Representation learning is fundamental to modern machine learning, powering applications such as text retrieval and multimodal understanding. However, learning robust and generalizable representations remains challenging. While prior work has demonstrated that active noise injection, a form of data augmentation, can enhance encoding performance, most existing methods rely on heuristic or static noise, overlooking the dynamic nature of feature distributions during training. In this work, we systematically study the role of noise in representation learning from both gradient-based and feature distribution perspectives, using InfoNCE loss as a representative example. Focusing on multimodal representation learning, we propose FANoise, a novel feature-adaptive noise injection strategy. By leveraging the dynamics of contrastive learning, FANoise effectively mitigates the negative impacts of noise while preserving its benefits. Under this theoretically grounded framework, comprehensive experiments demonstrate that FANoise consistently improves overall performance on multimodal tasks across various base VLM models.

Large-scale multimodal contrastive learning has recently achieved impressive success in learning rich and transferable representations, yet it remains fundamentally limited by the uniform treatment of feature dimensions and the neglect of the intrinsic spectral structure of the learned features. Empirical evidence indicates that high-dimensional embeddings tend to collapse into narrow cones, concentrating task-relevant semantics in a small subspace, while the majority of dimensions remain occupied by noise and spurious correlations. Such spectral imbalance and entanglement undermine model generalization. We propose Spectral Disentanglement and Enhancement (SDE), a novel framework that bridges the gap between the geometry of the embedded spaces and their spectral properties. Our approach leverages singular value decomposition to adaptively partition feature dimensions into strong signals that capture task-critical semantics, weak signals that reflect ancillary correlations, and noise representing irrelevant perturbations. A curriculum-based spectral enhancement strategy is then applied, selectively amplifying informative components with theoretical guarantees on training stability. Building upon the enhanced features, we further introduce a dual-domain contrastive loss that jointly optimizes alignment in both the feature and spectral spaces, effectively integrating spectral regularization into the training process and encouraging richer, more robust representations. Extensive experiments on large-scale multimodal benchmarks demonstrate that SDE consistently improves representation robustness and generalization, outperforming state-of-the-art methods. SDE integrates seamlessly with existing contrastive pipelines, offering an effective solution for multimodal representation learning.

Guoqiang Gong, Jiaxing Wang, Jin Xu, Deping Xiang, Zicheng Zhang, Leqi Shen, Yifeng Zhang, JunhuaShu JunhuaShu, ZhaolongXing ZhaolongXing, Zhen Chen, Pengzhang Liu, Ke Zhang. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

The parameter-efficient adaptation of the image-text pre-training model CLIP for video-text retrieval is a prominent area of research. While CLIP is focused on image-level vision-language matching, video-text retrieval demands comprehensive understanding at the video level. Three key discrepancies emerge in the transfer from image-level to video-level: vision, language, and alignment. However, existing methods mainly focus on vision while neglecting language and alignment. In this paper, we propose Discrepancy Reduction in Vision, Language, and Alignment (DiscoVLA), which simultaneously mitigates all three discrepancies. Specifically, we introduce Image-Video Features Fusion to integrate image-level and video-level features, effectively tackling both vision and language discrepancies. Additionally, we generate pseudo image captions to learn fine-grained image-level alignment. To mitigate alignment discrepancies, we propose Image-To-Video Alignment Distillation, which leverages image-level alignment knowledge to enhance video-level alignment. Extensive experiments demonstrate the superiority of our DiscoVLA. In particular, on MSRVTT with CLIP (ViT-B/16), DiscoVLA outperforms previous methods by 2.2% R@1 and 7.5% R@sum. The code is available at https://github.com/LunarShen/DsicoVLA.

Most text-video retrieval methods utilize the text-image pre-trained models like CLIP as a backbone. These methods process each sampled frame independently by the image encoder, resulting in high computational overhead and limiting practical deployment. Addressing this, we focus on efficient text-video retrieval by tackling two key challenges: 1. From the perspective of trainable parameters, current parameter-efficient fine-tuning methods incur high inference costs; 2. From the perspective of model complexity, current token compression methods are mainly designed for images to reduce spatial redundancy but overlook temporal redundancy in consecutive frames of a video. To tackle these challenges, we propose Temporal Token Merging (TempMe), a parameter-efficient and training-inference efficient text-video retrieval architecture that minimizes trainable parameters and model complexity. Specifically, we introduce a progressive multi-granularity framework. By gradually combining neighboring clips, we reduce spatio-temporal redundancy and enhance temporal modeling across different frames, leading to improved efficiency and performance. Extensive experiments validate the superiority of our TempMe. Compared to previous parameter-efficient text-video retrieval methods, TempMe achieves superior performance with just 0.50M trainable parameters. It significantly reduces output tokens by 95% and GFLOPs by 51%, while achieving a 1.8X speedup and a 4.4% R-Sum improvement. With full fine-tuning, TempMe achieves a significant 7.9% R-Sum improvement, trains 1.57X faster, and utilizes 75.2% GPU memory usage. The code is available at https://github.com/LunarShen/TempMe.

Video Large Language Models have demonstrated strong video understanding capabilities, yet their practical deployment is hindered by substantial inference costs caused by redundant video tokens. Existing pruning techniques fail to effectively exploit the spatiotemporal redundancy present in video data. To bridge this gap, we perform a systematic analysis of video redundancy from two perspectives: temporal context and visual context. Leveraging these insights, we propose Dynamic Density Pruning for Fast Video LLMs termed FastVID. Specifically, FastVID dynamically partitions videos into temporally ordered segments to preserve temporal structure and applies a density-based token pruning strategy to maintain essential spatial and temporal information. Our method significantly reduces computational overhead while maintaining temporal and visual integrity. Extensive evaluations show that FastVID achieves state-of-the-art performance across various short- and long-video benchmarks on leading Video LLMs, including LLaVA-OneVision, LLaVA-Video, Qwen2-VL, and Qwen2.5-VL. Notably, on LLaVA-OneVision-7B, FastVID effectively prunes $\textbf{90.3\%}$ of video tokens, reduces FLOPs to $\textbf{8.3\%}$, and accelerates the LLM prefill stage by $\textbf{7.1}\times$, while maintaining $\textbf{98.0\%}$ of the original accuracy. The code is available at https://github.com/LunarShen/FastVID.

The capabilities of large language models (LLMs) significantly depend on training data drawn from various domains. Optimizing domain-specific mixture ratios can be modeled as a bi-level optimization problem, which we simplify into a single-level penalized form and solve with twin networks: a proxy model trained on primary data and a dynamically updated reference model trained with additional data. Our proposed method, Twin Networks for bi-level DatA mixturE optiMization (TANDEM), measures the data efficacy through the difference between the twin models and up-weights domains that benefit more from the additional data. TANDEM provides theoretical guarantees and wider applicability, compared to prior approaches. Furthermore, our bi-level perspective suggests new settings to study domain reweighting such as data-restricted scenarios and supervised fine-tuning, where optimized mixture ratios significantly improve the performance. Extensive experiments validate TANDEM's effectiveness in all scenarios.

Low-Rank Adaptation (LoRA) offers a parameter-efficient paradigm for tuning large models. While recent spectral initialization methods improve convergence and performance over the naive “Noise \& Zeros” scheme, their extra computational and storage overhead undermines efficiency. In this paper, we establish update magnitude as the fundamental driver of LoRA performance and propose LoRAM, a magnitude-driven “Basis \& Basis” initialization scheme that matches spectral methods without their inefficiencies. Our key contributions are threefold: (i) Magnitude of weight updates determines convergence. We prove low-rank structures intrinsically bound update magnitudes, unifying hyperparameter tuning in learning rate, scaling factor, and initialization as mechanisms to optimize magnitude regulation. (ii) Spectral initialization succeeds via magnitude amplification. We demystify that the presumed knowledge-driven benefit of spectral component essentially arises from the boost in the weight update magnitude. (iii) A novel and compact initialization strategy, LoRAM, scales deterministic orthogonal bases using pretrained weight magnitudes to simulate spectral gains. Extensive experiments show that LoRAM serves as a strong baseline, retaining the full efficiency of LoRA while matching or outperforming spectral initialization across benchmarks.

Prompt tuning achieves superior performance across a wide range of tasks, including multi-label zero-shot classification. Existing approaches employ multiple prompts to acquire comprehensive knowledge from categories, demonstrating state-of-the-art performance and significant computational efficiency. However, two main challenges still exist in these methods that impede the full potential of generalization. First, the class imbalance is not carefully addressed. Despite some efforts to adopt re-weighted loss functions to alleviate the positive-negative imbalance, such strategies tend to exacerbate the class imbalance by over-suppression of labels with fewer samples and overfitting to dominant classes. Second, the multi-prompt methods neglect the interactions between prompts during parameter optimization, underestimating the potential of prompts and leading to suboptimal performance. To address these issues, we present a novel framework named Dynamic Regulation in Prompt Tuning (DAR-Prompt). DAR-Prompt introduces three dynamic components: semantic regulator and debiased regulator to address the class imbalance, along with contrastive gradient regularization to enhance feature separation through prompt interactions during the backward pass. Specifically, the semantic regulator generates class-adaptive thresholds to compensate for tail classes and mitigate over-suppression, while the debiased regulator focuses on learning biased classes by rectifying overconfident predictions. Moreover, we apply dynamic regularization to the gradient update directions of prompts to promote orthogonality, thereby enhancing feature distinctiveness. Extensive experiments on several benchmarks show that our method can achieve state-of-the-art performance, well demonstrating its effectiveness and superiority. Code is available at https://github.com/Evelyn1ywliang/DAR-Prompt.

Temporal modeling plays an important role in the effective adaption of the powerful pretrained text-image foundation model into text-video retrieval. However, existing methods often rely on additional heavy trainable modules, such as transformer or BiLSTM, which are inefficient. In contrast, we avoid introducing such heavy components by leveraging frozen foundation models. To this end, we propose temporal modeling with frozen vision-language foundation models (TFVL) to model the temporal dynamics with fixed encoders. Specifically, text encoder temporal modeling (TextTemp) and image encoder temporal modeling (ImageTemp) apply frozen text and image encoders within the video head and video backbone, respectively. TextTemp uses a frozen text encoder to interpret frame representations as "visual words" within a temporal "sentence," capturing temporal dependencies. On the other hand, ImageTemp uses a frozen image encoder to treat all frame tokens as a unified visual entity, learning spatiotemporal information. The total trainable parameters of our method, comprising a lightweight projection and several prompt tokens, are significantly fewer than those in other existing methods. We evaluate the effectiveness of our method on MSR-VTT, DiDeMo, ActivityNet, and LSMDC. Compared with full fine-tuning on MSR-VTT, our TFVL achieves an average 3.25% gain in R@1 with merely 0.35% of the parameters. Extensive experiments demonstrate that the proposed TFVL outperforms state-of-the-art methods with significantly fewer parameters.

Text-video retrieval, a fundamental task for associating textual descriptions with video content, has become increasingly important in the video domain. Most existing methods focus on the single-modality features only considering the knowledge within individual video or text modalities, often neglecting cross-modal interactions. However, a text description corresponds to a specific spatio-temporal content within a video, involving a certain segment of a frame sequence and distinct sub-regions within these frames. Therefore, we focus on the text-conditioned video features to bridge the modality gap. In this article, we propose Spatio-Temporal Attention for video-text retrieval, termed STAttn, which utilizes textual information to focus on the spatio-temporal video content. Our final text-conditioned video features are generated from the text-related video frames and the text-related regions within these frames. First, we propose the Spatial Text-Attention Module (STAM) to learn the spatial information within video frames. STAM introduces the text-related salient patches to capture more fine-grained details. Second, we propose the Temporal Text-Attention Module (TTAM) to learn the temporal relationships between video frames. Temporal Triplet loss is proposed in TTAM to enhance the attention toward the text-related frames. Thus, the two modules learn the text-related spatio-temporal content from both intra-frame and inter-frame aspects. Extensive experiments on three benchmark datasets, MSRVTT, ActivityNet, and DiDeMo, demonstrate that our STAttn outperforms the state-of-the-art methods.

Collecting large-scale multi-label data with full labels is difficult for real-world scenarios. Many existing studies have tried to address the issue of missing labels caused by annotation but ignored the difficulties encountered during the annotation process. We find that the high annotation workload can be attributed to two reasons: (1) Annotators are required to identify labels on widely varying visual concepts. (2) Exhaustively annotating the entire dataset with all the labels becomes notably difficult and time-consuming. In this paper, we propose a new setting, i.e. block diagonal labels, to reduce the workload on both sides. The numerous categories can be divided into different subsets based on semantics and relevance. Each annotator can only focus on its own subset of labels so that only a small set of highly relevant labels are required to be annotated per image. To deal with the issue of such missing labels, we introduce a simple yet effective method that does not require any prior knowledge of the dataset. In practice, we propose an Adaptive Pseudo-Labeling method to predict the unknown labels with less noise. Formal analysis is conducted to evaluate the superiority of our setting. Extensive experiments are conducted to verify the effectiveness of our method on multiple widely used benchmarks.

Multi-label recognition (MLR) with incomplete labels is very challenging. Recent works strive to explore the image-to-label correspondence in the vision-language model, i.e., CLIP [22], to compensate for insufficient annotations. In spite of promising performance, they generally overlook the valuable prior about the label-to-label correspondence. In this paper, we advocate remedying the deficiency of label supervision for the MLR with incomplete labels by deriving a structured semantic prior about the label-to-label corre-spondence via a semantic prior prompter. We then present a novel Semantic Correspondence Prompt Network (SCP-Net), which can thoroughly explore the structured semantic prior. A Prior-Enhanced Self-Supervised Learning method is further introduced to enhance the use of the prior. Comprehensive experiments and analyses on several widely used benchmark datasets show that our method significantly out-performs existing methods on all datasets, well demonstrating the effectiveness and the superiority of our method. Our code will be available at https://github.com/jameslahm/SCPNet.

Collecting full annotations to construct multi-label datasets is difficult and labor-consuming. As an effective solution to relieve the annotation burden, single positive multi-label learning (SPML) draws increasing attention from both academia and industry. It only annotates each image with one positive label, leaving other labels unobserved. Therefore, existing methods strive to explore the cue of unobserved labels to compensate for the insufficiency of label supervision. Though achieving promising performance, they generally consider labels independently, leaving out the inherent hierarchical semantic relationship among labels which reveals that labels can be clustered into groups. In this paper, we propose a hierarchical prompt learning method with a novel Hierarchical Semantic Prompt Network (HSPNet) to harness such hierarchical semantic relationships using a large-scale pretrained vision and language model, i.e., CLIP, for SPML. We first introduce a Hierarchical Conditional Prompt (HCP) strategy to grasp the hierarchical label-group dependency. Then we equip a Hierarchical Graph Convolutional Network (HGCN) to capture the high-order inter-label and inter-group dependencies. Comprehensive experiments and analyses on several benchmark datasets show that our method significantly outperforms the state-of-the-art methods, well demonstrating its superiority and effectiveness. Our code will be available at https://github.com/jameslahm/HSPNet.

Aspect-based sentiment analysis (ABSA) has received increasing attention recently. ABSA can be divided into multiple tasks according to the different extracted elements. Existing generative methods usually treat the output as a whole string rather than the combination of different elements and only focus on a single task at once. This paper proposes a unified generative multi-task framework that can solve multiple ABSA tasks by controlling the type of task prompts consisting of multiple element prompts. Further, the proposed approach can train on simple tasks and transfer to difficult tasks by assembling task prompts, like assembling Lego bricks. We conduct experiments on six ABSA tasks across multiple benchmarks. Our proposed multi-task approach achieves new state-of-the-art results in almost all tasks and competitive results in task transfer scenarios.