Zhenbang Sun

Minghao Xu, Yuanfan Guo, Xuanyu Zhu et al.

Learning self-supervised image representations has been broadly studied to boost various visual understanding tasks. Existing methods typically learn a single level of image semantics like pairwise semantic similarity or image clustering patterns. However, these methods can hardly capture multiple levels of semantic information that naturally exists in an image dataset, e.g., the semantic hierarchy of "Persian cat to cat to mammal" encoded in an image database for species. It is thus unknown whether an arbitrary image self-supervised learning (SSL) approach can benefit from learning such hierarchical semantics. To answer this question, we propose a general framework for Hierarchical Image Representation Learning (HIRL). This framework aims to learn multiple semantic representations for each image, and these representations are structured to encode image semantics from fine-grained to coarse-grained. Based on a probabilistic factorization, HIRL learns the most fine-grained semantics by an off-the-shelf image SSL approach and learns multiple coarse-grained semantics by a novel semantic path discrimination scheme. We adopt six representative image SSL methods as baselines and study how they perform under HIRL. By rigorous fair comparison, performance gain is observed on all the six methods for diverse downstream tasks, which, for the first time, verifies the general effectiveness of learning hierarchical image semantics. All source code and model weights are available at https://github.com/hirl-team/HIRL

Minyi Zhao, Jie Wang, Zhaoyang Li et al.

Recent studies have shown that Vision Language Large Models (VLLMs) may output content not relevant to the input images. This problem, called the hallucination phenomenon, undoubtedly degrades VLLM performance. Therefore, various anti-hallucination techniques have been proposed to make model output more reasonable and accurate. Despite their successes, from extensive tests we found that augmenting the prompt (e.g. word appending, rewriting, and spell error etc.) may change model output and make the output hallucinate again. To cure this drawback, we propose a new instruct-tuning framework called Prompt Augmentation and Caption Utilization (PACU) to boost VLLM's generation ability under the augmented prompt scenario. Concretely, on the one hand, PACU exploits existing LLMs to augment and evaluate diverse prompts automatically. The resulting high-quality prompts are utilized to enhance VLLM's ability to process different prompts. On the other hand, PACU exploits image captions to jointly work with image features as well as the prompts for response generation. When the visual feature is inaccurate, LLM can capture useful information from the image captions for response generation. Extensive experiments on hallucination evaluation and prompt-augmented datasets demonstrate that our PACU method can work well with existing schemes to effectively boost VLLM model performance. Code is available in https://github.com/zhaominyiz/PACU.

Hongyu Hu, Jiyuan Zhang, Minyi Zhao et al.

Nowadays, the research on Large Vision-Language Models (LVLMs) has been significantly promoted thanks to the success of Large Language Models (LLM). Nevertheless, these Vision-Language Models (VLMs) are suffering from the drawback of hallucination -- due to insufficient understanding of vision and language modalities, VLMs may generate incorrect perception information when doing downstream applications, for example, captioning a non-existent entity. To address the hallucination phenomenon, on the one hand, we introduce a Contrastive Instruction Evaluation Method (CIEM), which is an automatic pipeline that leverages an annotated image-text dataset coupled with an LLM to generate factual/contrastive question-answer pairs for the evaluation of the hallucination of VLMs. On the other hand, based on CIEM, we further propose a new instruction tuning method called CIT (the abbreviation of Contrastive Instruction Tuning) to alleviate the hallucination of VLMs by automatically producing high-quality factual/contrastive question-answer pairs and corresponding justifications for model tuning. Through extensive experiments on CIEM and CIT, we pinpoint the hallucination issues commonly present in existing VLMs, the disability of the current instruction-tuning dataset to handle the hallucination phenomenon and the superiority of CIT-tuned VLMs over both CIEM and public datasets.

Yangming Shi, Shixiang Zhu, Tao Shen et al.

We present Mamoda2.5, a unified AR-Diffusion framework that seamlessly integrates multimodal understanding and generation within a single architecture. To efficiently enhance the model's generation capability, we equip the Diffusion Transformer backbone with a fine-grained Mixture-of-Experts (MoE) design (128 experts, Top-8 routing), yielding a 25B-parameter model that activates only 3B parameters, significantly reducing training costs while scaling up the model capacity. Mamoda2.5 achieves top-tier generation performance on VBench 2.0 and sets a new record in video editing quality, surpassing evaluated open-source models and matching the performance of current top-tier proprietary models, including the Kling O1 on OpenVE-Bench. Furthermore, we introduce a joint few-step distillation and reinforcement learning framework that compresses the 30-step editing model into a 4-step model and greatly accelerates model inference. Compared to open-source baselines, Mamoda2.5 achieves up to $95.9\times$ faster video editing inference. In real-world applications, Mamoda2.5 has been successfully deployed for content moderation and creative restoration tasks in advertising scenarios, achieving a 98% success rate in internal advertising video editing scenario.

Hongyu Hu, Tiancheng Lin, Jie Wang et al.

Large-scale vision-language models (VLMs), e.g., CLIP, learn broad visual concepts from tedious training data, showing superb generalization ability. Amount of prompt learning methods have been proposed to efficiently adapt the VLMs to downstream tasks with only a few training samples. We introduce a novel method to improve the prompt learning of vision-language models by incorporating pre-trained large language models (LLMs), called Dual-Aligned Prompt Tuning (DuAl-PT). Learnable prompts, like CoOp, implicitly model the context through end-to-end training, which are difficult to control and interpret. While explicit context descriptions generated by LLMs, like GPT-3, can be directly used for zero-shot classification, such prompts are overly relying on LLMs and still underexplored in few-shot domains. With DuAl-PT, we propose to learn more context-aware prompts, benefiting from both explicit and implicit context modeling. To achieve this, we introduce a pre-trained LLM to generate context descriptions, and we encourage the prompts to learn from the LLM's knowledge by alignment, as well as the alignment between prompts and local image features. Empirically, DuAl-PT achieves superior performance on 11 downstream datasets on few-shot recognition and base-to-new generalization. Hopefully, DuAl-PT can serve as a strong baseline. Code will be available.

Lequan Lin, Dai Shi, Andi Han et al.

Supervised learning relies on high-quality labeled data, but obtaining such data through human annotation is both expensive and time-consuming. Recent work explores using large language models (LLMs) for annotation, but LLM-generated labels still fall short of human-level quality. To address this problem, we propose the Annotation with Critical Thinking (ACT) data pipeline, where LLMs serve not only as annotators but also as judges to critically identify potential errors. Human effort is then directed towards reviewing only the most "suspicious" cases, significantly improving the human annotation efficiency. Our major contributions are as follows: (1) ACT is applicable to a wide range of domains, including natural language processing (NLP), computer vision (CV), and multimodal understanding, by leveraging multimodal-LLMs (MLLMs). (2) Through empirical studies, we derive 7 insights on how to enhance annotation quality while efficiently reducing the human cost, and then translate these findings into user-friendly guidelines. (3) We theoretically analyze how to modify the loss function so that models trained on ACT data achieve similar performance to those trained on fully human-annotated data. Our experiments show that the performance gap can be reduced to less than 2% on most benchmark datasets while saving up to 90% of human costs.

Feng Chen, Chenhui Gou, Jing Liu et al.

As multimodal large language models (MLLMs) advance rapidly, rigorous evaluation has become essential, providing further guidance for their development. In this work, we focus on a unified and robust evaluation of \textbf{vision perception} abilities, the foundational skill of MLLMs. We find that existing perception benchmarks, each focusing on different question types, domains, and evaluation metrics, introduce significant evaluation variance, complicating comprehensive assessments of perception abilities when relying on any single benchmark. To address this, we introduce \textbf{AbilityLens}, a unified benchmark designed to evaluate MLLMs in six key perception abilities (ranging from counting, OCR, to understanding structural data), focusing on both accuracy and stability, with each ability encompassing diverse types of questions, domains, and metrics. With the assistance of AbilityLens, we: (1) identify the strengths and weaknesses of current main-stream MLLMs, highlighting stability patterns and revealing a notable performance gap between state-of-the-art open-source and closed-source models; (2) uncover interesting ability conflict and early convergence phenomena during MLLM training; (3) reveal the primary reason of ability conflict is data mixing ratio and LLM model size; and (4) discuss the effectiveness of some straightforward strategies \eg, fine-tuning and model merging, to solve the ability conflict. The benchmark and online leaderboard is released in https://github.com/Chenfeng1271/AbilityLens.

Feng Chen, Yefei He, Shaoxuan He et al.

Existing sparse attention methods primarily target inference-time acceleration by selecting critical tokens under predefined sparsity patterns. However, they often fail to bridge the training-inference gap and lack the capacity for fine-grained token selection across multiple dimensions such as queries, key-values (KV), and heads, leading to suboptimal performance and limited acceleration gains. In this paper, we introduce OmniSparse, a training-aware fine-grained sparse attention framework for long-video MLLMs, which operates in both training and inference with dynamic token budget allocation. Specifically, OmniSparse contains three adaptive and complementary mechanisms: (1) query selection via lazy-active classification, retaining active queries that capture broad semantic similarity while discarding most lazy ones that focus on limited local context and exhibit high functional redundancy; (2) KV selection with head-level dynamic budget allocation, where a shared budget is determined based on the flattest head and applied uniformly across all heads to ensure attention recall; and (3) KV cache slimming to reduce head-level redundancy by selectively fetching visual KV cache according to the head-level decoding query pattern. Experimental results show that OmniSparse matches the performance of full attention while achieving up to 2.7x speedup during prefill and 2.4x memory reduction during decoding.

Yuanfan Guo, Minghao Xu, Jiawen Li et al.

Hierarchical semantic structures naturally exist in an image dataset, in which several semantically relevant image clusters can be further integrated into a larger cluster with coarser-grained semantics. Capturing such structures with image representations can greatly benefit the semantic understanding on various downstream tasks. Existing contrastive representation learning methods lack such an important model capability. In addition, the negative pairs used in these methods are not guaranteed to be semantically distinct, which could further hamper the structural correctness of learned image representations. To tackle these limitations, we propose a novel contrastive learning framework called Hierarchical Contrastive Selective Coding (HCSC). In this framework, a set of hierarchical prototypes are constructed and also dynamically updated to represent the hierarchical semantic structures underlying the data in the latent space. To make image representations better fit such semantic structures, we employ and further improve conventional instance-wise and prototypical contrastive learning via an elaborate pair selection scheme. This scheme seeks to select more diverse positive pairs with similar semantics and more precise negative pairs with truly distinct semantics. On extensive downstream tasks, we verify the superior performance of HCSC over state-of-the-art contrastive methods, and the effectiveness of major model components is proved by plentiful analytical studies. We build a comprehensive model zoo in Sec. D. Our source code and model weights are available at https://github.com/gyfastas/HCSC

Minghao Xu, Hang Wang, Bingbing Ni et al.

Autonomous highlight detection is crucial for enhancing the efficiency of video browsing on social media platforms. To attain this goal in a data-driven way, one may often face the situation where highlight annotations are not available on the target video category used in practice, while the supervision on another video category (named as source video category) is achievable. In such a situation, one can derive an effective highlight detector on target video category by transferring the highlight knowledge acquired from source video category to the target one. We call this problem cross-category video highlight detection, which has been rarely studied in previous works. For tackling such practical problem, we propose a Dual-Learner-based Video Highlight Detection (DL-VHD) framework. Under this framework, we first design a Set-based Learning module (SL-module) to improve the conventional pair-based learning by assessing the highlight extent of a video segment under a broader context. Based on such learning manner, we introduce two different learners to acquire the basic distinction of target category videos and the characteristics of highlight moments on source video category, respectively. These two types of highlight knowledge are further consolidated via knowledge distillation. Extensive experiments on three benchmark datasets demonstrate the superiority of the proposed SL-module, and the DL-VHD method outperforms five typical Unsupervised Domain Adaptation (UDA) algorithms on various cross-category highlight detection tasks. Our code is available at https://github.com/ChrisAllenMing/Cross_Category_Video_Highlight .

Yun Li, Zhe Liu, Yajing Kong et al.

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.

Tao Shen, Xin Wan, Taicai Chen et al.

Unified multimodal models aim to integrate understanding and generation within a single framework, yet bridging the gap between discrete semantic reasoning and high-fidelity visual synthesis remains challenging. We present MammothModa2 (Mammoth2), a unified autoregressive-diffusion (AR-Diffusion) framework designed to effectively couple autoregressive semantic planning with diffusion-based generation. Mammoth2 adopts a serial design: an AR path equipped with generation experts performs global semantic modeling over discrete tokens, while a single-stream Diffusion Transformer (DiT) decoder handles high-fidelity image synthesis. A carefully designed AR-Diffusion feature alignment module combines multi-layer feature aggregation, unified condition encoding, and in-context conditioning to stably align AR's representations with the diffusion decoder's continuous latents. Mammoth2 is trained end-to-end with joint Next-Token Prediction and Flow Matching objectives, followed by supervised fine-tuning and reinforcement learning over both generation and editing. With roughly 60M supervised generation samples and no reliance on pre-trained generators, Mammoth2 delivers strong text-to-image and instruction-based editing performance on public benchmarks, achieving 0.87 on GenEval, 87.2 on DPGBench, and 4.06 on ImgEdit, while remaining competitive with understanding-only backbones (e.g., Qwen3-VL-8B) on multimodal understanding tasks. These results suggest that a carefully coupled AR-Diffusion architecture can provide high-fidelity generation and editing while maintaining strong multimodal comprehension within a single, parameter- and data-efficient model.

Ziye Chen, Yiqun Duan, Riheng Zhu et al.

Personalized multiple clustering aims to generate diverse partitions of a dataset based on different user-specific aspects, rather than a single clustering. It has recently drawn research interest for accommodating varying user preferences. Recent approaches primarily use CLIP embeddings with proxy learning to extract representations biased toward user clustering preferences. However, CLIP primarily focuses on coarse image-text alignment, lacking a deep contextual understanding of user interests. To overcome these limitations, we propose an agent-centric personalized clustering framework that leverages multi-modal large language models (MLLMs) as agents to comprehensively traverse a relational graph to search for clusters based on user interests. Due to the advanced reasoning mechanism of MLLMs, the obtained clusters align more closely with user-defined criteria than those obtained from CLIP-based representations. To reduce computational overhead, we shorten the agents' traversal path by constructing a relational graph using user-interest-biased embeddings extracted by MLLMs. A large number of weakly connected edges can be filtered out based on embedding similarity, facilitating an efficient traversal search for agents. Experimental results show that the proposed method achieves NMI scores of 0.9667 and 0.9481 on the Card Order and Card Suits benchmarks, respectively, largely improving the SOTA model by over 140%.