Abstract:The differing representation spaces required for visual understanding and generation pose a challenge in unifying them within the autoregressive paradigm of large language models. A vision tokenizer trained for reconstruction excels at capturing low-level perceptual details, making it well-suited for visual generation but lacking high-level semantic representations for understanding tasks. Conversely, a vision encoder trained via contrastive learning aligns well with language but struggles to decode back into the pixel space for generation tasks. To bridge this gap, we propose DualToken, a method that unifies representations for both understanding and generation within a single tokenizer. However, directly integrating reconstruction and semantic objectives in a single tokenizer creates conflicts, leading to degraded performance in both reconstruction quality and semantic performance. Instead of forcing a single codebook to handle both semantic and perceptual information, DualToken disentangles them by introducing separate codebooks for high and low-level features, effectively transforming their inherent conflict into a synergistic relationship. As a result, DualToken achieves state-of-the-art performance in both reconstruction and semantic tasks while demonstrating remarkable effectiveness in downstream MLLM understanding and generation tasks. Notably, we also show that DualToken, as a unified tokenizer, surpasses the naive combination of two distinct types vision encoders, providing superior performance within a unified MLLM.
Abstract:Multimodal fusion breaks through the barriers between diverse modalities and has already yielded numerous impressive performances. However, in various specialized fields, it is struggling to obtain sufficient alignment data for the training process, which seriously limits the use of previously elegant models. Thus, semi-supervised learning attempts to achieve multimodal alignment with fewer matched pairs but traditional methods like pseudo-labeling are difficult to apply in domains with no label information. To address these problems, we transform semi-supervised multimodal alignment into a manifold matching problem and propose a new method based on CLIP, named Gentle-CLIP. Specifically, we design a novel semantic density distribution loss to explore implicit semantic alignment information from unpaired multimodal data by constraining the latent representation distribution with fine granularity, thus eliminating the need for numerous strictly matched pairs. Meanwhile, we introduce multi-kernel maximum mean discrepancy as well as self-supervised contrastive loss to pull separate modality distributions closer and enhance the stability of the representation distribution. In addition, the contrastive loss used in CLIP is employed on the supervised matched data to prevent negative optimization. Extensive experiments conducted on a range of tasks in various fields, including protein, remote sensing, and the general vision-language field, demonstrate the effectiveness of our proposed Gentle-CLIP.
Abstract:Metagenomic data, comprising mixed multi-species genomes, are prevalent in diverse environments like oceans and soils, significantly impacting human health and ecological functions. However, current research relies on K-mer representations, limiting the capture of structurally relevant gene contexts. To address these limitations and further our understanding of complex relationships between metagenomic sequences and their functions, we introduce a protein-based gene representation as a context-aware and structure-relevant tokenizer. Our approach includes Masked Gene Modeling (MGM) for gene group-level pre-training, providing insights into inter-gene contextual information, and Triple Enhanced Metagenomic Contrastive Learning (TEM-CL) for gene-level pre-training to model gene sequence-function relationships. MGM and TEM-CL constitute our novel metagenomic language model {\NAME}, pre-trained on 100 million metagenomic sequences. We demonstrate the superiority of our proposed {\NAME} on eight datasets.