Fengyun Rao

Yong Li, Furong Jia, Dacheng Yin et al.

Image geo-localization aims to determine where a photograph was taken, a task that often requires more than recognizing visible landmarks. Human experts typically solve it through an iterative workflow: they inspect informative regions, form location hypotheses, seek external evidence, and revise their judgments as new clues appear. Existing methods only partially capture this process: direct prediction methods bypass evidence acquisition altogether, while retrieval-augmented methods introduce external evidence but usually provide limited supervision on the intermediate decisions of where to search, how to query, and how to filter noisy results. We present REVERSE, a framework that reinforces the interplay between evidence search and verification to enable multi-turn agentic reasoning. REVERSE teaches three intermediate decisions: where to look, what to query, and what evidence to trust. To support this, we construct tool-grounded trajectories with annotated region selections, search observations, and geo-informative evidence labels, and introduce process rewards for visual grounding, query utility, and evidence discrimination. An offline search cache makes retrieval observations stable and reusable during reinforcement learning, enabling dense supervision over noisy search results. With a 4B model, REVERSE outperforms strong retrieval-augmented baselines and rivals substantially larger models on Im2GPS3k and YFCC4k. Code is available at https://github.com/yonglleee/REVERSE.

Xinli Yue, Jianhui Sun, Liangchao Yao et al.

The increasing popularity of short video platforms such as YouTube Shorts, TikTok, and Kwai has led to a surge in User-Generated Content (UGC), which presents significant challenges for the generalization performance of Video Quality Assessment (VQA) tasks. These challenges not only affect performance on test sets but also impact the ability to generalize across different datasets. While prior research has primarily focused on enhancing feature extractors, sampling methods, and network branches, it has largely overlooked the generalization capabilities of VQA tasks. In this work, we reevaluate the VQA task from a generalization standpoint. We begin by analyzing the weight loss landscape of VQA models, identifying a strong correlation between this landscape and the generalization gaps. We then investigate various techniques to regularize the weight loss landscape. Our results reveal that adversarial weight perturbations can effectively smooth this landscape, significantly improving the generalization performance, with cross-dataset generalization and fine-tuning performance enhanced by up to 1.8% and 3%, respectively. Through extensive experiments across various VQA methods and datasets, we validate the effectiveness of our approach. Furthermore, by leveraging our insights, we achieve state-of-the-art performance in Image Quality Assessment (IQA) tasks. Our code is available at https://github.com/XinliYue/VQA-Generalization.

Shannan Yan, Leqi Zheng, Keyu Lv et al.

We study the task of establishing object-level visual correspondence across different viewpoints in videos, focusing on the challenging egocentric-to-exocentric and exocentric-to-egocentric scenarios. We propose a simple yet effective framework based on conditional binary segmentation, where an object query mask is encoded into a latent representation to guide the localization of the corresponding object in a target video. To encourage robust, view-invariant representations, we introduce a cycle-consistency training objective: the predicted mask in the target view is projected back to the source view to reconstruct the original query mask. This bidirectional constraint provides a strong self-supervisory signal without requiring ground-truth annotations and enables test-time training (TTT) at inference. Experiments on the Ego-Exo4D and HANDAL-X benchmarks demonstrate the effectiveness of our optimization objective and TTT strategy, achieving state-of-the-art performance. The code is available at https://github.com/shannany0606/CCMP.

Tao Zhang, Ziqi Zhang, Zongyang Ma et al.

The ability to perform multi-modal multi-hop reasoning by iteratively integrating information across various modalities and external knowledge is critical for addressing complex real-world challenges. However, existing Multi-modal Large Language Models (MLLMs) are predominantly limited to single-step reasoning, as existing benchmarks lack the complexity needed to evaluate and drive multi-hop abilities. To bridge this gap, we introduce MMhops, a novel, large-scale benchmark designed to systematically evaluate and foster multi-modal multi-hop reasoning. MMhops dataset comprises two challenging task formats, Bridging and Comparison, which necessitate that models dynamically construct complex reasoning chains by integrating external knowledge. To tackle the challenges posed by MMhops, we propose MMhops-R1, a novel multi-modal Retrieval-Augmented Generation (mRAG) framework for dynamic reasoning. Our framework utilizes reinforcement learning to optimize the model for autonomously planning reasoning paths, formulating targeted queries, and synthesizing multi-level information. Comprehensive experiments demonstrate that MMhops-R1 significantly outperforms strong baselines on MMhops, highlighting that dynamic planning and multi-modal knowledge integration are crucial for complex reasoning. Moreover, MMhops-R1 demonstrates strong generalization to tasks requiring fixed-hop reasoning, underscoring the robustness of our dynamic planning approach. In conclusion, our work contributes a challenging new benchmark and a powerful baseline model, and we will release the associated code, data, and weights to catalyze future research in this critical area.

Feipeng Ma, Yizhou Zhou, Zheyu Zhang et al.

Recent advancements in Multimodal Large Language Models (MLLMs) have demonstrated satisfactory performance across various vision-language tasks. Current approaches for vision and language interaction fall into two categories: self-attention-based and cross-attention-based methods. However, both approaches present inherent limitations, forcing a trade-off between data and computational efficiency. To address this issue, we introduce the Data-$\textbf{E}$fficient and Compute-$\textbf{E}$fficient $\textbf{MLLM}$ ($\textbf{EE-MLLM}$). Specifically, we modify the original self-attention mechanism in MLLM to a composite attention mechanism. This mechanism has two key characteristics: 1) eliminating the computational overhead of self-attention among visual tokens to achieve $\textbf{compute efficiency}$, and 2) reusing the weights from each layer of LLM to facilitate effective vision-language modality alignment for $\textbf{data efficiency}$. As a result, EE-MLLM significantly outperforms Flamingo with limited training data, and reduces the prefilling time to 79 ms on an H800 GPU, compared to LLaVA's 277 ms. To further investigate the efficiency of EE-MLLM, we present a training-free variant named EE-MLLM-F, which reduces the computation cost of self-attention-based method without additional training. Experimental results demonstrate the effectiveness of EE-MLLM across a range of benchmarks, including general-purpose datasets like MMBench and SeedBench, as well as fine-grained tasks such as TextVQA and DocVQA.

Jian Yang, Dacheng Yin, Xiaoxuan He et al.

Recent progress in multimodal large language models (MLLMs) has highlighted the challenge of efficiently bridging pre-trained Vision-Language Models (VLMs) with Diffusion Models. While methods using a fixed number of learnable query tokens offer computational efficiency, they suffer from task generalization collapse, failing to adapt to new tasks that are distant from their pre-training tasks. To overcome this, we propose Noisy Query Tokens, which learn a distributed representation space between the VLM and Diffusion Model via end-to-end optimization, enhancing continual learning. Additionally, we introduce a VAE branch with linear projection to recover fine-grained image details. Experimental results confirm our approach mitigates generalization collapse and enables stable continual learning across diverse tasks.

Yongliang Wu, Xinting Hu, Yuyang Sun et al.

Video Large Language Models (Vid-LLMs) have made remarkable advancements in comprehending video content for QA dialogue. However, they struggle to extend this visual understanding to tasks requiring precise temporal localization, known as Video Temporal Grounding (VTG). To address this gap, we introduce Number-Prompt (NumPro), a novel method that empowers Vid-LLMs to bridge visual comprehension with temporal grounding by adding unique numerical identifiers to each video frame. Treating a video as a sequence of numbered frame images, NumPro transforms VTG into an intuitive process: flipping through manga panels in sequence. This allows Vid-LLMs to "read" event timelines, accurately linking visual content with corresponding temporal information. Our experiments demonstrate that NumPro significantly boosts VTG performance of top-tier Vid-LLMs without additional computational cost. Furthermore, fine-tuning on a NumPro-enhanced dataset defines a new state-of-the-art for VTG, surpassing previous top-performing methods by up to 6.9\% in mIoU for moment retrieval and 8.5\% in mAP for highlight detection. The code will be available at https://github.com/yongliang-wu/NumPro.

Zijie Xin, Jie Yang, Ruixiang Zhao et al.

Omni-modal large language models (om-LLMs) achieve unified audio-visual understanding by encoding video and audio into temporally aligned token sequences interleaved at the window level. However, processing these dense non-textual tokens throughout the LLM incurs substantial computational overhead. Although training-free token selection can reduce this cost, existing methods either focus on visual-only inputs or prune om-LLM tokens only before the LLM with fixed per-modality ratios, failing to capture how cross-modal token importance evolves across layers. To address this limitation, we first analyze the layer-wise token dependency of om-LLMs. We find that visual and audio dependencies follow a block-wise pattern and gradually weaken with depth, indicating that many late-layer non-textual tokens become redundant after cross-modal fusion. Motivated by this observation, we propose SEATS, a training-free, stage-adaptive token selection method for efficient om-LLM inference. Before the LLM, SEATS removes spatiotemporal redundancy via attention-weighted diversity selection. Inside the LLM, it progressively prunes tokens across blocks and dynamically allocates the retention budget from temporal windows to modalities using query relevance scores. In late layers, it removes all remaining non-textual tokens once cross-modal fusion is complete. Experiments on Qwen2.5-Omni and Qwen3-Omni demonstrate that SEATS effectively improves inference efficiency. Retaining only 10% of visual and audio tokens, it achieves a 9.3x FLOPs reduction and a 4.8x prefill speedup while preserving 96.3% of the original performance.

Tianrun Xu, Yue Sun, Qixun Wang et al.

Multimodal latent-space reasoning aims to replace explicit thinking with images by performing visual reasoning directly in a compact latent space. However, existing approaches largely rely on visual supervision and produce latent representations that lack sufficient semantic richness, limiting their ability to support diverse region-level reasoning tasks. In this work, we introduce Semantic-Enriched Latent Visual Reasoning (SLVR), a two-stage learning framework that enriches latent representations with attribute-level visual semantics and aligns them with diverse reasoning objectives. In the first stage, SLVR learns semantically enriched region-centric latents under fine-grained attribute supervision. In the second stage, we design Multi-query Group Relative Policy Optimization (M-GRPO) to align latent representations across multiple queries grounded in the same region. To support this framework, we construct SLV-Set, comprising approximately 400K region-level attribute annotations and 800K multi-query question answering samples, and introduce SV-QA, a benchmark that evaluates latent reasoning under semantic variation. Experiments demonstrate that SLVR improves the robustness and semantic consistency of latent visual reasoning compared to existing baselines.

Peixi Wu, Ke Mei, Feipeng Ma et al.

Multimodal Large Language Models (MLLMs) have emerged as a promising foundation for universal multimodal embeddings. Recent studies have shown that reasoning-driven generative multimodal embeddings can outperform discriminative embeddings on several embedding tasks. However, Chain-of-Thought (CoT) reasoning tends to generate redundant thinking steps and introduce semantic ambiguity in the summarized answers in broader retrieval scenarios. To address this limitation, we propose Rewrite-driven Multimodal Embedding (RIME), a unified framework that jointly optimizes generation and embedding through a retrieval-friendly rewrite. Meanwhile, we present the Cross-Mode Alignment (CMA) to bridge the generative and discriminative embedding spaces, enabling flexible mutual retrieval to trade off efficiency and accuracy. Based on this, we also introduce Refine Reinforcement Learning (Refine-RL) that treats discriminative embeddings as stable semantic anchors to guide the rewrite optimization. Extensive experiments on MMEB-V2, MRMR and UVRB demonstrate that RIME substantially outperforms prior generative embedding models while significantly reducing the length of thinking.

Ruixiang Zhao, Jie Yang, Zijie Xin et al.

Omni-proactive streaming video understanding, i.e., autonomously deciding when to speak and what to say from continuous audio-visual streams, is an emerging capability of omni-modal large language models. Existing benchmarks fall short in three key aspects: they rely primarily on visual signals, adopt polling or fixed-timestamp protocols instead of true proactive evaluation, and cover only a limited range of tasks, preventing reliable assessment and differentiation of omni-proactive streaming models. We present OmniPro, the first benchmark to jointly evaluate omni-modal perception, proactive responding, and diverse video understanding tasks. It comprises 2,700 human-verified samples spanning 9 sub-tasks and 3 cognitive levels, covering 6 basic video understanding capabilities. Notably, 84% of samples require audio signals (speech or non-speech), and each sample is annotated with modality-isolation labels to enable fine-grained multimodal analysis. We further introduce a dual-mode evaluation protocol: Probe mode assesses content understanding by querying the model before and after each ground-truth trigger, while Online mode evaluates full proactive ability by requiring models to autonomously decide when to respond in streaming input. Evaluating 11 representative models reveals three key findings: (1) audio provides consistent gains but with highly variable utilization across models, (2) performance degrades significantly over time, indicating limited long-horizon robustness, and (3) non-speech audio perception remains the weakest dimension.

Xinli Yue, Jianhui Sun, Han Kong et al.

In recent years, AI generative models have made remarkable progress across various domains, including text generation, image generation, and video generation. However, assessing the quality of text-to-video generation is still in its infancy, and existing evaluation frameworks fall short when compared to those for natural videos. Current video quality assessment (VQA) methods primarily focus on evaluating the overall quality of natural videos and fail to adequately account for the substantial quality discrepancies between frames in generated videos. To address this issue, we propose a novel loss function that combines mean absolute error with cross-entropy loss to mitigate inter-frame quality inconsistencies. Additionally, we introduce the innovative S2CNet technique to retain critical content, while leveraging adversarial training to enhance the model's generalization capabilities. Experimental results demonstrate that our method outperforms existing VQA techniques on the AIGC Video dataset, surpassing the previous state-of-the-art by 3.1% in terms of PLCC.

Jie Yang, Feipeng Ma, Zitian Wang et al.

Building on the success of text-based reasoning models like DeepSeek-R1, extending these capabilities to multimodal reasoning holds great promise. While recent works have attempted to adapt DeepSeek-R1-style reinforcement learning (RL) training paradigms to multimodal large language models (MLLM), focusing on domain-specific tasks like math and visual perception, a critical question remains: How can we achieve the general-purpose visual-language reasoning through RL? To address this challenge, we make three key efforts: (1) A novel Scalable Multimodal QA Synthesis pipeline that autonomously generates context-aware, reasoning-centric question-answer (QA) pairs directly from the given images. (2) The open-source WeThink dataset containing over 120K multimodal QA pairs with annotated reasoning paths, curated from 18 diverse dataset sources and covering various question domains. (3) A comprehensive exploration of RL on our dataset, incorporating a hybrid reward mechanism that combines rule-based verification with model-based assessment to optimize RL training efficiency across various task domains. Across 14 diverse MLLM benchmarks, we demonstrate that our WeThink dataset significantly enhances performance, from mathematical reasoning to diverse general multimodal tasks. Moreover, we show that our automated data pipeline can continuously increase data diversity to further improve model performance.

Shenghao Fu, Yukun Su, Fengyun Rao et al.

Aligning objects with corresponding textual descriptions is a fundamental challenge and a realistic requirement in vision-language understanding. While recent multimodal embedding models excel at global image-text alignment, they often struggle with fine-grained alignment between image regions and specific phrases. In this work, we present ObjEmbed, a novel MLLM embedding model that decomposes the input image into multiple regional embeddings, each corresponding to an individual object, along with global embeddings. It supports a wide range of visual understanding tasks like visual grounding, local image retrieval, and global image retrieval. ObjEmbed enjoys three key properties: (1) Object-Oriented Representation: It captures both semantic and spatial aspects of objects by generating two complementary embeddings for each region: an object embedding for semantic matching and an IoU embedding that predicts localization quality. The final object matching score combines semantic similarity with the predicted IoU, enabling more accurate retrieval. (2) Versatility: It seamlessly handles both region-level and image-level tasks. (3) Efficient Encoding: All objects in an image, along with the full image, are encoded in a single forward pass for high efficiency. Superior performance on 18 diverse benchmarks demonstrates its strong semantic discrimination.

Yi Yang, Xiaoxuan He, Hongkun Pan et al.

Large Language Models have demonstrated remarkable reasoning capability in complex textual tasks. However, multimodal reasoning, which requires integrating visual and textual information, remains a significant challenge. Existing visual-language models often struggle to effectively analyze and reason visual content, resulting in suboptimal performance on complex reasoning tasks. Moreover, the absence of comprehensive benchmarks hinders the accurate assessment of multimodal reasoning capabilities. In this paper, we introduce R1-Onevision, a multimodal reasoning model designed to bridge the gap between visual perception and deep reasoning. To achieve this, we propose a cross-modal reasoning pipeline that transforms images into formal textural representations, enabling precise language-based reasoning. Leveraging this pipeline, we construct the R1-Onevision dataset which provides detailed, step-by-step multimodal reasoning annotations across diverse domains. We further develop the R1-Onevision model through supervised fine-tuning and reinforcement learning to cultivate advanced reasoning and robust generalization abilities. To comprehensively evaluate multimodal reasoning performance across different grades, we introduce R1-Onevision-Bench, a benchmark aligned with human educational stages, covering exams from junior high school to university and beyond. Experimental results show that R1-Onevision achieves state-of-the-art performance, outperforming models such as GPT-4o and Qwen2.5-VL on multiple challenging multimodal reasoning benchmarks.

Shannan Yan, Jingchen Ni, Leqi Zheng et al.

Large language model (LLM) agents increasingly rely on external memory to support long-horizon interaction, personalized assistance, and multi-step reasoning. However, existing memory systems still face three core challenges: they often rely too heavily on semantic similarity, which can miss evidence crucial for user-centric understanding; they frequently store related experiences as isolated fragments, weakening temporal and causal coherence; and they typically use static memory granularities that do not adapt well to the requirements of different questions. We propose AdaMem, an adaptive user-centric memory framework for long-horizon dialogue agents. AdaMem organizes dialogue history into working, episodic, persona, and graph memories, enabling the system to preserve recent context, structured long-term experiences, stable user traits, and relation-aware connections within a unified framework. At inference time, AdaMem first resolves the target participant, then builds a question-conditioned retrieval route that combines semantic retrieval with relation-aware graph expansion only when needed, and finally produces the answer through a role-specialized pipeline for evidence synthesis and response generation. We evaluate AdaMem on the LoCoMo and PERSONAMEM benchmarks for long-horizon reasoning and user modeling. Experimental results show that AdaMem achieves state-of-the-art performance on both benchmarks. The code will be released upon acceptance.

Xiaoxuan He, Siming Fu, Wanli Li et al.

Aligning diffusion models with human preferences remains challenging, particularly when reward models are unavailable or impractical to obtain, and collecting large-scale preference datasets is prohibitively expensive. \textit{This raises a fundamental question: can we achieve effective alignment using only minimal human feedback, without auxiliary reward models, by unlocking the latent capabilities within diffusion models themselves?} In this paper, we propose \textbf{SAIL} (\textbf{S}elf-\textbf{A}mplified \textbf{I}terative \textbf{L}earning), a novel framework that enables diffusion models to act as their own teachers through iterative self-improvement. Starting from a minimal seed set of human-annotated preference pairs, SAIL operates in a closed-loop manner where the model progressively generates diverse samples, self-annotates preferences based on its evolving understanding, and refines itself using this self-augmented dataset. To ensure robust learning and prevent catastrophic forgetting, we introduce a ranked preference mixup strategy that carefully balances exploration with adherence to initial human priors. Extensive experiments demonstrate that SAIL consistently outperforms state-of-the-art methods across multiple benchmarks while using merely 6\% of the preference data required by existing approaches, revealing that diffusion models possess remarkable self-improvement capabilities that, when properly harnessed, can effectively replace both large-scale human annotation and external reward models.

Changli Tang, Qinfan Xiao, Ke Mei et al.

While embeddings from multimodal large language models (LLMs) excel as general-purpose representations, their application to dynamic modalities like audio and video remains underexplored. We introduce WAVE (\textbf{u}nified \& \textbf{v}ersatile \textbf{a}udio-\textbf{v}isual \textbf{e}mbeddings), the first LLM-based embedding that creates a unified representation space for text, audio, and video modalities. WAVE employs a novel hierarchical feature fusion strategy and a joint multi-modal, multi-task training approach to enable two key capabilities: any-to-any cross-modal retrieval and the generation of prompt-aware embeddings tailored to user instructions. Experimentally, WAVE sets a new state-of-the-art on the MMEB-v2 video benchmark and achieves superior results in audio and video-to-audio retrieval. Its prompt-aware nature also yields remarkable performance in multimodal question answering, significantly outperforming existing embedding models. Ablation studies validate our joint training strategy, demonstrating improved performance across all modalities. With a newly introduced benchmark for versatile audio-visual learning, WAVE opens up broad possibilities for cross-modal, any-to-any applications. Our code, checkpoints, and data will be released.

Yukun Su, Yiwen Cao, Jingliang Deng et al.

A large amount of User Generated Content (UGC) is uploaded to the Internet daily and displayed to people world-widely through the client side (e.g., mobile and PC). This requires the cropping algorithms to produce the aesthetic thumbnail within a specific aspect ratio on different devices. However, existing image cropping works mainly focus on landmark or landscape images, which fail to model the relations among the multi-objects with the complex background in UGC. Besides, previous methods merely consider the aesthetics of the cropped images while ignoring the content integrity, which is crucial for UGC cropping. In this paper, we propose a Spatial-Semantic Collaborative cropping network (S2CNet) for arbitrary user generated content accompanied by a new cropping benchmark. Specifically, we first mine the visual genes of the potential objects. Then, the suggested adaptive attention graph recasts this task as a procedure of information association over visual nodes. The underlying spatial and semantic relations are ultimately centralized to the crop candidate through differentiable message passing, which helps our network efficiently to preserve both the aesthetics and the content integrity. Extensive experiments on the proposed UGCrop5K and other public datasets demonstrate the superiority of our approach over state-of-the-art counterparts. Our project is available at https://github.com/suyukun666/S2CNet.

Zhenhua Liu, Feipeng Ma, Tianyi Wang et al.

With the development of multimedia technology, Video Copy Detection has been a crucial problem for social media platforms. Meta AI hold Video Similarity Challenge on CVPR 2023 to push the technology forward. In this report, we share our winner solutions on Matching Track. We propose a Similarity Alignment Model(SAM) for video copy segment matching. Our SAM exhibits superior performance compared to other competitors, with a 0.108 / 0.144 absolute improvement over the second-place competitor in Phase 1 / Phase 2. Code is available at https://github.com/FeipengMa6/VSC22-Submission/tree/main/VSC22-Matching-Track-1st.

Tianyi Wang, Feipeng Ma, Zhenhua Liu et al.

With the development of multimedia technology, Video Copy Detection has been a crucial problem for social media platforms. Meta AI hold Video Similarity Challenge on CVPR 2023 to push the technology forward. In this paper, we share our winner solutions on both tracks to help progress in this area. For Descriptor Track, we propose a dual-level detection method with Video Editing Detection (VED) and Frame Scenes Detection (FSD) to tackle the core challenges on Video Copy Detection. Experimental results demonstrate the effectiveness and efficiency of our proposed method. Code is available at https://github.com/FeipengMa6/VSC22-Submission.

Liang Xu, Xintao Lv, Yichao Yan et al.

The analysis of the ubiquitous human-human interactions is pivotal for understanding humans as social beings. Existing human-human interaction datasets typically suffer from inaccurate body motions, lack of hand gestures and fine-grained textual descriptions. To better perceive and generate human-human interactions, we propose Inter-X, a currently largest human-human interaction dataset with accurate body movements and diverse interaction patterns, together with detailed hand gestures. The dataset includes ~11K interaction sequences and more than 8.1M frames. We also equip Inter-X with versatile annotations of more than 34K fine-grained human part-level textual descriptions, semantic interaction categories, interaction order, and the relationship and personality of the subjects. Based on the elaborate annotations, we propose a unified benchmark composed of 4 categories of downstream tasks from both the perceptual and generative directions. Extensive experiments and comprehensive analysis show that Inter-X serves as a testbed for promoting the development of versatile human-human interaction analysis. Our dataset and benchmark will be publicly available for research purposes.

Liang Xu, Yizhou Zhou, Yichao Yan et al.

Humans constantly interact with their surrounding environments. Current human-centric generative models mainly focus on synthesizing humans plausibly interacting with static scenes and objects, while the dynamic human action-reaction synthesis for ubiquitous causal human-human interactions is less explored. Human-human interactions can be regarded as asymmetric with actors and reactors in atomic interaction periods. In this paper, we comprehensively analyze the asymmetric, dynamic, synchronous, and detailed nature of human-human interactions and propose the first multi-setting human action-reaction synthesis benchmark to generate human reactions conditioned on given human actions. To begin with, we propose to annotate the actor-reactor order of the interaction sequences for the NTU120, InterHuman, and Chi3D datasets. Based on them, a diffusion-based generative model with a Transformer decoder architecture called ReGenNet together with an explicit distance-based interaction loss is proposed to predict human reactions in an online manner, where the future states of actors are unavailable to reactors. Quantitative and qualitative results show that our method can generate instant and plausible human reactions compared to the baselines, and can generalize to unseen actor motions and viewpoint changes.

Cong Chen, Mingyu Liu, Chenchen Jing et al.

This paper aims to address the challenge of hallucinations in Multimodal Large Language Models (MLLMs) particularly for dense image captioning tasks. To tackle the challenge, we identify the current lack of a metric that finely measures the caption quality in concept level. We hereby introduce HalFscore, a novel metric built upon the language graph and is designed to evaluate both the accuracy and completeness of dense captions at a granular level. Additionally, we identify the root cause of hallucination as the model's over-reliance on its language prior. To address this, we propose PerturboLLaVA, which reduces the model's reliance on the language prior by incorporating adversarially perturbed text during training. This method enhances the model's focus on visual inputs, effectively reducing hallucinations and producing accurate, image-grounded descriptions without incurring additional computational overhead. PerturboLLaVA significantly improves the fidelity of generated captions, outperforming existing approaches in handling multimodal hallucinations and achieving improved performance across general multimodal benchmarks.

Jian Yang, Dacheng Yin, Yizhou Zhou et al.

Recent advancements in multi-modal large language models have propelled the development of joint probabilistic models capable of both image understanding and generation. However, we have identified that recent methods suffer from loss of image information during understanding task, due to either image discretization or diffusion denoising steps. To address this issue, we propose a novel Multi-Modal Auto-Regressive (MMAR) probabilistic modeling framework. Unlike discretization line of method, MMAR takes in continuous-valued image tokens to avoid information loss in an efficient way. Differing from diffusion-based approaches, we disentangle the diffusion process from auto-regressive backbone model by employing a light-weight diffusion head on top each auto-regressed image patch embedding. In this way, when the model transits from image generation to understanding through text generation, the backbone model's hidden representation of the image is not limited to the last denoising step. To successfully train our method, we also propose a theoretically proven technique that addresses the numerical stability issue and a training strategy that balances the generation and understanding task goals. Extensive evaluations on 18 image understanding benchmarks show that MMAR significantly outperforms most of the existing joint multi-modal models, surpassing the method that employs pre-trained CLIP vision encoder. Meanwhile, MMAR is able to generate high quality images. We also show that our method is scalable with larger data and model size.

Xiaoxuan He, Siming Fu, Yuke Zhao et al.

Recent flow matching models for text-to-image generation have achieved remarkable quality, yet their integration with reinforcement learning for human preference alignment remains suboptimal, hindering fine-grained reward-based optimization. We observe that the key impediment to effective GRPO training of flow models is the temporal uniformity assumption in existing approaches: sparse terminal rewards with uniform credit assignment fail to capture the varying criticality of decisions across generation timesteps, resulting in inefficient exploration and suboptimal convergence. To remedy this shortcoming, we introduce \textbf{TempFlow-GRPO} (Temporal Flow GRPO), a principled GRPO framework that captures and exploits the temporal structure inherent in flow-based generation. TempFlow-GRPO introduces three key innovations: (i) a trajectory branching mechanism that provides process rewards by concentrating stochasticity at designated branching points, enabling precise credit assignment without requiring specialized intermediate reward models; (ii) a noise-aware weighting scheme that modulates policy optimization according to the intrinsic exploration potential of each timestep, prioritizing learning during high-impact early stages while ensuring stable refinement in later phases; and (iii) a seed group strategy that controls for initialization effects to isolate exploration contributions. These innovations endow the model with temporally-aware optimization that respects the underlying generative dynamics, leading to state-of-the-art performance in human preference alignment and text-to-image benchmarks.

Yucheng Suo, Fan Ma, Linchao Zhu et al.

Multi-modal Large language models (MLLMs) show remarkable ability in video understanding. Nevertheless, understanding long videos remains challenging as the models can only process a finite number of frames in a single inference, potentially omitting crucial visual information. To address the challenge, we propose generating multiple predictions through visual context sampling, followed by a scoring mechanism to select the final prediction. Specifically, we devise a bin-wise sampling strategy that enables MLLMs to generate diverse answers based on various combinations of keyframes, thereby enriching the visual context. To determine the final prediction from the sampled answers, we employ a self-reward by linearly combining three scores: (1) a frequency score indicating the prevalence of each option, (2) a marginal confidence score reflecting the inter-intra sample certainty of MLLM predictions, and (3) a reasoning score for different question types, including clue-guided answering for global questions and temporal self-refocusing for local questions. The frequency score ensures robustness through majority correctness, the confidence-aligned score reflects prediction certainty, and the typed-reasoning score addresses cases with sparse key visual information using tailored strategies. Experiments show that this approach covers the correct answer for a high percentage of long video questions, on seven datasets show that our method improves the performance of three MLLMs.

Yunzhu Zhang, Yu Lu, Tianyi Wang et al.

Long-form video understanding poses a significant challenge for video large language models (VideoLLMs) due to prohibitively high computational and memory demands. In this paper, we propose FlexSelect, a flexible and efficient token selection strategy for processing long videos. FlexSelect identifies and retains the most semantically relevant content by leveraging cross-modal attention patterns from a reference transformer layer. It comprises two key components: (1) a training-free token ranking pipeline that leverages faithful cross-modal attention weights to estimate each video token's importance, and (2) a rank-supervised lightweight selector that is trained to replicate these rankings and filter redundant tokens. This generic approach can be seamlessly integrated into various VideoLLM architectures, such as LLaVA-Video, InternVL and Qwen-VL, serving as a plug-and-play module to extend their temporal context length. Empirically, FlexSelect delivers strong gains across multiple long-video benchmarks including VideoMME, MLVU, LongVB, and LVBench. Moreover, it achieves significant speed-ups (for example, up to 9 times on a LLaVA-Video-7B model), highlighting FlexSelect's promise for efficient long-form video understanding. Project page available at: https://yunzhuzhang0918.github.io/flex_select

Zitang Zhou, Ke Mei, Yu Lu et al.

This paper introduces HarmonySet, a comprehensive dataset designed to advance video-music understanding. HarmonySet consists of 48,328 diverse video-music pairs, annotated with detailed information on rhythmic synchronization, emotional alignment, thematic coherence, and cultural relevance. We propose a multi-step human-machine collaborative framework for efficient annotation, combining human insights with machine-generated descriptions to identify key transitions and assess alignment across multiple dimensions. Additionally, we introduce a novel evaluation framework with tasks and metrics to assess the multi-dimensional alignment of video and music, including rhythm, emotion, theme, and cultural context. Our extensive experiments demonstrate that HarmonySet, along with the proposed evaluation framework, significantly improves the ability of multimodal models to capture and analyze the intricate relationships between video and music.

Xinli Yue, JianHui Sun, Junda Lu et al.

With the rapid advancement of text-to-image (T2I) generation models, assessing the semantic alignment between generated images and text descriptions has become a significant research challenge. Current methods, including those based on Visual Question Answering (VQA), still struggle with fine-grained assessments and precise quantification of image-text alignment. This paper presents an improved evaluation method named Instruction-augmented Multimodal Alignment for Image-Text and Element Matching (iMatch), which evaluates image-text semantic alignment by fine-tuning multimodal large language models. We introduce four innovative augmentation strategies: First, the QAlign strategy creates a precise probabilistic mapping to convert discrete scores from multimodal large language models into continuous matching scores. Second, a validation set augmentation strategy uses pseudo-labels from model predictions to expand training data, boosting the model's generalization performance. Third, an element augmentation strategy integrates element category labels to refine the model's understanding of image-text matching. Fourth, an image augmentation strategy employs techniques like random lighting to increase the model's robustness. Additionally, we propose prompt type augmentation and score perturbation strategies to further enhance the accuracy of element assessments. Our experimental results show that the iMatch method significantly surpasses existing methods, confirming its effectiveness and practical value. Furthermore, our iMatch won first place in the CVPR NTIRE 2025 Text to Image Generation Model Quality Assessment - Track 1 Image-Text Alignment.

Zhixiang Wei, Guangting Wang, Xiaoxiao Ma et al.

Large-scale but noisy image-text pair data have paved the way for the success of Contrastive Language-Image Pretraining (CLIP). As the foundation vision encoder, CLIP in turn serves as the cornerstone for most large vision-language models (LVLMs). This interdependence naturally raises an interesting question: Can we reciprocally leverage LVLMs to enhance the quality of image-text pair data, thereby opening the possibility of a self-reinforcing cycle for continuous improvement? In this work, we take a significant step toward this vision by introducing an LVLM-driven data refinement pipeline. Our framework leverages LVLMs to process images and their raw alt-text, generating four complementary textual formulas: long positive descriptions, long negative descriptions, short positive tags, and short negative tags. Applying this pipeline to the curated DFN-Large dataset yields VLM-150M, a refined dataset enriched with multi-grained annotations. Based on this dataset, we further propose a training paradigm that extends conventional contrastive learning by incorporating negative descriptions and short tags as additional supervised signals. The resulting model, namely HQ-CLIP, demonstrates remarkable improvements across diverse benchmarks. Within a comparable training data scale, our approach achieves state-of-the-art performance in zero-shot classification, cross-modal retrieval, and fine-grained visual understanding tasks. In retrieval benchmarks, HQ-CLIP even surpasses standard CLIP models trained on the DFN-2B dataset, which contains 10$\times$ more training data than ours. All code, data, and models are available at https://zxwei.site/hqclip.

Shenghao Fu, Yukun Su, Fengyun Rao et al.

Open-vocabulary object detection aims to detect arbitrary classes via text prompts. Methods without cross-modal fusion layers (non-fusion) offer faster inference by treating recognition as a retrieval problem, \ie, matching regions to text queries in a shared embedding space. In this work, we fully explore this retrieval philosophy and demonstrate its unique advantages in efficiency and versatility through a model family named WeDetect: (1) State-of-the-art performance. WeDetect is a real-time detector with a dual-tower architecture. We show that, with well-curated data and full training, the non-fusion WeDetect surpasses other fusion models and establishes a strong open-vocabulary foundation. (2) Fast backtrack of historical data. WeDetect-Uni is a universal proposal generator based on WeDetect. We freeze the entire detector and only finetune an objectness prompt to retrieve generic object proposals across categories. Importantly, the proposal embeddings are class-specific and enable a new application, object retrieval, supporting retrieval objects in historical data. (3) Integration with LMMs for referring expression comprehension (REC). We further propose WeDetect-Ref, an LMM-based object classifier to handle complex referring expressions, which retrieves target objects from the proposal list extracted by WeDetect-Uni. It discards next-token prediction and classifies objects in a single forward pass. Together, the WeDetect family unifies detection, proposal generation, object retrieval, and REC under a coherent retrieval framework, achieving state-of-the-art performance across 15 benchmarks with high inference efficiency.

Zitian Wang, Yue Liao, Kang Rong et al.

Preference alignment has emerged as an effective strategy to enhance the performance of Multimodal Large Language Models (MLLMs) following supervised fine-tuning. While existing preference alignment methods predominantly target hallucination factors, they overlook the factors essential for multi-modal comprehension capabilities, often narrowing their improvements on hallucination mitigation. To bridge this gap, we propose Instruction-oriented Preference Alignment (IPA), a scalable framework designed to automatically construct alignment preferences grounded in instruction fulfillment efficacy. Our method involves an automated preference construction coupled with a dedicated verification process that identifies instruction-oriented factors, avoiding significant variability in response representations. Additionally, IPA incorporates a progressive preference collection pipeline, further recalling challenging samples through model self-evolution and reference-guided refinement. Experiments conducted on Qwen2VL-7B demonstrate IPA's effectiveness across multiple benchmarks, including hallucination evaluation, visual question answering, and text understanding tasks, highlighting its capability to enhance general comprehension.

Feipeng Ma, Yizhou Zhou, Fengyun Rao et al.

Image captioning requires numerous annotated image-text pairs, resulting in substantial annotation costs. Recently, large models (e.g. diffusion models and large language models) have excelled in producing high-quality images and text. This potential can be harnessed to create synthetic image-text pairs for training captioning models. Synthetic data can improve cost and time efficiency in data collection, allow for customization to specific domains, bootstrap generalization capability for zero-shot performance, and circumvent privacy concerns associated with real-world data. However, existing methods struggle to attain satisfactory performance solely through synthetic data. We identify the issue as generated images from simple descriptions mostly capture a solitary perspective with limited context, failing to align with the intricate scenes prevalent in real-world imagery. To tackle this, we present an innovative pipeline that introduces multi-context data generation. Beginning with an initial text corpus, our approach employs a large language model to extract multiple sentences portraying the same scene from diverse viewpoints. These sentences are then condensed into a single sentence with multiple contexts. Subsequently, we generate intricate images using the condensed captions through diffusion models. Our model is exclusively trained on synthetic image-text pairs crafted through this process. The effectiveness of our pipeline is validated through experimental results in both the in-domain and cross-domain settings, where it achieves state-of-the-art performance on well-known datasets such as MSCOCO, Flickr30k, and NoCaps.

Lu Qi, Jason Kuen, Zhe Lin et al.

To improve instance-level detection/segmentation performance, existing self-supervised and semi-supervised methods extract either task-unrelated or task-specific training signals from unlabeled data. We show that these two approaches, at the two extreme ends of the task-specificity spectrum, are suboptimal for the task performance. Utilizing too little task-specific training signals causes underfitting to the ground-truth labels of downstream tasks, while the opposite causes overfitting to the ground-truth labels. To this end, we propose a novel Class-Agnostic Semi-Supervised Learning (CA-SSL) framework to achieve a more favorable task-specificity balance in extracting training signals from unlabeled data. CA-SSL has three training stages that act on either ground-truth labels (labeled data) or pseudo labels (unlabeled data). This decoupling strategy avoids the complicated scheme in traditional SSL methods that balances the contributions from both data types. Especially, we introduce a warmup training stage to achieve a more optimal balance in task specificity by ignoring class information in the pseudo labels, while preserving localization training signals. As a result, our warmup model can better avoid underfitting/overfitting when fine-tuned on the ground-truth labels in detection and segmentation tasks. Using 3.6M unlabeled data, we achieve a significant performance gain of 4.7% over ImageNet-pretrained baseline on FCOS object detection. In addition, our warmup model demonstrates excellent transferability to other detection and segmentation frameworks.

Mingkang Tang, Zhanyu Wang, Zhenhua Liu et al.

Video captioning is a challenging task since it requires generating sentences describing various diverse and complex videos. Existing video captioning models lack adequate visual representation due to the neglect of the existence of gaps between videos and texts. To bridge this gap, in this paper, we propose a CLIP4Caption framework that improves video captioning based on a CLIP-enhanced video-text matching network (VTM). This framework is taking full advantage of the information from both vision and language and enforcing the model to learn strongly text-correlated video features for text generation. Besides, unlike most existing models using LSTM or GRU as the sentence decoder, we adopt a Transformer structured decoder network to effectively learn the long-range visual and language dependency. Additionally, we introduce a novel ensemble strategy for captioning tasks. Experimental results demonstrate the effectiveness of our method on two datasets: 1) on MSR-VTT dataset, our method achieved a new state-of-the-art result with a significant gain of up to 10% in CIDEr; 2) on the private test data, our method ranking 2nd place in the ACM MM multimedia grand challenge 2021: Pre-training for Video Understanding Challenge. It is noted that our model is only trained on the MSR-VTT dataset.

Mingkang Tang, Zhanyu Wang, Zhaoyang Zeng et al.

This report describes our solution to the VALUE Challenge 2021 in the captioning task. Our solution, named CLIP4Caption++, is built on X-Linear/X-Transformer, which is an advanced model with encoder-decoder architecture. We make the following improvements on the proposed CLIP4Caption++: We employ an advanced encoder-decoder model architecture X-Transformer as our main framework and make the following improvements: 1) we utilize three strong pre-trained CLIP models to extract the text-related appearance visual features. 2) we adopt the TSN sampling strategy for data enhancement. 3) we involve the video subtitle information to provide richer semantic information. 3) we introduce the subtitle information, which fuses with the visual features as guidance. 4) we design word-level and sentence-level ensemble strategies. Our proposed method achieves 86.5, 148.4, 64.5 CIDEr scores on VATEX, YC2C, and TVC datasets, respectively, which shows the superior performance of our proposed CLIP4Caption++ on all three datasets.