CITER: Collaborative Inference for Efficient Large Language Model Decoding with Token-Level Routing

Large language models have achieved remarkable success in various tasks but suffer from high computational costs during inference, limiting their deployment in resource-constrained applications. To address this issue, we propose a novel CITER (\textbf{C}ollaborative \textbf{I}nference with \textbf{T}oken-l\textbf{E}vel \textbf{R}outing) framework that enables efficient collaboration between small and large language models (SLMs & LLMs) through a token-level routing strategy. Specifically, CITER routes non-critical tokens to an SLM for efficiency and routes critical tokens to an LLM for generalization quality. We formulate router training as a policy optimization, where the router receives rewards based on both the quality of predictions and the inference costs of generation. This allows the router to learn to predict token-level routing scores and make routing decisions based on both the current token and the future impact of its decisions. To further accelerate the reward evaluation process, we introduce a shortcut which significantly reduces the costs of the reward estimation and improving the practicality of our approach. Extensive experiments on five benchmark datasets demonstrate that CITER reduces the inference costs while preserving high-quality generation, offering a promising solution for real-time and resource-constrained applications.

Exploring the Role of Explicit Temporal Modeling in Multimodal Large Language Models for Video Understanding

Applying Multimodal Large Language Models (MLLMs) to video understanding presents significant challenges due to the need to model temporal relations across frames. Existing approaches adopt either implicit temporal modeling, relying solely on the LLM decoder, or explicit temporal modeling, employing auxiliary temporal encoders. To investigate this debate between the two paradigms, we propose the Stackable Temporal Encoder (STE). STE enables flexible explicit temporal modeling with adjustable temporal receptive fields and token compression ratios. Using STE, we systematically compare implicit and explicit temporal modeling across dimensions such as overall performance, token compression effectiveness, and temporal-specific understanding. We also explore STE's design considerations and broader impacts as a plug-in module and in image modalities. Our findings emphasize the critical role of explicit temporal modeling, providing actionable insights to advance video MLLMs.

VisionLLM-based Multimodal Fusion Network for Glottic Carcinoma Early Detection

The early detection of glottic carcinoma is critical for improving patient outcomes, as it enables timely intervention, preserves vocal function, and significantly reduces the risk of tumor progression and metastasis. However, the similarity in morphology between glottic carcinoma and vocal cord dysplasia results in suboptimal detection accuracy. To address this issue, we propose a vision large language model-based (VisionLLM-based) multimodal fusion network for glottic carcinoma detection, known as MMGC-Net. By integrating image and text modalities, multimodal models can capture complementary information, leading to more accurate and robust predictions. In this paper, we collect a private real glottic carcinoma dataset named SYSU1H from the First Affiliated Hospital of Sun Yat-sen University, with 5,799 image-text pairs. We leverage an image encoder and additional Q-Former to extract vision embeddings and the Large Language Model Meta AI (Llama3) to obtain text embeddings. These modalities are then integrated through a laryngeal feature fusion block, enabling a comprehensive integration of image and text features, thereby improving the glottic carcinoma identification performance. Extensive experiments on the SYSU1H dataset demonstrate that MMGC-Net can achieve state-of-the-art performance, which is superior to previous multimodal models.

Cross-Lingual Text-Rich Visual Comprehension: An Information Theory Perspective

Recent Large Vision-Language Models (LVLMs) have shown promising reasoning capabilities on text-rich images from charts, tables, and documents. However, the abundant text within such images may increase the model's sensitivity to language. This raises the need to evaluate LVLM performance on cross-lingual text-rich visual inputs, where the language in the image differs from the language of the instructions. To address this, we introduce XT-VQA (Cross-Lingual Text-Rich Visual Question Answering), a benchmark designed to assess how LVLMs handle language inconsistency between image text and questions. XT-VQA integrates five existing text-rich VQA datasets and a newly collected dataset, XPaperQA, covering diverse scenarios that require faithful recognition and comprehension of visual information despite language inconsistency. Our evaluation of prominent LVLMs on XT-VQA reveals a significant drop in performance for cross-lingual scenarios, even for models with multilingual capabilities. A mutual information analysis suggests that this performance gap stems from cross-lingual questions failing to adequately activate relevant visual information. To mitigate this issue, we propose MVCL-MI (Maximization of Vision-Language Cross-Lingual Mutual Information), where a visual-text cross-lingual alignment is built by maximizing mutual information between the model's outputs and visual information. This is achieved by distilling knowledge from monolingual to cross-lingual settings through KL divergence minimization, where monolingual output logits serve as a teacher. Experimental results on the XT-VQA demonstrate that MVCL-MI effectively reduces the visual-text cross-lingual performance disparity while preserving the inherent capabilities of LVLMs, shedding new light on the potential practice for improving LVLMs. Codes are available at: https://github.com/Stardust-y/XTVQA.git

Compositional Zero-Shot Learning with Contextualized Cues and Adaptive Contrastive Training

Compositional Zero-Shot Learning (CZSL) aims to recognize unseen combinations of seen attributes and objects. Current CLIP-based methods in CZSL, despite their advancements, often fail to effectively understand and link the attributes and objects due to inherent limitations in CLIP's pretraining mechanisms. To address these shortcomings, this paper introduces a novel framework, Understanding and Linking Attributes and Objects (ULAO) in CZSL, which comprises two innovative modules. The Understanding Attributes and Objects (UAO) module improves primitive understanding by sequential primitive prediction and leveraging recognized objects as contextual hints for attribute classification. Concurrently, the Linking Attributes and Objects (LAO) module improves the attribute-object linkage understanding through a new contrastive learning strategy that incorporates tailored hard negative generation and adaptive loss adjustments. We demonstrate our model's superiority by showcasing its state-of-the-art performance across three benchmark datasets in both Closed-World (CW) and Open-World (OW) scenarios.

MMedPO: Aligning Medical Vision-Language Models with Clinical-Aware Multimodal Preference Optimization

The advancement of Large Vision-Language Models (LVLMs) has propelled their application in the medical field. However, Medical LVLMs (Med-LVLMs) encounter factuality challenges due to modality misalignment, where the models prioritize textual knowledge over visual input, leading to hallucinations that contradict information in medical images. Previous attempts to enhance modality alignment in Med-LVLMs through preference optimization have inadequately mitigated clinical relevance in preference data, making these samples easily distinguishable and reducing alignment effectiveness. To address this challenge, we propose MMedPO, a novel multimodal medical preference optimization approach that considers the clinical relevance of preference samples to enhance Med-LVLM alignment. MMedPO curates multimodal preference data by introducing two types of dispreference: (1) plausible hallucinations injected through target Med-LVLMs or GPT-4o to produce medically inaccurate responses, and (2) lesion region neglect achieved through local lesion-noising, disrupting visual understanding of critical areas. We then calculate clinical relevance for each sample based on scores from multiple Med-LLMs and visual tools, and integrate these scores into the preference optimization process as weights, enabling effective alignment. Our experiments demonstrate that MMedPO significantly enhances factual accuracy in Med-LVLMs, achieving substantial improvements over existing preference optimization methods by averaging 14.2% and 51.7% across the Med-VQA and report generation tasks. Our code are available in https://github.com/aiming-lab/MMedPO.

LoCo: Low-Contrast-Enhanced Contrastive Learning for Semi-Supervised Endoscopic Image Segmentation

The segmentation of endoscopic images plays a vital role in computer-aided diagnosis and treatment. The advancements in deep learning have led to the employment of numerous models for endoscopic tumor segmentation, achieving promising segmentation performance. Despite recent advancements, precise segmentation remains challenging due to limited annotations and the issue of low contrast. To address these issues, we propose a novel semi-supervised segmentation framework termed LoCo via low-contrast-enhanced contrastive learning (LCC). This innovative approach effectively harnesses the vast amounts of unlabeled data available for endoscopic image segmentation, improving both accuracy and robustness in the segmentation process. Specifically, LCC incorporates two advanced strategies to enhance the distinctiveness of low-contrast pixels: inter-class contrast enhancement (ICE) and boundary contrast enhancement (BCE), enabling models to segment low-contrast pixels among malignant tumors, benign tumors, and normal tissues. Additionally, a confidence-based dynamic filter (CDF) is designed for pseudo-label selection, enhancing the utilization of generated pseudo-labels for unlabeled data with a specific focus on minority classes. Extensive experiments conducted on two public datasets, as well as a large proprietary dataset collected over three years, demonstrate that LoCo achieves state-of-the-art results, significantly outperforming previous methods. The source code of LoCo is available at the URL of https://github.com/AnoK3111/LoCo.

SAM-Swin: SAM-Driven Dual-Swin Transformers with Adaptive Lesion Enhancement for Laryngo-Pharyngeal Tumor Detection

Laryngo-pharyngeal cancer (LPC) is a highly lethal malignancy in the head and neck region. Recent advancements in tumor detection, particularly through dual-branch network architectures, have significantly improved diagnostic accuracy by integrating global and local feature extraction. However, challenges remain in accurately localizing lesions and fully capitalizing on the complementary nature of features within these branches. To address these issues, we propose SAM-Swin, an innovative SAM-driven Dual-Swin Transformer for laryngo-pharyngeal tumor detection. This model leverages the robust segmentation capabilities of the Segment Anything Model 2 (SAM2) to achieve precise lesion segmentation. Meanwhile, we present a multi-scale lesion-aware enhancement module (MS-LAEM) designed to adaptively enhance the learning of nuanced complementary features across various scales, improving the quality of feature extraction and representation. Furthermore, we implement a multi-scale class-aware guidance (CAG) loss that delivers multi-scale targeted supervision, thereby enhancing the model's capacity to extract class-specific features. To validate our approach, we compiled three LPC datasets from the First Affiliated Hospital (FAHSYSU), the Sixth Affiliated Hospital (SAHSYSU) of Sun Yat-sen University, and Nanfang Hospital of Southern Medical University (NHSMU). The FAHSYSU dataset is utilized for internal training, while the SAHSYSU and NHSMU datasets serve for external evaluation. Extensive experiments demonstrate that SAM-Swin outperforms state-of-the-art methods, showcasing its potential for advancing LPC detection and improving patient outcomes. The source code of SAM-Swin is available at the URL of \href{https://github.com/VVJia/SAM-Swin}{https://github.com/VVJia/SAM-Swin}.

Reinforcement Learning Based Bidding Framework with High-dimensional Bids in Power Markets

Over the past decade, bidding in power markets has attracted widespread attention. Reinforcement Learning (RL) has been widely used for power market bidding as a powerful AI tool to make decisions under real-world uncertainties. However, current RL methods mostly employ low dimensional bids, which significantly diverge from the N price-power pairs commonly used in the current power markets. The N-pair bidding format is denoted as High Dimensional Bids (HDBs), which has not been fully integrated into the existing RL-based bidding methods. The loss of flexibility in current RL bidding methods could greatly limit the bidding profits and make it difficult to tackle the rising uncertainties brought by renewable energy generations. In this paper, we intend to propose a framework to fully utilize HDBs for RL-based bidding methods. First, we employ a special type of neural network called Neural Network Supply Functions (NNSFs) to generate HDBs in the form of N price-power pairs. Second, we embed the NNSF into a Markov Decision Process (MDP) to make it compatible with most existing RL methods. Finally, experiments on Energy Storage Systems (ESSs) in the PJM Real-Time (RT) power market show that the proposed bidding method with HDBs can significantly improve bidding flexibility, thereby improving the profit of the state-of-the-art RL bidding methods.

Forgetting Through Transforming: Enabling Federated Unlearning via Class-Aware Representation Transformation

Federated Unlearning (FU) enables clients to selectively remove the influence of specific data from a trained federated learning model, addressing privacy concerns and regulatory requirements. However, existing FU methods often struggle to balance effective erasure with model utility preservation, especially for class-level unlearning in non-IID settings. We propose Federated Unlearning via Class-aware Representation Transformation (FUCRT), a novel method that achieves unlearning through class-aware representation transformation. FUCRT employs two key components: (1) a transformation class selection strategy to identify optimal forgetting directions, and (2) a transformation alignment technique using dual class-aware contrastive learning to ensure consistent transformations across clients. Extensive experiments on four datasets demonstrate FUCRT's superior performance in terms of erasure guarantee, model utility preservation, and efficiency. FUCRT achieves complete (100\%) erasure of unlearning classes while maintaining or improving performance on remaining classes, outperforming state-of-the-art baselines across both IID and Non-IID settings. Analysis of the representation space reveals FUCRT's ability to effectively merge unlearning class representations with the transformation class from remaining classes, closely mimicking the model retrained from scratch.