Dynamic Base model Shift for Delta Compression

Transformer-based models with the pretrain-finetune paradigm bring about significant progress, along with the heavy storage and deployment costs of finetuned models on multiple tasks. Delta compression attempts to lower the costs by reducing the redundancy of delta parameters (i.e., the difference between the finetuned and pre-trained model weights) through pruning or quantization. However, existing methods by default employ the pretrained model as the base model and compress the delta parameters for every task, which may causes significant performance degradation, especially when the compression rate is extremely high. To tackle this issue, we investigate the impact of different base models on the performance of delta compression and find that the pre-trained base model can hardly be optimal. To this end, we propose Dynamic Base Model Shift (DBMS), which dynamically adapts the base model to the target task before performing delta compression. Specifically, we adjust two parameters, which respectively determine the magnitude of the base model shift and the overall scale of delta compression, to boost the compression performance on each task. Through low-cost learning of these two parameters, our DBMS can maintain most of the finetuned model's performance even under an extremely high compression ratio setting, significantly surpassing existing methods. Moreover, our DBMS is orthogonal and can be integrated with a variety of other methods, and it has been evaluated across different types of models including language, vision transformer, and multi-modal models.

Generalizable Pancreas Segmentation via a Dual Self-Supervised Learning Framework

Recently, numerous pancreas segmentation methods have achieved promising performance on local single-source datasets. However, these methods don't adequately account for generalizability issues, and hence typically show limited performance and low stability on test data from other sources. Considering the limited availability of distinct data sources, we seek to improve the generalization performance of a pancreas segmentation model trained with a single-source dataset, i.e., the single source generalization task. In particular, we propose a dual self-supervised learning model that incorporates both global and local anatomical contexts. Our model aims to fully exploit the anatomical features of the intra-pancreatic and extra-pancreatic regions, and hence enhance the characterization of the high-uncertainty regions for more robust generalization. Specifically, we first construct a global-feature contrastive self-supervised learning module that is guided by the pancreatic spatial structure. This module obtains complete and consistent pancreatic features through promoting intra-class cohesion, and also extracts more discriminative features for differentiating between pancreatic and non-pancreatic tissues through maximizing inter-class separation. It mitigates the influence of surrounding tissue on the segmentation outcomes in high-uncertainty regions. Subsequently, a local-image restoration self-supervised learning module is introduced to further enhance the characterization of the high uncertainty regions. In this module, informative anatomical contexts are actually learned to recover randomly corrupted appearance patterns in those regions.

OmniCaptioner: One Captioner to Rule Them All

We propose OmniCaptioner, a versatile visual captioning framework for generating fine-grained textual descriptions across a wide variety of visual domains. Unlike prior methods limited to specific image types (e.g., natural images or geometric visuals), our framework provides a unified solution for captioning natural images, visual text (e.g., posters, UIs, textbooks), and structured visuals (e.g., documents, tables, charts). By converting low-level pixel information into semantically rich textual representations, our framework bridges the gap between visual and textual modalities. Our results highlight three key advantages: (i) Enhanced Visual Reasoning with LLMs, where long-context captions of visual modalities empower LLMs, particularly the DeepSeek-R1 series, to reason effectively in multimodal scenarios; (ii) Improved Image Generation, where detailed captions improve tasks like text-to-image generation and image transformation; and (iii) Efficient Supervised Fine-Tuning (SFT), which enables faster convergence with less data. We believe the versatility and adaptability of OmniCaptioner can offer a new perspective for bridging the gap between language and visual modalities.

Local Information Matters: Inference Acceleration For Grounded Conversation Generation Models Through Adaptive Local-Aware Token Pruning

Grounded Conversation Generation (GCG) is an emerging vision-language task that requires models to generate natural language responses seamlessly intertwined with corresponding object segmentation masks. Recent models, such as GLaMM and OMG-LLaVA, achieve pixel-level grounding but incur significant computational costs due to processing a large number of visual tokens. Existing token pruning methods, like FastV and PyramidDrop, fail to preserve the local visual features critical for accurate grounding, leading to substantial performance drops in GCG tasks. To address this, we propose Adaptive Local-Aware Token Pruning (ALTP), a simple yet effective framework that accelerates GCG models by prioritizing local object information. ALTP introduces two key components: (1) Detail Density Capture (DDC), which uses superpixel segmentation to retain tokens in object-centric regions, preserving fine-grained details, and (2) Dynamic Density Formation (DDF), which dynamically allocates tokens based on information density, ensuring higher retention in semantically rich areas. Extensive experiments on the GranDf dataset demonstrate that ALTP significantly outperforms existing token pruning methods, such as FastV and PyramidDrop, on both GLaMM and OMG-LLaVA models. Notably, when applied to GLaMM, ALTP achieves a 90% reduction in visual tokens with a 4.9% improvement in AP50 and a 5.0% improvement in Recall compared to PyramidDrop. Similarly, on OMG-LLaVA, ALTP improves AP by 2.1% and mIOU by 3.0% at a 90% token reduction compared with PDrop.

Consistency-aware Self-Training for Iterative-based Stereo Matching

Iterative-based methods have become mainstream in stereo matching due to their high performance. However, these methods heavily rely on labeled data and face challenges with unlabeled real-world data. To this end, we propose a consistency-aware self-training framework for iterative-based stereo matching for the first time, leveraging real-world unlabeled data in a teacher-student manner. We first observe that regions with larger errors tend to exhibit more pronounced oscillation characteristics during model prediction.Based on this, we introduce a novel consistency-aware soft filtering module to evaluate the reliability of teacher-predicted pseudo-labels, which consists of a multi-resolution prediction consistency filter and an iterative prediction consistency filter to assess the prediction fluctuations of multiple resolutions and iterative optimization respectively. Further, we introduce a consistency-aware soft-weighted loss to adjust the weight of pseudo-labels accordingly, relieving the error accumulation and performance degradation problem due to incorrect pseudo-labels. Extensive experiments demonstrate that our method can improve the performance of various iterative-based stereo matching approaches in various scenarios. In particular, our method can achieve further enhancements over the current SOTA methods on several benchmark datasets.

FAVOR-Bench: A Comprehensive Benchmark for Fine-Grained Video Motion Understanding

Multimodal Large Language Models (MLLMs) have shown remarkable capabilities in video content understanding but still struggle with fine-grained motion comprehension. To comprehensively assess the motion understanding ability of existing MLLMs, we introduce FAVOR-Bench, comprising 1,776 videos with structured manual annotations of various motions. Our benchmark includes both close-ended and open-ended tasks. For close-ended evaluation, we carefully design 8,184 multiple-choice question-answer pairs spanning six distinct sub-tasks. For open-ended evaluation, we develop both a novel cost-efficient LLM-free and a GPT-assisted caption assessment method, where the former can enhance benchmarking interpretability and reproducibility. Comprehensive experiments with 21 state-of-the-art MLLMs reveal significant limitations in their ability to comprehend and describe detailed temporal dynamics in video motions. To alleviate this limitation, we further build FAVOR-Train, a dataset consisting of 17,152 videos with fine-grained motion annotations. The results of finetuning Qwen2.5-VL on FAVOR-Train yield consistent improvements on motion-related tasks of TVBench, MotionBench and our FAVOR-Bench. Comprehensive assessment results demonstrate that the proposed FAVOR-Bench and FAVOR-Train provide valuable tools to the community for developing more powerful video understanding models. Project page: \href{https://favor-bench.github.io/}{https://favor-bench.github.io/}.

TokenCarve: Information-Preserving Visual Token Compression in Multimodal Large Language Models

Multimodal Large Language Models (MLLMs) are becoming increasingly popular, while the high computational cost associated with multimodal data input, particularly from visual tokens, poses a significant challenge. Existing training-based token compression methods improve inference efficiency but require costly retraining, while training-free methods struggle to maintain performance when aggressively reducing token counts. In this study, we reveal that the performance degradation of MLLM closely correlates with the accelerated loss of information in the attention output matrix. This insight introduces a novel information-preserving perspective, making it possible to maintain performance even under extreme token compression. Based on this finding, we propose TokenCarve, a training-free, plug-and-play, two-stage token compression framework. The first stage employs an Information-Preservation-Guided Selection (IPGS) strategy to prune low-information tokens, while the second stage further leverages IPGS to guide token merging, minimizing information loss. Extensive experiments on 11 datasets and 2 model variants demonstrate the effectiveness of TokenCarve. It can even reduce the number of visual tokens to 22.2% of the original count, achieving a 1.23x speedup in inference, a 64% reduction in KV cache storage, and only a 1.54% drop in accuracy. Our code is available at https://github.com/ShawnTan86/TokenCarve.

Attention Reallocation: Towards Zero-cost and Controllable Hallucination Mitigation of MLLMs

Multi-Modal Large Language Models (MLLMs) stand out in various tasks but still struggle with hallucinations. While recent training-free mitigation methods mostly introduce additional inference overhead via retrospection strategy and contrastive decoding, we propose attention reallocation (AttnReal) to mitigate hallucinations with nearly zero extra cost. Our approach is motivated by the key observations that, MLLM's unreasonable attention distribution causes features to be dominated by historical output tokens, which further contributes to hallucinated responses because of the distribution gap between different token types. Based on the observations, AttnReal recycles excessive attention from output tokens and reallocates it to visual tokens, which reduces MLLM's reliance on language priors and ensures the decoding process depends more on the visual inputs. More interestingly, we find that, by controlling the intensity of AttnReal, we can achieve a wide-range trade-off between the response faithfulness and overall performance. Comprehensive results from different benchmarks validate the effectiveness of AttnReal across six open-source MLLMs and three decoding strategies.

Seeing Delta Parameters as JPEG Images: Data-Free Delta Compression with Discrete Cosine Transform

With transformer-based models and the pretrain-finetune paradigm becoming mainstream, the high storage and deployment costs of individual finetuned models on multiple tasks pose critical challenges. Delta compression attempts to lower the costs by reducing the redundancy of delta parameters (i.e., the difference between the finetuned and pre-trained model weights). However, existing methods usually face problems including data accessibility and training requirements. To tackle this issue, we introduce Delta-DCT, the first data-free delta compression method inspired by classic JPEG image compression, leveraging the Discrete Cosine Transform (DCT). We first (a) group delta parameters within a layer into patches. Then we (b) assess the importance of each patch and allocate them with different quantization bit-widths. Afterwards, we (c) convert these patches to the DCT domain and conduct quantization to each patch based on the allocated bit-width. The proposed Delta-DCT does not require any training or data calibration, while achieving performance comparable to or even surpassing original finetuned models under 1-bit equivalent delta compression ratios on different kinds of models including: (1) recently-released LLMs of different sizes from 7B to 13B, (2) relatively smaller language models including RoBERTa and T5 models, (3) variants of vision transformer models, and (4) multi-modal BEiT-3 models.

DeRS: Towards Extremely Efficient Upcycled Mixture-of-Experts Models

Upcycled Mixture-of-Experts (MoE) models have shown great potential in various tasks by converting the original Feed-Forward Network (FFN) layers in pre-trained dense models into MoE layers. However, these models still suffer from significant parameter inefficiency due to the introduction of multiple experts. In this work, we propose a novel DeRS (Decompose, Replace, and Synthesis) paradigm to overcome this shortcoming, which is motivated by our observations about the unique redundancy mechanisms of upcycled MoE experts. Specifically, DeRS decomposes the experts into one expert-shared base weight and multiple expert-specific delta weights, and subsequently represents these delta weights in lightweight forms. Our proposed DeRS paradigm can be applied to enhance parameter efficiency in two different scenarios, including: 1) DeRS Compression for inference stage, using sparsification or quantization to compress vanilla upcycled MoE models; and 2) DeRS Upcycling for training stage, employing lightweight sparse or low-rank matrixes to efficiently upcycle dense models into MoE models. Extensive experiments across three different tasks show that the proposed methods can achieve extreme parameter efficiency while maintaining the performance for both training and compression of upcycled MoE models.