approach for further improvement. Experimental results on three publicly available SR image quality databases demonstrate the effectiveness and generalization ability of our proposed DeepSRQ method compared with state-of-the-art image quality assessment algorithms.

Prior-enlightened and Motion-robust Video Deblurring

Various blur distortions in video will cause negative impact on both human viewing and video-based applications, which makes motion-robust deblurring methods urgently needed. Most existing works have strong dataset dependency and limited generalization ability in handling challenging scenarios, like blur in low contrast or severe motion areas, and non-uniform blur. Therefore, we propose a PRiOr-enlightened and MOTION-robust video deblurring model (PROMOTION) suitable for challenging blurs. On the one hand, we use 3D group convolution to efficiently encode heterogeneous prior information, explicitly enhancing the scenes' perception while mitigating the output's artifacts. On the other hand, we design the priors representing blur distribution, to better handle non-uniform blur in spatio-temporal domain. Besides the classical camera shake caused global blurry, we also prove the generalization for the downstream task suffering from local blur. Extensive experiments demonstrate we can achieve the state-of-the-art performance on well-known REDS and GoPro datasets, and bring machine task gain.

Region Normalization for Image Inpainting

Feature Normalization (FN) is an important technique to help neural network training, which typically normalizes features across spatial dimensions. Most previous image inpainting methods apply FN in their networks without considering the impact of the corrupted regions of the input image on normalization, e.g. mean and variance shifts. In this work, we show that the mean and variance shifts caused by full-spatial FN limit the image inpainting network training and we propose a spatial region-wise normalization named Region Normalization (RN) to overcome the limitation. RN divides spatial pixels into different regions according to the input mask, and computes the mean and variance in each region for normalization. We develop two kinds of RN for our image inpainting network: (1) Basic RN (RN-B), which normalizes pixels from the corrupted and uncorrupted regions separately based on the original inpainting mask to solve the mean and variance shift problem; (2) Learnable RN (RN-L), which automatically detects potentially corrupted and uncorrupted regions for separate normalization, and performs global affine transformation to enhance their fusion. We apply RN-B in the early layers and RN-L in the latter layers of the network respectively. Experiments show that our method outperforms current state-of-the-art methods quantitatively and qualitatively. We further generalize RN to other inpainting networks and achieve consistent performance improvements.