Multimodal Large Language Models (MLLMs) have been shown to be vulnerable to malicious queries that can elicit unsafe responses. Recent works use prompt engineering, response classification, or finetuning to improve MLLM safety. Nevertheless, such approaches are often ineffective against evolving malicious patterns, may require rerunning the query, or demand heavy computational resources. Steering the intermediate activations of a frozen model at inference time has recently emerged as a flexible and effective solution. However, existing steering methods for MLLMs typically handle only a narrow set of safety-related concepts or struggle to adjust specific concepts without affecting others. To address these challenges, we introduce Dictionary-Aligned Concept Control (DACO), a framework that jointly utilizes a curated concept dictionary with a Sparse Autoencoder (SAE) to provide granular control over MLLM activations. First, we curate a dictionary of 15,000 multimodal concepts by retrieving over 400,000 caption-image stimuli (we name the dataset DACO-400K) and summarizing their activations into per-concept directions. Second, we show that the curated dictionary can be directly applied to interventions via sparse coding. Third, we propose a new steering approach that uses our dictionary to initialize the training of an SAE and automatically annotate the semantics of the SAE atoms for safeguarding MLLMs. Experiments on multiple MLLMs (e.g., QwenVL, LLaVA, InternVL) across safety benchmarks (e.g., MM-SafetyBench, JailBreakV) show that DACO significantly improves the MLLM safety while maintaining general-purpose capabilities.

Composed Image Retrieval (CIR) is a complex task that aims to retrieve images based on a multimodal query. Typical training data consists of triplets containing a reference image, a textual description of desired modifications, and the target image, which are expensive and time-consuming to acquire. The scarcity of CIR datasets has led to zero-shot approaches utilizing synthetic triplets or leveraging vision-language models (VLMs) with ubiquitous web-crawled image-caption pairs. However, these methods have significant limitations: synthetic triplets suffer from limited scale, lack of diversity, and unnatural modification text, while image-caption pairs hinder joint embedding learning of the multimodal query due to the absence of triplet data. Moreover, existing approaches struggle with complex and nuanced modification texts that demand sophisticated fusion and understanding of vision and language modalities. We present CoLLM, a one-stop framework that effectively addresses these limitations. Our approach generates triplets on-the-fly from image-caption pairs, enabling supervised training without manual annotation. We leverage Large Language Models (LLMs) to generate joint embeddings of reference images and modification texts, facilitating deeper multimodal fusion. Additionally, we introduce Multi-Text CIR (MTCIR), a large-scale dataset comprising 3.4M samples, and refine existing CIR benchmarks (CIRR and Fashion-IQ) to enhance evaluation reliability. Experimental results demonstrate that CoLLM achieves state-of-the-art performance across multiple CIR benchmarks and settings. MTCIR yields competitive results, with up to 15% performance improvement. Our refined benchmarks provide more reliable evaluation metrics for CIR models, contributing to the advancement of this important field. Project page is at collm-cvpr25.github.io.

Image to image matching has been well studied in the computer vision community. Previous studies mainly focus on training a deep metric learning model matching visual patterns between the query image and gallery images. In this study, we show that pure image-to- image matching suffers from false positives caused by matching to local visual patterns. To alleviate this issue, we propose to leverage recent advances in vision-language pretraining research. Specifically, we introduce additional image-text alignment losses into deep metric learning, which serve as constraints to the image-to-image matching loss. With additional alignments between the text (e.g., product title) and image pairs, the model can learn concepts from both modalities explicitly, which avoids matching low-level visual features. We progressively develop two variants, a 3-tower and a 4-tower model, where the latter takes one more short text query input. Through extensive experiments, we show that this change leads to a substantial improvement to the image to image matching problem. We further leveraged this model for multimodal search, which takes both image and reformulation text queries to improve search quality. Both offline and online experiments show strong improvements on the main metrics. Specifically, we see 4.95% relative improvement on image matching click through rate with the 3-tower model and 1.13% further improvement from the 4-tower model.

Acquiring meaningful representations of gene expression is essential for the accurate prediction of downstream regulatory tasks, such as identifying promoters and transcription factor binding sites. However, the current dependency on supervised learning, constrained by the limited availability of labeled genomic data, impedes the ability to develop robust predictive models with broad generalization capabilities. In response, recent advancements have pivoted towards the application of self-supervised training for DNA sequence modeling, enabling the adaptation of pre-trained genomic representations to a variety of downstream tasks. Departing from the straightforward application of masked language learning techniques to DNA sequences, approaches such as MoDNA enrich genome language modeling with prior biological knowledge. In this study, we advance DNA language models by utilizing the Motif-oriented DNA (MoDNA) pre-training framework, which is established for self-supervised learning at the pre-training stage and is flexible enough for application across different downstream tasks. MoDNA distinguishes itself by efficiently learning semantic-level genomic representations from an extensive corpus of unlabeled genome data, offering a significant improvement in computational efficiency over previous approaches. The framework is pre-trained on a comprehensive human genome dataset and fine-tuned for targeted downstream tasks. Our enhanced analysis and evaluation in promoter prediction and transcription factor binding site prediction have further validated MoDNA’s exceptional capabilities, emphasizing its contribution to advancements in genomic predictive modeling.

Recent advancements in Multimodal Large Language Models (MLLMs) have revolutionized the field of vision-language understanding by integrating visual perception capabilities into Large Language Models (LLMs). The prevailing trend in this field involves the utilization of a vision encoder derived from vision-language contrastive learning (CL), showing expertise in capturing overall representations while facing difficulties in capturing detailed local patterns. In this work, we focus on enhancing the visual representations for MLLMs by combining high-frequency and detailed visual representations, obtained through masked image modeling (MIM), with semantically-enriched low-frequency representations captured by CL. To achieve this goal, we introduce X-Former which is a lightweight transformer module designed to exploit the complementary strengths of CL and MIM through an innovative interaction mechanism. Specifically, X-Former first bootstraps vision-language representation learning and multimodal-to-multimodal generative learning from two frozen vision encoders, i.e., CLIP-ViT (CL-based) and MAE-ViT (MIM-based). It further bootstraps vision-to-language generative learning from a frozen LLM to ensure visual features from X-Former can be interpreted by the LLM. To demonstrate the effectiveness of our approach, we assess its performance on tasks demanding detailed visual understanding. Extensive evaluations indicate that X-Former excels in visual reasoning tasks involving both structural and semantic categories in the GQA dataset. Assessment on fine-grained visual perception benchmark further confirms its superior capabilities in visual understanding.

Bile acids (BAs) facilitate the absorption of dietary lipids and vitamins and have also been identified as signaling molecules involved in regulating their own metabolism, glucose and lipid metabolism, as well as immunity. Disturbances in BA homeostasis are associated with various enterohepatic and metabolic diseases, such as cholestasis, nonalcoholic steatohepatitis, inflammatory bowel disease, and obesity. As a key regulator, the nuclear orphan receptor farnesoid X receptor (FXR, NR1H4) precisely regulates BA homeostasis by transcriptional regulation of genes involved in BA synthesis, metabolism, and enterohepatic circulation. FXR is widely regarded as the most potential therapeutic target. Obeticholic acid is the only FXR agonist approved to treat patients with primary biliary cholangitis, but its non-specific activation of systemic FXR also causes high-frequency side effects. In recent years, developing tissue-specific FXR-targeting drugs has become a research highlight. This article provides a comprehensive overview of the role of tissue-specific intestine/liver FXR in regulating genes involved in BA homeostasis and briefly discusses tissue-specific FXR as a therapeutic target for treating diseases. These findings provide the basis for the development of tissue-specific FXR modulators for the treatment of enterohepatic and metabolic diseases associated with BA dysfunction.

Identifying similar products in e-commerce is useful in discovering relationships between products, making recommendations, and increasing diversity in search results. Product representation learning is the first step to define a generalized product similarity metric for search. The second step is to extend similarity search to a large scale (e.g., e-commerce catalog scale) without sacrificing quality. In this work, we present a solution that interweaves both steps, i.e., learn representations suited to high quality retrieval using contrastive learning (CL) and retrieve similar items from a large search space using approximate nearest neighbor search (ANNS) to trade-off quality for speed. We propose a CL training strategy for learning uni-modal encoders suited to multi-modal similarity search for e-commerce. We study ANNS retrieval by generating Pareto Frontiers (PFs) without requiring labels. Our CL training strategy doubles retrieval@1 metric across categories (e.g., from 36% to 88% in category C). We also demonstrate that ANNS engine optimization using PFs help select configurations appropriately (e.g., we achieve 6.8× search speed with just 2% drop from the maximum retrieval accuracy in medium size datasets).

Electrocardiogram (ECG)-gated multi-phase computed tomography angiography (MP-CTA) is frequently used for diagnosis of coronary artery disease. Radiation dose may become a potential concern as the scan needs to cover a wide range of cardiac phases during a heart cycle. A common method to reduce radiation is to limit the full-dose acquisition to a predefined range of phases while reducing the radiation dose for the rest. Our goal in this study is to develop a spatiotemporal deep learning method to enhance the quality of low-dose CTA images at phases acquired at reduced radiation dose. Recently, we demonstrated that a deep learning method, Cycle-Consistent generative adversarial networks (CycleGAN), could effectively denoise low-dose CT images through spatial image translation without labeled image pairs in both low-dose and full-dose image domains. As CycleGAN does not utilize the temporal information in its denoising mechanism, we propose to use RecycleGAN, which could translate a series of images ordered in time from the low-dose domain to the full-dose domain through an additional recurrent network. To evaluate RecycleGAN, we use the XCAT phantom program, a highly realistic simulation tool based on real patient data, to generate MP-CTA image sequences for 18 patients (14 for training, 2 for validation and 2 for test). Our simulation results show that RecycleGAN can achieve better denoising performance than CycleGAN based on both visual inspection and quantitative metrics. We further demonstrate the superior denoising performance of RecycleGAN using clinical MP-CTA images from 50 patients.

Unsupervised domain adaptation (UDA) aims to transfer the knowledge learnt from a labeled source domain to an unlabeled target domain. Previous work is mainly built upon convolutional neural networks (CNNs) to learn domain-invariant representations. With the recent exponential increase in applying Vision Transformer (ViT) to vision tasks, the capability of ViT in adapting cross-domain knowledge, however, remains unexplored in the literature. To fill this gap, this paper first comprehensively investigates the performance of ViT on a variety of domain adaptation tasks. Surprisingly, ViT demonstrates superior generalization ability, while the performance can be further improved by incorporating adversarial adaptation. Notwithstanding, directly using CNNs-based adaptation strategies fails to take the advantage of ViT’s intrinsic merits (e.g., attention mechanism and sequential image representation) which play an important role in knowledge transfer. To remedy this, we propose an unified framework, namely Transferable Vision Transformer (TVT), to fully exploit the transferability of ViT for domain adaptation. Specifically, we delicately devise a novel and effective unit, which we term Transferability Adaption Module (TAM). By injecting learned trans- ferabilities into attention blocks, TAM compels ViT focus on both transferable and discriminative features. Besides, we leverage discriminative clustering to enhance feature diversity and separation which are undermined during adversarial domain alignment. To verify its versatility, we perform extensive studies of TVT on four benchmarks and the experimental results demonstrate that TVT attains significant improvements compared to existing state-of-the-art UDA methods.

Obtaining informative representations of gene expression is crucial in predicting various downstream regulatory-related tasks such as promoter prediction and transcription factor binding sites prediction. Nevertheless, current supervised learning with insufficient labeled genomes limits the generalization capability of training a robust predictive model. Recently researchers model DNA sequences by self-supervised training and transfer the pre-trained genome representations to various downstream tasks. Instead of directly shifting the mask language learning to DNA sequence learning, we incorporate prior knowledge into genome language modeling representations. We propose a novel Motif-oriented DNA (MoDNA) pre-training framework, which is designed self-supervised and can be fine-tuned for different downstream tasks MoDNA effectively learns the semantic level genome representations from enormous unlabelled genome data, and is more computationally efficient than previous methods. We pre-train MoDNA on human genome data and fine-tune it on downstream tasks. Extensive experimental results on promoter prediction and transcription factor binding sites prediction demonstrate the state-of-the-art performance of MoDNA.

Molecule generation is the procedure to generate initial novel molecule proposals for molecule design. Molecules are first projected into continuous vectors in chemical latent space, and then, these embedding vectors are decoded into molecules under the variational autoencoder (VAE) framework. The continuous latent space of VAE can be utilized to generate novel molecules with desired chemical properties and further optimize the desired chemical properties of molecules. However, there is a posterior collapse problem with the conventional recurrent neural network-based VAEs for the molecule sequence generation, which deteriorates the generation performance. We investigate the posterior collapse problem and find that the underestimated reconstruction loss is the main factor in the posterior collapse problem in molecule sequence generation. To support our conclusion, we present both analytical and experimental evidence. What is more, we propose an efficient and effective solution to fix the problem and prevent posterior collapse. As a result, our method achieves competitive reconstruction accuracy and validity score on the benchmark data sets.

Aligning signals from different modalities is an important step in vision-language representation learning as it affects the performance of later stages such as cross-modality fusion. Since image and text typically reside in different regions of the feature space, directly aligning them at instance level is challenging especially when features are still evolving during training. In this paper, we propose to align at a higher and more stable level using cluster representation. Specifically, we treat image and text as two "views" of the same entity, and encode them into a joint vision-language coding space spanned by a dictionary of cluster centers (codebook). We contrast positive and negative samples via their cluster assignments while simultaneously optimizing the cluster centers. To further smooth out the learning process, we adopt a teacher-student distillation paradigm, where the momentum teacher of one view guides the student learning of the other. We evaluated our approach on common vision language benchmarks and obtain new SoTA on zero-shot cross modality retrieval while being competitive on various other transfer tasks.

Vision-language representation learning largely benefits from image-text alignment through contrastive losses (e.g., InfoNCE loss). The success of this alignment strategy is attributed to its capability in maximizing the mutual information (MI) between an image and its matched text. However, simply performing cross-modal alignment (CMA) ignores data potential within each modality, which may result in degraded representations. For instance, although CMA-based models are able to map image-text pairs close together in the embedding space, they fail to ensure that similar inputs from the same modality stay close by. This problem can get even worse when the pre-training data is noisy. In this paper, we propose triple contrastive learning (TCL) for vision-language pre-training by leveraging both cross-modal and intra-modal self-supervision. Besides CMA, TCL introduces an intra-modal contrastive objective to provide complementary benefits in representation learning. To take advantage of localized and structural information from image and text input, TCL further maximizes the average MI between local regions of image/text and their global summary. To the best of our knowledge, ours is the first work that takes into account local structure information for multi-modality representation learning. Experimental evaluations show that our approach is competitive and achieves the new state of the art on various common down-stream vision-language tasks such as image-text retrieval and visual question answering.

Recent studies imply that deep neural networks are vulnerable to adversarial examples, i.e., inputs with a slight but intentional perturbation are incorrectly classified by the network. Such vulnerability makes it risky for some security-related applications (e.g., semantic segmentation in autonomous cars) and triggers tremendous concerns on the model reliability. For the first time, we comprehensively evaluate the robustness of existing UDA methods and propose a robust UDA approach. It is rooted in two observations: i) the robustness of UDA methods in semantic segmentation remains unexplored, which poses a security concern in this field; and ii) although commonly used self-supervision (e.g., rotation and jigsaw) benefits model robustness in classification and recognition tasks, they fail to provide the critical supervision signals that are essential in semantic segmentation. These observations motivate us to propose adversarial self-supervision UDA (or ASSUDA) that maximizes the agreement between clean images and their adversarial examples by a contrastive loss in the output space. Extensive empirical studies on commonly used benchmarks demonstrate that ASSUDA is resistant to adversarial attacks.

In this paper, we consider the problem of unsupervised domain adaptation in the semantic segmentation. There are two primary issues in this field, i.e., what and how to transfer domain knowledge across two domains. Existing methods mainly focus on adapting domain-invariant features (what to transfer) through adversarial learning (how to transfer). Context dependency is essential for semantic segmentation, however, its transferability is still not well understood. Furthermore, how to transfer contextual information across two domains remains unexplored. Motivated by this, we propose a cross-attention mechanism based on self-attention to capture context dependencies between two domains and adapt transferable context. To achieve this goal, we design two cross-domain attention modules to adapt context dependencies from both spatial and channel views. Specifically, the spatial attention module captures local feature dependencies between each position in the source and target image. The channel attention module models semantic dependencies between each pair of cross-domain channel maps. To adapt context dependencies, we further selectively aggregate the context information from two domains. The superiority of our method over existing state-of-the-art methods is empirically proved on "GTA5 to Cityscapes" and "SYNTHIA to Cityscapes".

Graph Neural Networks (GNNs) draw their strength from explicitly modeling the topological information of structured data. However, existing GNNs suffer from limited capability in capturing the hierarchical graph representation which plays an important role in graph classification. In this paper, we innovatively propose hierarchical graph capsule network (HGCN) that can jointly learn node embeddings and extract graph hierarchies. Specifically, disentangled graph capsules are established by identifying heterogeneous factors underlying each node, such that their instantiation parameters represent different properties of the same entity. To learn the hierarchical representation, HGCN characterizes the part-whole relationship between lower-level capsules (part) and higher-level capsules (whole) by explicitly considering the structure information among the parts. Experimental studies demonstrate the effectiveness of HGCN and the contribution of each component. Code: https://github.com/uta-smile/HGCN

Retrosynthesis is the process of recursively decomposing target molecules into available building blocks. It plays an important role in solving problems in organic synthesis planning. To automate the retrosynthesis analysis, many retrosynthesis prediction methods have been proposed. However, most of them are cumbersome and lack interpretability about their predictions. In this paper, we devise a novel template-free algorithm, RetroXpert, for automatic retrosynthetic expansion by automating the procedure that chemists used to do. Our method disassembles retrosynthesis into two steps: i) we identify the potential reaction center within the target molecule through a graph neural network and generate intermediate synthons; and ii) we predict the associated reactants based on the obtained synthons via a reactant generation model. While outperforming the state-of-the-art baselines by a significant margin, our model also provides chemically reasonable interpretation.

Retrosynthesis is the process of recursively decomposing target molecules into available building blocks. Computer-aided retrosynthesis can potentially assist chemists in designing synthetic routes; however, at present, it is cumbersome and lacks interpretability behind their predictions. In this study, we devise a novel template-free model for retrosynthetic expansion by automating the procedure that chemistsusedtodo. Our method plans synthesis in two steps, by first identifying the potential disconnection bonds of the molecule graph with a graph neural network and thereafter generating synthons according to the identified disconnection bonds of the target molecule graph, and then predicting the associated reactants SMILES based on the obtained synthons with a reactant prediction model. While outperforming previous state-of-the-art baselines by a significant margin on the benchmark datasets, our model also provides predictions with high diversity and chemically reasonable interpretation.

Unsupervised domain adaptation enables to alleviate the need for pixel-wise annotation in the semantic segmentation. One of the most common strategies is to translate images from the source domain to the target domain and then align their marginal distributions in the feature space using adversarial learning. However, source-to-target translation enlarges the bias in translated images and introduces extra computations, owing to the dominant data size of the source domain. Furthermore, consistency of the joint distribution in source and target domains cannot be guaranteed through global feature alignment. Here, we present an innovative framework, designed to mitigate the image translation bias and align cross-domain features with the same category. This is achieved by 1) performing the target-to-source translation and 2) reconstructing both source and target images from their predicted labels. Extensive experiments on adapting from synthetic to real urban scene understanding demonstrate that our framework competes favorably against existing state-of-the-art methods.

Molecule generation is to design new molecules with specific chemical properties and further to optimize the desired chemical properties. Following previous work, we encode molecules into continuous vectors in the latent space and then decode the embedding vectors into molecules under the variational autoencoder (VAE) framework. We investigate the posterior collapse problem of the current widely-used RNN-based VAEs for the molecule sequence generation. For the first time, we point out that the underestimated reconstruction loss of VAEs leads to the posterior collapse, and we also provide both analytical and experimental evidences to support our findings. To fix the problem and avoid the posterior collapse, we propose an effective and efficient solution in this work. Without bells and whistles, our method achieves the state-of-the-art reconstruction accuracy and competitive validity score on the ZINC 250K dataset. When generating 10,000 unique valid molecule sequences from the random prior sampling, it costs the JT-VAE 1450 seconds while our method only needs 9 seconds on a regular desktop machine.