Feiteng Fang, Dingwei Chen, Xiang Huang, Ting-En Lin, Yuchuan Wu, Xiong Liu, Jing Ye, Ziqiang Liu, Haonan Zhang, Liang Zhu, Hamid Alinejad-Rokny, Min Yang, Yongbin Li. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Haonan Zhang, Run Luo, Xiong Liu, Yuchuan Wu, Ting-En Lin, Pengpeng Zeng, Qiang Qu, Feiteng Fang, Min Yang, Lianli Gao, Jingkuan Song, Fei Huang, Yongbin Li. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Recent advancements in omnimodal learning have significantly improved understanding and generation across images, text, and speech, yet these developments remain predominantly confined to proprietary models. The lack of high-quality omnimodal datasets and the challenges of real-time emotional speech synthesis have notably hindered progress in open-source research. To address these limitations, we introduce OpenOmni, a two-stage training framework that integrates omnimodal alignment and speech generation to develop a state-of-the-art omnimodal large language model. In the alignment phase, a pretrained speech model undergoes further training on image-text tasks, enabling (near) zero-shot generalization from vision to speech, outperforming models trained on tri-modal datasets. In the speech generation phase, a lightweight decoder is trained on speech tasks with direct preference optimization, which enables real-time emotional speech synthesis with high fidelity. Extensive experiments demonstrate that OpenOmni surpasses state-of-the-art models across omnimodal, vision-language, and speech-language benchmarks. It achieves a 4-point absolute improvement on OmniBench over the leading open-source model VITA, despite using 5$\times$ fewer training examples and a smaller model size (7B vs. 7$\times$8B). Besides, OpenOmni achieves real-time speech generation with less than 1 second latency at non-autoregressive mode, reducing inference time by 5$\times$ compared to autoregressive methods, and improves emotion classification accuracy by 7.7\%. The codebase is available at https://github.com/RainBowLuoCS/OpenOmni.

Video Question Answering (VideoQA) is a challenging task in the vision-language field. Due to the time-consuming and labor-intensive labeling process of the question-answer pairs, fully supervised methods are no longer suitable for the current increasing demand for data. This has led to the rise of zero-shot VideoQA, and some works propose to adapt large language models (LLMs) to assist zero-shot learning. Despite recent progress, the inadequacy of LLMs in comprehending temporal information in videos and the neglect of temporal differences, e.g., the different dynamic changes between scenes or objects, remain insufficiently addressed by existing attempts in zero-shot VideoQA. In light of these challenges, a novel Temporal-guided Mixture-of-Experts Network (T-MoENet) for zero-shot video question answering is proposed in this paper. Specifically, we apply a temporal module to imbue language models with the capacity to perceive temporal information. Then a temporal-guided mixture-of-experts module is proposed to further learn the temporal differences presented in different videos. It enables the model to effectively improve the capacity of generalization. Our proposed method achieves state-of-the-art performance on multiple zero-shot VideoQA benchmarks, notably improving accuracy by 5.6% on TGIF-FrameQA and 2.3% on MSRVTT-QA while remaining competitive with other methods in the fully supervised setting. The codes and models developed in this study will be made publicly available athttps://github.com/qyx1121/T-MoENet.

Adapting large-scale image-text pre-training models, e.g., CLIP, to the video domain represents the current state-of-the-art for text-video retrieval. The primary approaches involve transferring text-video pairs to a common embedding space and leveraging cross-modal interactions on specific entities for semantic alignment. Though effective, these paradigms entail prohibitive computational costs, leading to inefficient retrieval. To address this, we propose a simple yet effective method, Global-Local Semantic Consistent Learning (GLSCL), which capitalizes on latent shared semantics across modalities for text-video retrieval. Specifically, we introduce a parameter-free global interaction module to explore coarse-grained alignment. Then, we devise a shared local interaction module that employs several learnable queries to capture latent semantic concepts for learning fine-grained alignment. Furthermore, an Inter-Consistency Loss (ICL) is devised to accomplish the concept alignment between the visual query and corresponding textual query, and an Intra-Diversity Loss (IDL) is developed to repulse the distribution within visual (textual) queries to generate more discriminative concepts. Extensive experiments on five widely used benchmarks (i.e., MSR-VTT, MSVD, DiDeMo, LSMDC, and ActivityNet) substantiate the superior effectiveness and efficiency of the proposed method. Remarkably, our method achieves comparable performance with SOTA as well as being nearly 220 times faster in terms of computational cost. Code is available at: https://github.com/zchoi/GLSCL.

Generating consecutive descriptions for videos, that is, video captioning, requires taking full advantage of visual representation along with the generation process. Existing video captioning methods focus on an exploration of spatial-temporal representations and their relationships to produce inferences. However, such methods only exploit the superficial association contained in a video itself without considering the intrinsic visual commonsense knowledge that exists in a video dataset, which may hinder their capabilities of knowledge cognitive to reason accurate descriptions. To address this problem, we propose a simple, yet effective method, called visual commonsense-aware representation network (VCRN), for video captioning. Specifically, we construct a Video Dictionary, a plug-and-play component, obtained by clustering all video features from the total dataset into multiple clustered centers without additional annotation. Each center implicitly represents a visual commonsense concept in a video domain, which is utilized in our proposed visual concept selection (VCS) component to obtain a video-related concept feature. Next, a concept-integrated generation (CIG) component is proposed to enhance caption generation. Extensive experiments on three public video captioning benchmarks: MSVD, MSR-VTT, and VATEX, demonstrate that our method achieves state-of-the-art performance, indicating the effectiveness of our method. In addition, our method is integrated into the existing method of video question answering (VideoQA) and improves this performance, which further demonstrates the generalization capability of our method. The source code has been released at https://github.com/zchoi/VCRN.

Recently, significant advancements have been made in supporting text-video retrieval by transferring large-scale image-text pre-training models through model adaptation, i.e., full fine-tuning, or prompt tuning, a parameter-efficient fine-tuning strategy. While full fine-tuning involves high computational costs, particularly with increasing model size, prompt tuning offers greater flexibility and efficiency by adjusting only a few learnable parameters. However, current prompt tuning methods rely on coarse visual and textual cues for text-video retrieval task, neglecting the domain-specific features when performing the adaptation. This approach may lead to sub-optimal performance due to the incorporation of irrelevant and indiscriminate knowledge. To address such an issue, we present a Multi-grained Prompt Tuning (MPT) for text-video retrieval, that designs a variety of specific prompts to effectively explore semantic interaction across different modalities with diverse granularity. Specifically, we devise a multi-grained video encoder that employs spatial, temporal, and global prompts to transfer the base-generic knowledge from the image-text pre-trained model while comprehensively excavating determinative video-specific characteristics. Meanwhile, we introduce a novel multi-grained text encoder aimed at capturing various levels of textual clues through the utilization of word and phrase prompts. Extensive experiments on four benchmark datasets, i.e., MSR-VTT, ActivityNet, DiDeMo, and LSMDC, demonstrate that MPT achieves outstanding performance, surpassing state-of-the-art methods with negligible computational cost. The codebase is publicly available at: https://github.com/zchoi/MPT.

The existing approaches to image captioning tend to adopt Transformer-based architectures with grid features, which represent the state-of-the-art. However, the strategies are prone to address the grid features with a fixed resolution, which often hampers the perception of entities with various scales. In addition, directly applying them may also result in spatial and fine-grained semantic information loss. To this end, we propose a simple yet effective method, named Spatial Pyramid Transformer (SPT). Specifically, it adopts several parameter-shared pyramid structures to perform semantic interactions across different grid resolutions. In each layer, we design a Spatial-aware Pseudo-supervised (SP) module, which aims to adaptively resort to disrupted spatial information among flatted grid features. Moreover, to maintain the model size and enhance semantics, we build a simple weighted residual connection termed as Scale-wise Reinforcement (SR) module to simultaneously explore both low- and high-level encoded features. Extensive experiments on the MS-COCO benchmark demonstrate that our method achieves new state-of-the-art performance without bringing excessive parameters compared with vanilla transformer. In addition, our method is extended to the video captioning task, which further proves the practicability of the proposed method. Code is available at https://github.com/zchoi/SPT.

Prompt tuning, an emerging parameter-efficient strategy, leverages the powerful knowledge of large-scale pre-trained image-text models (e.g., CLIP) to swiftly adapt to downstream tasks. Despite its effectiveness, adapting prompt tuning to text-video retrieval encounters two limitations: i) existing methods adopt two isolated prompt tokens to prompt two modal branches separately, making it challenging to learn a well-aligned unified representation, i.e., modality gap; ii) video encoders typically utilize a fixed pre-trained visual backbone, neglecting the incorporation of spatial-temporal information. To this end, we propose a simple yet effective method, named Unified Modality-aware Prompt Tuning (UMP), for text-video retrieval. Concretely, we first introduce a Unified Prompt Generation (UPG) module to dynamically produce modality-aware prompt tokens, enabling the perception of prior semantic information on both video and text inputs. These prompt tokens are simultaneously injected into two branches that can bridge the semantics gap between two modalities in a unified-adjusting manner. Then, we design a parameter-free Spatial-Temporal Shift (STS) module to facilitate both intra- and inter-communication among video tokens and prompt tokens in the spatial-temporal dimension. Notably, extensive experiments on four widely used benchmarks show that UMP achieves new state-of-the-art performance compared to existing prompt-tuning methods without bringing excessive parameters. Code is available at:https://github.com/zchoi/UMP_TVR.

Video captioning focuses on generating natural language descriptions according to the video content. Existing works mainly explore this multimodal learning with the paired source video and corresponding sentence, which have achieved competitive performances. Nonetheless, learning from video-description pair cannot capture implicit external knowledge, i.e., multiple visual context information and linguistic clues existing in the video-language dataset, which may limit the cognitive capability of the model to generate diverse descriptions. To this end, we propose a Memory-based Augmentation Network (MAN), in which a memory structure is designed to augment the current encoder-decoder framework by incorporating implicit external knowledge with a neural memory. Specifically, we first propose a visual memory for the encoder to store multiple visual contexts across videos in the dataset, which is utilized to obtain memory-augmented contextual features for the source video. In addition, a textual memory is introduced for the decoder to capture the external language clues across sentences in the dataset. It is adapted to capture memory-augmented language features in each time step. The proposed approach is able to capture comprehensive contextual understanding compared to the basic encoder-decoder framework, which is more compatible with the human cognitive process. Extensive experiments on three video captioning datasets including MSVD, MSR-VTT, and VATEX demonstrate the effectiveness of the proposed method.

In Video-Language (VL) learning tasks, a massive amount of text annotations are describing geometrical relationships of instances (e.g. 19.6% to 45.0% in MSVD, MSR-VTT, MSVD-QA and MSVRTT-QA), which often become the bottleneck of the current VL tasks (e.g. 60.8% vs. 98.2% CIDEr in MSVD for geometrical and non-geometrical annotations). Considering the rich spatial information of depth map, an intuitive way is to enrich the conventional 2D visual representations with depth information through current SOTA models, e.g. transformer. However, it is cumbersome to compute the self-attention on a long-range sequence and heterogeneous video-level representations with regard to computation cost and flexibility on various frame scales. To tackle this, we propose a hierarchical transformer, termed Depth-Aware Sparse Transformer (DAST). Specifically, to guarantee computational efficiency, a depth-aware sparse attention modular with linear computational complexity is designed for each transformer layer to learn depth-aware 2D representations. Furthermore, we design a hierarchical structure to maintain multi-scale temporal coherence across long-range dependencies. These qualities of DAST make it compatible with a broad range of video-language tasks, including video captioning (achieving MSVD 107.8%, MSR-VTT 52.5% for CIDEr), video question answering (MSVD-QA 44.1%, MSRVTT-QA 39.4%), and video-text matching (MSR-VTT 215.7 for SumR). Our code is available at https://github.com/zchoi/DAST

Transformer-based architectures with grid features represent the state-of-the-art in visual and language reasoning tasks, such as visual question answering and image-text matching. However, directly applying them to image captioning may result in spatial and fine-grained semantic information loss. Their applicability to image captioning is still largely under-explored. Towards this goal, we propose a simple yet effective method, Spatial- and Scale-aware Transformer (S2 Transformer) for image captioning. Specifically, we firstly propose a Spatial-aware Pseudo-supervised (SP) module, which resorts to feature clustering to help preserve spatial information for grid features. Next, to maintain the model size and produce superior results, we build a simple weighted residual connection, named Scale-wise Reinforcement (SR) module, to simultaneously explore both low- and high-level encoded features with rich semantics. Extensive experiments on the MSCOCO benchmark demonstrate that our method achieves new state-of-art performance without bringing excessive parameters compared with the vanilla transformer. The source code is available at https://github.com/zchoi/S2-Transformer

Cross-modal retrieval aims to build correspondence between multiple modalities by learning a common representation space. Typically, an image can match multiple texts semantically and vice versa, which significantly increases the difficulty of this task. To address this problem, probabilistic embedding is proposed to quantify these many-to-many relationships. However, existing datasets (e.g., MS-COCO) and metrics (e.g., Recall@K) cannot fully represent these diversity correspondences due to non-exhaustive annotations. Based on this observation, we utilize semantic correlation computed by CIDEr to find the potential correspondences. Then we present an effective metric, named Average Semantic Precision (ASP), which can measure the ranking precision of semantic correlation for retrieval sets. Additionally, we introduce a novel and concise objective, coined Differentiable ASP Approximation (DAA). Concretely, DAA can optimize ASP directly by making the ranking function of ASP differentiable through a sigmoid function. To verify the effectiveness of our approach, extensive experiments are conducted on MS-COCO, CUB Captions, and Flickr30K, which are commonly used in cross-modal retrieval. The results show that our approach obtains superior performance over the state-of-the-art approaches on all metrics. The code and trained models are released at https://github.com/leolee99/2022-NeurIPS-DAA.

Video Question Answering (VideoQA), which explores spatial-temporal visual information of videos given a linguistic query, has received unprecedented attention over recent years. One of the main challenges lies in locating relevant visual and linguistic information, and therefore various attention-based approaches are proposed. Despite the impressive progress, two aspects are not fully explored by current methods to get proper attention. Firstly, prior knowledge, which in the human cognitive process plays an important role in assisting the reasoning process of VideoQA, is not fully utilized. Secondly, structured visual information (e.g., object) instead of the raw video is underestimated. To address the above two issues, we propose a Prior Knowledge and Object-sensitive Learning (PKOL) by exploring the effect of prior knowledge and learning object-sensitive representations to boost the VideoQA task. Specifically, we first propose a Prior Knowledge Exploring (PKE) module that aims to acquire and integrate prior knowledge into a question feature for feature enriching, where an information retriever is constructed to retrieve related sentences as prior knowledge from the massive corpus. In addition, we propose an Object-sensitive Representation Learning (ORL) module to generate object-sensitive features by interacting object-level features with frame and clip-level features. Our proposed PKOL achieves consistent improvements on three competitive benchmarks (i.e., MSVD-QA, MSRVTT-QA, and TGIF-QA) and gains state-of-the-art performance. The source code is available at https://github.com/zchoi/PKOL.