SpeCache: Speculative Key-Value Caching for Efficient Generation of LLMs

Transformer-based large language models (LLMs) have already achieved remarkable results on long-text tasks, but the limited GPU memory (VRAM) resources struggle to accommodate the linearly growing demand for key-value (KV) cache as the sequence length increases, which has become a bottleneck for the application of LLMs on long sequences. Existing KV cache compression methods include eviction, merging, or quantization of the KV cache to reduce its size. However, compression results in irreversible information forgetting, potentially affecting the accuracy of subsequent decoding. In this paper, we propose SpeCache, which takes full advantage of the large and easily expandable CPU memory to offload the complete KV cache, and dynamically fetches KV pairs back in each decoding step based on their importance measured by low-bit KV cache copy in VRAM. To avoid inference latency caused by CPU-GPU communication, SpeCache speculatively predicts the KV pairs that the next token might attend to, allowing us to prefetch them before the next decoding step which enables parallelization of prefetching and computation. Experiments on LongBench and Needle-in-a-Haystack benchmarks verify that SpeCache effectively reduces VRAM usage while avoiding information forgetting for long sequences without re-training, even with a 10x high KV cache compression ratio.

Mixture of Lookup Experts

Mixture-of-Experts (MoE) activates only a subset of experts during inference, allowing the model to maintain low inference FLOPs and latency even as the parameter count scales up. However, since MoE dynamically selects the experts, all the experts need to be loaded into VRAM. Their large parameter size still limits deployment, and offloading, which load experts into VRAM only when needed, significantly increase inference latency. To address this, we propose Mixture of Lookup Experts (MoLE), a new MoE architecture that is efficient in both communication and VRAM usage. In MoLE, the experts are Feed-Forward Networks (FFNs) during training, taking the output of the embedding layer as input. Before inference, these experts can be re-parameterized as lookup tables (LUTs) that retrieves expert outputs based on input ids, and offloaded to storage devices. Therefore, we do not need to perform expert computations during inference. Instead, we directly retrieve the expert's computation results based on input ids and load them into VRAM, and thus the resulting communication overhead is negligible. Experiments show that, with the same FLOPs and VRAM usage, MoLE achieves inference speeds comparable to dense models and significantly faster than MoE with experts offloading, while maintaining performance on par with MoE.

Adaptive Label Correction for Robust Medical Image Segmentation with Noisy Labels

Deep learning has shown remarkable success in medical image analysis, but its reliance on large volumes of high-quality labeled data limits its applicability. While noisy labeled data are easier to obtain, directly incorporating them into training can degrade model performance. To address this challenge, we propose a Mean Teacher-based Adaptive Label Correction (ALC) self-ensemble framework for robust medical image segmentation with noisy labels. The framework leverages the Mean Teacher architecture to ensure consistent learning under noise perturbations. It includes an adaptive label refinement mechanism that dynamically captures and weights differences across multiple disturbance versions to enhance the quality of noisy labels. Additionally, a sample-level uncertainty-based label selection algorithm is introduced to prioritize high-confidence samples for network updates, mitigating the impact of noisy annotations. Consistency learning is integrated to align the predictions of the student and teacher networks, further enhancing model robustness. Extensive experiments on two public datasets demonstrate the effectiveness of the proposed framework, showing significant improvements in segmentation performance. By fully exploiting the strengths of the Mean Teacher structure, the ALC framework effectively processes noisy labels, adapts to challenging scenarios, and achieves competitive results compared to state-of-the-art methods.

Post-Training Quantization for Diffusion Transformer via Hierarchical Timestep Grouping

Diffusion Transformer (DiT) has now become the preferred choice for building image generation models due to its great generation capability. Unlike previous convolution-based UNet models, DiT is purely composed of a stack of transformer blocks, which renders DiT excellent in scalability like large language models. However, the growing model size and multi-step sampling paradigm bring about considerable pressure on deployment and inference. In this work, we propose a post-training quantization framework tailored for Diffusion Transforms to tackle these challenges. We firstly locate that the quantization difficulty of DiT mainly originates from the time-dependent channel-specific outliers. We propose a timestep-aware shift-and-scale strategy to smooth the activation distribution to reduce the quantization error. Secondly, based on the observation that activations of adjacent timesteps have similar distributions, we utilize a hierarchical clustering scheme to divide the denoising timesteps into multiple groups. We further design a re-parameterization scheme which absorbs the quantization parameters into nearby module to avoid redundant computations. Comprehensive experiments demonstrate that out PTQ method successfully quantize the Diffusion Transformer into 8-bit weight and 8-bit activation (W8A8) with state-of-the-art FiD score. And our method can further quantize DiT model into 4-bit weight and 8-bit activation (W4A8) without sacrificing generation quality.

DynCIM: Dynamic Curriculum for Imbalanced Multimodal Learning

Multimodal learning integrates complementary information from diverse modalities to enhance the decision-making process. However, the potential of multimodal collaboration remains under-exploited due to disparities in data quality and modality representation capabilities. To address this, we introduce DynCIM, a novel dynamic curriculum learning framework designed to quantify the inherent imbalances from both sample and modality perspectives. DynCIM employs a sample-level curriculum to dynamically assess each sample's difficulty according to prediction deviation, consistency, and stability, while a modality-level curriculum measures modality contributions from global and local. Furthermore, a gating-based dynamic fusion mechanism is introduced to adaptively adjust modality contributions, minimizing redundancy and optimizing fusion effectiveness. Extensive experiments on six multimodal benchmarking datasets, spanning both bimodal and trimodal scenarios, demonstrate that DynCIM consistently outperforms state-of-the-art methods. Our approach effectively mitigates modality and sample imbalances while enhancing adaptability and robustness in multimodal learning tasks. Our code is available at https://github.com/Raymond-Qiancx/DynCIM.

ELIP: Enhanced Visual-Language Foundation Models for Image Retrieval

The objective in this paper is to improve the performance of text-to-image retrieval. To this end, we introduce a new framework that can boost the performance of large-scale pre-trained vision-language models, so that they can be used for text-to-image re-ranking. The approach, Enhanced Language-Image Pre-training (ELIP), uses the text query to predict a set of visual prompts to condition the ViT image encoding. ELIP can easily be applied to the commonly used CLIP/SigLIP and the state-of-the-art BLIP-2 architectures. To train the architecture with limited computing resources, we develop a 'student friendly' best practice involving global hard sample mining, and selection and curation of a large-scale dataset. On the evaluation side, we set up two new out-of-distribution benchmarks, Occluded COCO and ImageNet-R, to assess the zero-shot generalisation of the models to different domains. Benefiting from the novel architecture and data curation, experiments show our enhanced network significantly boosts CLIP/SigLIP performance and outperforms the state-of-the-art BLIP-2 model on text-to-image retrieval.

VipDiff: Towards Coherent and Diverse Video Inpainting via Training-free Denoising Diffusion Models

Recent video inpainting methods have achieved encouraging improvements by leveraging optical flow to guide pixel propagation from reference frames either in the image space or feature space. However, they would produce severe artifacts in the mask center when the masked area is too large and no pixel correspondences can be found for the center. Recently, diffusion models have demonstrated impressive performance in generating diverse and high-quality images, and have been exploited in a number of works for image inpainting. These methods, however, cannot be applied directly to videos to produce temporal-coherent inpainting results. In this paper, we propose a training-free framework, named VipDiff, for conditioning diffusion model on the reverse diffusion process to produce temporal-coherent inpainting results without requiring any training data or fine-tuning the pre-trained diffusion models. VipDiff takes optical flow as guidance to extract valid pixels from reference frames to serve as constraints in optimizing the randomly sampled Gaussian noise, and uses the generated results for further pixel propagation and conditional generation. VipDiff also allows for generating diverse video inpainting results over different sampled noise. Experiments demonstrate that VipDiff can largely outperform state-of-the-art video inpainting methods in terms of both spatial-temporal coherence and fidelity.

Parallel Sequence Modeling via Generalized Spatial Propagation Network

We present the Generalized Spatial Propagation Network (GSPN), a new attention mechanism optimized for vision tasks that inherently captures 2D spatial structures. Existing attention models, including transformers, linear attention, and state-space models like Mamba, process multi-dimensional data as 1D sequences, compromising spatial coherence and efficiency. GSPN overcomes these limitations by directly operating on spatially coherent image data and forming dense pairwise connections through a line-scan approach. Central to GSPN is the Stability-Context Condition, which ensures stable, context-aware propagation across 2D sequences and reduces the effective sequence length to $\sqrt{N}$ for a square map with N elements, significantly enhancing computational efficiency. With learnable, input-dependent weights and no reliance on positional embeddings, GSPN achieves superior spatial fidelity and state-of-the-art performance in vision tasks, including ImageNet classification, class-guided image generation, and text-to-image generation. Notably, GSPN accelerates SD-XL with softmax-attention by over $84\times$ when generating 16K images.

Eve: Efficient Multimodal Vision Language Models with Elastic Visual Experts

Multimodal vision language models (VLMs) have made significant progress with the support of continuously increasing model sizes and data volumes. Running VLMs on edge devices has become a challenge for their widespread application. There are several efficient VLM efforts, but they often sacrifice linguistic capabilities to enhance multimodal abilities, or require extensive training. To address this quandary,we introduce the innovative framework of Efficient Vision Language Models with Elastic Visual Experts (Eve). By strategically incorporating adaptable visual expertise at multiple stages of training, Eve strikes a balance between preserving linguistic abilities and augmenting multimodal capabilities. This balanced approach results in a versatile model with only 1.8B parameters that delivers significant improvements in both multimodal and linguistic tasks. Notably, in configurations below 3B parameters, Eve distinctly outperforms in language benchmarks and achieves state-of-the-art results 68.87% in VLM Benchmarks. Additionally, its multimodal accuracy outstrips that of the larger 7B LLaVA-1.5 model.

PruneVid: Visual Token Pruning for Efficient Video Large Language Models

In this paper, we introduce PruneVid, a visual token pruning method designed to enhance the efficiency of multi-modal video understanding. Large Language Models (LLMs) have shown promising performance in video tasks due to their extended capabilities in comprehending visual modalities. However, the substantial redundancy in video data presents significant computational challenges for LLMs. To address this issue, we introduce a training-free method that 1) minimizes video redundancy by merging spatial-temporal tokens, and 2) leverages LLMs' reasoning capabilities to selectively prune visual features relevant to question tokens, enhancing model efficiency. We validate our method across multiple video benchmarks, which demonstrate that PruneVid can prune over 80% of tokens while maintaining competitive performance combined with different model networks. This highlights its superior effectiveness and efficiency compared to existing pruning methods. Code: https://github.com/Visual-AI/PruneVid.