University of Science and Technology of China
Abstract:Personalized image generation has emerged as a promising direction in multimodal content creation. It aims to synthesize images tailored to individual style preferences (e.g., color schemes, character appearances, layout) and semantic intentions (e.g., emotion, action, scene contexts) by leveraging user-interacted history images and multimodal instructions. Despite notable progress, existing methods -- whether based on diffusion models, large language models, or Large Multimodal Models (LMMs) -- struggle to accurately capture and fuse user style preferences and semantic intentions. In particular, the state-of-the-art LMM-based method suffers from the entanglement of visual features, leading to Guidance Collapse, where the generated images fail to preserve user-preferred styles or reflect the specified semantics. To address these limitations, we introduce DRC, a novel personalized image generation framework that enhances LMMs through Disentangled Representation Composition. DRC explicitly extracts user style preferences and semantic intentions from history images and the reference image, respectively, to form user-specific latent instructions that guide image generation within LMMs. Specifically, it involves two critical learning stages: 1) Disentanglement learning, which employs a dual-tower disentangler to explicitly separate style and semantic features, optimized via a reconstruction-driven paradigm with difficulty-aware importance sampling; and 2) Personalized modeling, which applies semantic-preserving augmentations to effectively adapt the disentangled representations for robust personalized generation. Extensive experiments on two benchmarks demonstrate that DRC shows competitive performance while effectively mitigating the guidance collapse issue, underscoring the importance of disentangled representation learning for controllable and effective personalized image generation.
Abstract:Imaging the human body's morphological and angiographic information is essential for diagnosing, monitoring, and treating medical conditions. Ultrasonography performs the morphological assessment of the soft tissue based on acoustic impedance variations, whereas photoacoustic tomography (PAT) can visualize blood vessels based on intrinsic hemoglobin absorption. Three-dimensional (3D) panoramic imaging of the vasculature is generally not practical in conventional ultrasonography with limited field-of-view (FOV) probes, and PAT does not provide sufficient scattering-based soft tissue morphological contrast. Complementing each other, fast panoramic rotational ultrasound tomography (RUST) and PAT are integrated for hybrid rotational ultrasound and photoacoustic tomography (RUS-PAT), which obtains 3D ultrasound structural and PAT angiographic images of the human body quasi-simultaneously. The RUST functionality is achieved in a cost-effective manner using a single-element ultrasonic transducer for ultrasound transmission and rotating arc-shaped arrays for 3D panoramic detection. RUST is superior to conventional ultrasonography, which either has a limited FOV with a linear array or is high-cost with a hemispherical array that requires both transmission and receiving. By switching the acoustic source to a light source, the system is conveniently converted to PAT mode to acquire angiographic images in the same region. Using RUS-PAT, we have successfully imaged the human head, breast, hand, and foot with a 10 cm diameter FOV, submillimeter isotropic resolution, and 10 s imaging time for each modality. The 3D RUS-PAT is a powerful tool for high-speed, 3D, dual-contrast imaging of the human body with potential for rapid clinical translation.
Abstract:Ranking models primarily focus on modeling the relative order of predictions while often neglecting the significance of the accuracy of their absolute values. However, accurate absolute values are essential for certain downstream tasks, necessitating the calibration of the original predictions. To address this, existing calibration approaches typically employ predefined transformation functions with order-preserving properties to adjust the original predictions. Unfortunately, these functions often adhere to fixed forms, such as piece-wise linear functions, which exhibit limited expressiveness and flexibility, thereby constraining their effectiveness in complex calibration scenarios. To mitigate this issue, we propose implementing a calibrator using an Unconstrained Monotonic Neural Network (UMNN), which can learn arbitrary monotonic functions with great modeling power. This approach significantly relaxes the constraints on the calibrator, improving its flexibility and expressiveness while avoiding excessively distorting the original predictions by requiring monotonicity. Furthermore, to optimize this highly flexible network for calibration, we introduce a novel additional loss function termed Smooth Calibration Loss (SCLoss), which aims to fulfill a necessary condition for achieving the ideal calibration state. Extensive offline experiments confirm the effectiveness of our method in achieving superior calibration performance. Moreover, deployment in Kuaishou's large-scale online video ranking system demonstrates that the method's calibration improvements translate into enhanced business metrics. The source code is available at https://github.com/baiyimeng/UMC.
Abstract:This work considers the three-dimensional (3-D) positioning problem in a Terahertz (THz) system enabled by a modular extra-large (XL) array with sub-connected architecture. Our purpose is to estimate the Cartesian Coordinates of multiple user equipments (UEs) with the received signal of the RF chains while considering the spatial non-stationarity (SNS). We apply the hybrid spherical-planar wave model (HSPWM) as the channel model owing to the structual feature of the modular array, and propose a 3-D localization algorithm with relatively high accuracy and low complexity. Specifically, we first distinguish the visible sub-arrays (SAs) located in the VR and estimate the angles-of-arrival (AoAs) from each UE to typical visible SAs with the largest receive power via compressed sensing (CS) method. In addition, we apply the weighted least square (WLS) method to obtain a coarse 3-D position estimation of each UE according to the AoA estimations. Then, we estimate the AoAs of the other SAs with a reduced dictionary (RD)-CS-based method for lower computational complexity, and utilize all the efficient AoA estimations to derive a fine position estimation. Simulation results indicate that the proposed positioning framework based on modular XL-array can achieve satisfactory accuracy with evident reduction in complexity. Furthermore, the deployment of SAs and the allocation of antenna elements need to be specially designed for better positioning performance.
Abstract:Vision Transformer (ViT) has achieved remarkable results in object detection for synthetic aperture radar (SAR) images, owing to its exceptional ability to extract global features. However, it struggles with the extraction of multi-scale local features, leading to limited performance in detecting small targets, especially when they are densely arranged. Therefore, we propose Density-Sensitive Vision Transformer with Adaptive Tokens (DenSe-AdViT) for dense SAR target detection. We design a Density-Aware Module (DAM) as a preliminary component that generates a density tensor based on target distribution. It is guided by a meticulously crafted objective metric, enabling precise and effective capture of the spatial distribution and density of objects. To integrate the multi-scale information enhanced by convolutional neural networks (CNNs) with the global features derived from the Transformer, Density-Enhanced Fusion Module (DEFM) is proposed. It effectively refines attention toward target-survival regions with the assist of density mask and the multiple sources features. Notably, our DenSe-AdViT achieves 79.8% mAP on the RSDD dataset and 92.5% on the SIVED dataset, both of which feature a large number of densely distributed vehicle targets.
Abstract:Reward Model (RM) has demonstrated impressive potential for enhancing Large Language Models (LLM), as RM can serve as a proxy for human preferences, providing signals to guide LLMs' behavior in various tasks. In this paper, we provide a comprehensive overview of relevant research, exploring RMs from the perspectives of preference collection, reward modeling, and usage. Next, we introduce the applications of RMs and discuss the benchmarks for evaluation. Furthermore, we conduct an in-depth analysis of the challenges existing in the field and dive into the potential research directions. This paper is dedicated to providing beginners with a comprehensive introduction to RMs and facilitating future studies. The resources are publicly available at github\footnote{https://github.com/JLZhong23/awesome-reward-models}.
Abstract:Watermarking has emerged as a promising technique for detecting texts generated by LLMs. Current research has primarily focused on three design criteria: high quality of the watermarked text, high detectability, and robustness against removal attack. However, the security against spoofing attacks remains relatively understudied. For example, a piggyback attack can maliciously alter the meaning of watermarked text-transforming it into hate speech-while preserving the original watermark, thereby damaging the reputation of the LLM provider. We identify two core challenges that make defending against spoofing difficult: (1) the need for watermarks to be both sensitive to semantic-distorting changes and insensitive to semantic-preserving edits, and (2) the contradiction between the need to detect global semantic shifts and the local, auto-regressive nature of most watermarking schemes. To address these challenges, we propose a semantic-aware watermarking algorithm that post-hoc embeds watermarks into a given target text while preserving its original meaning. Our method introduces a semantic mapping model, which guides the generation of a green-red token list, contrastively trained to be sensitive to semantic-distorting changes and insensitive to semantic-preserving changes. Experiments on two standard benchmarks demonstrate strong robustness against removal attacks and security against spoofing attacks, including sentiment reversal and toxic content insertion, while maintaining high watermark detectability. Our approach offers a significant step toward more secure and semantically aware watermarking for LLMs. Our code is available at https://github.com/UCSB-NLP-Chang/contrastive-watermark.
Abstract:Generative neural image compression supports data representation at extremely low bitrate, synthesizing details at the client and consistently producing highly realistic images. By leveraging the similarities between quantization error and additive noise, diffusion-based generative image compression codecs can be built using a latent diffusion model to "denoise" the artifacts introduced by quantization. However, we identify three critical gaps in previous approaches following this paradigm (namely, the noise level, noise type, and discretization gaps) that result in the quantized data falling out of the data distribution known by the diffusion model. In this work, we propose a novel quantization-based forward diffusion process with theoretical foundations that tackles all three aforementioned gaps. We achieve this through universal quantization with a carefully tailored quantization schedule and a diffusion model trained with uniform noise. Compared to previous work, our proposal produces consistently realistic and detailed reconstructions, even at very low bitrates. In such a regime, we achieve the best rate-distortion-realism performance, outperforming previous related works.
Abstract:We present ThinkPrune, a simple yet effective method for pruning the thinking length for long-thinking LLMs, which has been found to often produce inefficient and redundant thinking processes. Existing preliminary explorations of reducing thinking length primarily focus on forcing the thinking process to early exit, rather than adapting the LLM to optimize and consolidate the thinking process, and therefore the length-performance tradeoff observed so far is sub-optimal. To fill this gap, ThinkPrune offers a simple solution that continuously trains the long-thinking LLMs via reinforcement learning (RL) with an added token limit, beyond which any unfinished thoughts and answers will be discarded, resulting in a zero reward. To further preserve model performance, we introduce an iterative length pruning approach, where multiple rounds of RL are conducted, each with an increasingly more stringent token limit. We observed that ThinkPrune results in a remarkable performance-length tradeoff -- on the AIME24 dataset, the reasoning length of DeepSeek-R1-Distill-Qwen-1.5B can be reduced by half with only 2% drop in performance. We also observed that after pruning, the LLMs can bypass unnecessary steps while keeping the core reasoning process complete. Code is available at https://github.com/UCSB-NLP-Chang/ThinkPrune.
Abstract:Multimodal large language models (MLLMs), such as GPT-4o, are garnering significant attention. During the exploration of MLLM training, we identified Modality Composition Incoherence, a phenomenon that the proportion of a certain modality varies dramatically across different examples. It exacerbates the challenges of addressing mini-batch imbalances, which lead to uneven GPU utilization between Data Parallel (DP) instances and severely degrades the efficiency and scalability of MLLM training, ultimately affecting training speed and hindering further research on MLLMs. To address these challenges, we introduce OrchMLLM, a comprehensive framework designed to mitigate the inefficiencies in MLLM training caused by Modality Composition Incoherence. First, we propose Batch Post-Balancing Dispatcher, a technique that efficiently eliminates mini-batch imbalances in sequential data. Additionally, we integrate MLLM Global Orchestrator into the training framework to orchestrate multimodal data and tackle the issues arising from Modality Composition Incoherence. We evaluate OrchMLLM across various MLLM sizes, demonstrating its efficiency and scalability. Experimental results reveal that OrchMLLM achieves a Model FLOPs Utilization (MFU) of $41.6\%$ when training an 84B MLLM with three modalities on $2560$ H100 GPUs, outperforming Megatron-LM by up to $3.1\times$ in throughput.