Sitong Wu, Haoru Tan, Xichen Zhang, Bin Xia, Wenhu Zhang, Xiaojuan Qi, Bei Yu, Jiaya Jia. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Xichen Zhang, Ziyi He, Yinghao Zhu, Sitong Wu, Shaozuo Yu, Meng Chu, Wenhu Zhang, Haoru Tan, Jiaya Jia. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Data plays a pivotal role in the groundbreaking advancements in artificial intelligence. The quantitative analysis of data significantly contributes to model training, enhancing both the efficiency and quality of data utilization. However, existing data analysis tools often lag in accuracy. For instance, many of these tools even assume that the loss function of neural networks is convex. These limitations make it challenging to implement current methods effectively. In this paper, we introduce a new formulation to approximate a sample's influence by accumulating the differences in influence between consecutive learning steps, which we term Diff-In. Specifically, we formulate the sample-wise influence as the cumulative sum of its changes/differences across successive training iterations. By employing second-order approximations, we approximate these difference terms with high accuracy while eliminating the need for model convexity required by existing methods. Despite being a second-order method, Diff-In maintains computational complexity comparable to that of first-order methods and remains scalable. This efficiency is achieved by computing the product of the Hessian and gradient, which can be efficiently approximated using finite differences of first-order gradients. We assess the approximation accuracy of Diff-In both theoretically and empirically. Our theoretical analysis demonstrates that Diff-In achieves significantly lower approximation error compared to existing influence estimators. Extensive experiments further confirm its superior performance across multiple benchmark datasets in three data-centric tasks: data cleaning, data deletion, and coreset selection. Notably, our experiments on data pruning for large-scale vision-language pre-training show that Diff-In can scale to millions of data points and outperforms strong baselines.

The capacity of Large Language Models (LLMs) to comprehend and reason over long contexts is pivotal for advancements in diverse fields. Yet, they still stuggle with capturing long-distance dependencies within sequences to deeply understand semantics. To address this issue, we introduce Query-aware Inference for LLMs (Q-LLM), a system designed to process extensive sequences akin to human cognition. By focusing on memory data relevant to a given query, Q-LLM can accurately capture pertinent information within a fixed window size and provide precise answers to queries. It doesn’t require extra training and can be seamlessly integrated with any LLMs. Q-LLM using LLaMA3 (QuickLLaMA) can read Harry Potter within 30s and accurately answer the questions. On widely recognized benchmarks, Q-LLM improved by 7.17% compared to the current state-of-the-art on LLaMA3, and by 3.26% on Mistral on the \infty-bench. In the Needle-in-a-Haystack and BABILong task, Q-LLM improved upon the current SOTA by 7.0% and 6.1%. Our code is in https://github.com/dvlab-research/Q-LLM.

In recent years, developing compact and efficient large language models (LLMs) has emerged as a thriving area of research.Traditional Supervised Fine-Tuning (SFT), which relies on singular ground truth labels, often fails to capture token-level dependencies and linguistic diversity.To address these limitations, we propose a logits-based fine-tuning framework that integrates the strengths of supervised learning and knowledge distillation.Our approach constructs enriched training targets by combining teacher logits with ground truth labels, preserving both correctness and linguistic diversity.This ensures more reliable and effective training.We constructed a large-scale 1.2M logits dataset and trained a series of science-focused models.Experimental results demonstrate that our method achieves significant improvements, with accuracy gains of 18% on Mawps and 22.7% on TabMWP.Across nine widely used mathematical benchmarks, our method consistently outperforms prior SFT models, achieving an average improvement of 7.28%.Codes are available at https://github.com/dvlab- research/Logits-Based-Finetuning.

As Multi-modal Large Language Models (MLLMs) evolve, expanding beyond single-domain capabilities is essential to meet the demands for more versatile and efficient AI. However, previous omni-models have insufficiently explored speech, neglecting its integration with multi-modality. We introduce Lyra, an efficient MLLM that enhances multimodal abilities, including advanced long-speech comprehension, sound understanding, cross-modality efficiency, and seamless speech interaction. To achieve efficiency and speech-centric capabilities, Lyra employs three strategies: (1) leveraging existing open-source large models and a proposed multi-modality LoRA to reduce training costs and data requirements; (2) using a latent multi-modality regularizer and extractor to strengthen the relationship between speech and other modalities, thereby enhancing model performance; and (3) constructing a high-quality, extensive dataset that includes 1.5M multi-modal (language, vision, audio) data samples and 12K long speech samples, enabling Lyra to handle complex long speech inputs and achieve more robust omni-cognition. Compared to other omni-methods, Lyra achieves state-of-the-art performance on various vision-language, vision-speech, and speech-language benchmarks, while also using fewer computational resources and less training data.

In this paper, we present InfoMax, a novel data pruning method, also known as coreset selection, designed to maximize the information content of selected samples while minimizing redundancy. By doing so, InfoMax enhances the overall informativeness of the coreset. The information of individual samples is measured by importance scores, which capture their influence or difficulty in model learning. To quantify redundancy, we use pairwise sample similarities, based on the premise that similar samples contribute similarly to the learning process. We formalize the coreset selection problem as a discrete quadratic programming (DQP) task, with the objective of maximizing the total information content, represented as the sum of individual sample contributions minus the redundancies introduced by similar samples within the coreset. To ensure practical scalability, we introduce an efficient gradient-based solver, complemented by sparsification techniques applied to the similarity matrix and dataset partitioning strategies. This enables InfoMax to seamlessly scale to datasets with millions of samples. Extensive experiments demonstrate the superior performance of InfoMax in various data pruning tasks, including image classification, vision-language pre-training, and instruction tuning for large language models.

Dense visual contrastive learning (DRL) shows promise for learning localized information in dense prediction tasks, but struggles with establishing pixel/patch correspondence across different views (cross-contrasting). Existing methods primarily rely on self-contrasting the same view with variations, limiting input variance and hindering downstream performance. This paper delves into the mechanisms of self-contrasting and cross-contrasting, identifying the crux of the issue: transforming discrete positional embeddings to continuous representations. To address the correspondence problem, we propose a Continuous Relative Rotary Positional Query ({\mname}), enabling patch-level representation learning. Our extensive experiments on standard datasets demonstrate state-of-the-art (SOTA) results. Compared to the previous SOTA method (PQCL), our approach achieves significant improvements on COCO: with 300 epochs of pretraining, {\mname} obtains \textbf{3.4\%} mAP$^{bb}$ and \textbf{2.1\%} mAP$^{mk}$ improvements for detection and segmentation tasks, respectively. Furthermore, {\mname} exhibits faster convergence, achieving \textbf{10.4\%} mAP$^{bb}$ and \textbf{7.9\%} mAP$^{mk}$ improvements over SOTA with just 40 epochs of pretraining.

Reinforcement fine-tuning (RFT) is essential for enhancing the reasoning and generalization capabilities of large language models, but its success heavily relies on the quality of the training data. While data selection has been extensively studied in supervised learning, its role in reinforcement learning, particularly during the RFT stage, remains largely underexplored. In this work, we introduce RFT-Inf, the first influence estimator designed for data in reinforcement learning. RFT-Inf quantifies the importance of each training example by measuring how its removal affects the final training reward, offering a direct estimate of its contribution to model learning. To ensure scalability, we propose a first-order approximation of the RFT-Inf score by backtracking through the optimization process and applying temporal differentiation to the sample-wise influence term, along with a first-order Taylor approximation to adjacent time steps. This yields a lightweight, gradient-based estimator that evaluates the alignment between an individual sample’s gradient and the average gradient direction of all training samples, where a higher degree of alignment implies greater training utility. Extensive experiments demonstrate that RFT-Inf consistently improves reward performance and accelerates convergence in reinforcement fine-tuning.

Long chain-of-thought (LongCoT) has emerged as a powerful reasoning paradigm for enabling large language models (LLMs) to solve complex tasks through a systematic and thorough thinking phase. Although supervised fine-tuning (SFT) on high-quality LongCoT traces has proven effective to activate LongCoT abilities, we find that models trained in this way tend to overfit problem-specific knowledge and heuristics, leading to degraded out-of-distribution performance. To address this issue, we propose a Decoupled LongCoT Fine-Tuning (DLoFT) algorithm, which enables the model to learn generalizable LongCoT reasoning abilities while preventing overfitting to the reasoning content with problem-specific information. The key idea is to decouple the gradient into two orthogonal components: 1) a paradigm-relevant gradient corresponding to the general LongCoT paradigm and 2) a content-relevant gradient reflecting the problem-specific information, where only the former gradient is used to update model parameters. Specifically, by leveraging the unique two-phase composition (thinking and solution) of the LongCoT response, our gradient decoupling mechanism isolates the content-relevant gradient via a projection operation and separates the paradigm-relevant gradient through orthogonalization. Our DLoFT ensures the model concentrate on internalizing the LongCoT paradigm rather than memorizing problem-specific knowledge and heuristics. Extensive experiments demonstrate that our DLoFT significantly improves the generalization behavior of LongCoT abilities compared to SFT while maintaining strong in-distribution performance.

Vision transformers have gained popularity recently, leading to the development of new vision backbones with improved features and consistent performance gains. However, these advancements are not solely attributable to novel feature transformation designs; certain benefits also arise from advanced network-level and block-level architectures. This paper aims to identify the real gains of popular convolution and attention operators through a detailed study. We find that the key difference among these feature transformation modules, such as attention or convolution, lies in their spatial feature aggregation approach, known as the "spatial token mixer" (STM). To facilitate an impartial comparison, we introduce a unified architecture to neutralize the impact of divergent network-level and block-level designs. Subsequently, various STMs are integrated into this unified framework for comprehensive comparative analysis. Our experiments on various tasks and an analysis of inductive bias show a significant performance boost due to advanced network-level and block-level designs, but performance differences persist among different STMs. Our detailed analysis also reveals various findings about different STMs, including effective receptive fields, invariance, and adversarial robustness tests.

Vision-language pre-training (VLP) aims to learn joint representations of vision and language modalities. The contrastive paradigm is currently dominant in this field. However, we observe a notable misalignment phenomenon, that is, the affinity between samples has an obvious disparity across different modalities, namely “Affinity Inconsistency Problem”. Our intuition is that, for a well-aligned model, two images that look similar to each other should have the same level of similarity as their corresponding texts that describe them. In this paper, we first investigate the reason of this inconsistency problem. We discover that the lack of consideration for sample-wise affinity consistency across modalities in existing training objectives is the central cause. To address this problem, we propose a novel loss function, named Sample-wise affinity Consistency (SaCo) loss, which is designed to enhance such consistency by minimizing the distance between image embedding similarity and text embedding similarity for any two samples. Our SaCo loss can be easily incorporated into existing vision-language models as an additional loss due to its complementarity for most training objectives. In addition, considering that pre-training from scratch is computationally expensive, we also provide a more efficient way to continuously pre-train on a converged model by integrating our loss. Experimentally, the model trained with our SaCo loss significantly outperforms the baseline on a variety of vision and language tasks.

Vision Transformers have shown great potential in computer vision tasks. Most recent works have focused on elaborating the spatial token mixer for performance gains. However, we observe that a well-designed general architecture can significantly improve the performance of the entire backbone, regardless of which spatial token mixer is equipped. In this paper, we propose UniNeXt, an improved general architecture for the vision backbone. To verify its effectiveness, we instantiate the spatial token mixer with various typical and modern designs, including both convolution and attention modules. Compared with the architecture in which they are first proposed, our UniNeXt architecture can steadily boost the performance of all the spatial token mixers, and narrows the performance gap among them. Surprisingly, our UniNeXt equipped with naive local window attention even outperforms the previous state-of-the-art. Interestingly, the ranking of these spatial token mixers also changes under our UniNeXt, suggesting that an excellent spatial token mixer may be stifled due to a suboptimal general architecture, which further shows the importance of the study on the general architecture of vision backbone. Code is available at UniNeXt.

In this paper, we propose a novel data-pruning approach called moving-one-sample-out (MoSo), which aims to identify and remove the least informative samples from the training set. The core insight behind MoSo is to determine the importance of each sample by assessing its impact on the optimal empirical risk. This is achieved by measuring the extent to which the empirical risk changes when a particular sample is excluded from the training set. Instead of using the computationally expensive leaving-one-out-retraining procedure, we propose an efficient first-order approximator that only requires gradient information from different training stages. The key idea behind our approximation is that samples with gradients that are consistently aligned with the average gradient of the training set are more informative and should receive higher scores, which could be intuitively understood as follows: if the gradient from a specific sample is consistent with the average gradient vector, it implies that optimizing the network using the sample will yield a similar effect on all remaining samples. Experimental results demonstrate that MoSo effectively mitigates severe performance degradation at high pruning ratios and achieves satisfactory performance across various settings. Experimental results demonstrate that MoSo effectively mitigates severe performance degradation at high pruning ratios and outperforms state-of-the-art methods by a large margin across various settings.

In previous deep-learning-based methods, semantic segmentation has been regarded as a static or dynamic per-pixel classification task, i.e., classify each pixel representation to a specific category. However, these methods only focus on learning better pixel representations or classification kernels while ignoring the structural information of objects, which is critical to human decision-making mechanism. In this paper, we present a new paradigm for semantic segmentation, named structure-aware extraction. Specifically, it generates the segmentation results via the interactions between a set of learned structure tokens and the image feature, which aims to progressively extract the structural information of each category from the feature. Extensive experiments show that our StructToken outperforms the state-of-the-art on three widely-used benchmarks, including ADE20K, Cityscapes, and COCO-Stuff-10K.

The lightweight MLP-based decoder has become increasingly promising for semantic segmentation. However, the channel-wise MLP cannot expand the receptive fields, lacking the context modeling capacity, which is critical to semantic segmentation. In this paper, we propose a parametric-free patch rotate operation to reorganize the pixels spatially. It first divides the feature map into multiple groups and then rotates the patches within each group. Based on the proposed patch rotate operation, we design a novel segmentation network, named PRSeg, which includes an off-the-shelf backbone and a lightweight Patch Rotate MLP decoder containing multiple Dynamic Patch Rotate Blocks (DPR-Blocks). In each DPR-Block, the fully connected layer is performed following a Patch Rotate Module (PRM) to exchange spatial information between pixels. Specifically, in PRM, the feature map is first split into the reserved part and rotated part along the channel dimension according to the predicted probability of the Dynamic Channel Selection Module (DCSM), and our proposed patch rotate operation is only performed on the rotated part. Extensive experiments on ADE20K, Cityscapes and COCO-Stuff 10K datasets prove the effectiveness of our approach. We expect that our PRSeg can promote the development of MLP-based decoder in semantic segmentation.

Recently, Transformers have shown promising performance in various vision tasks. To reduce the quadratic computation complexity caused by the global self-attention, various methods constrain the range of attention within a local region to improve its efficiency. Consequently, their receptive fields in a single attention layer are not large enough, resulting in insufficient context modeling. To address this issue, we propose a Pale-Shaped self-Attention (PS-Attention), which performs self-attention within a pale-shaped region. Compared to the global self-attention, PS-Attention can reduce the computation and memory costs significantly. Meanwhile, it can capture richer contextual information under the similar computation complexity with previous local self-attention mechanisms. Based on the PS-Attention, we develop a general Vision Transformer backbone with a hierarchical architecture, named Pale Transformer, which achieves 83.4%, 84.3%, and 84.9% Top-1 accuracy with the model size of 22M, 48M, and 85M respectively for 224x224 ImageNet-1K classification, outperforming the previous Vision Transformer backbones. For downstream tasks, our Pale Transformer backbone performs better than the recent state-of-the-art CSWin Transformer by a large margin on ADE20K semantic segmentation and COCO object detection & instance segmentation. The code will be released on https://github.com/BR-IDL/PaddleViT.

Few-shot segmentation (FSS) aims to segment novel categories given scarce annotated support images. The crux of FSS is how to aggregate dense correlations between support and query images for query segmentation while being robust to the large variations in appearance and context. To this end, previous Transformer-based methods explore global consensus either on context similarity or affinity map between support-query pairs. In this work, we effectively integrate the context and affinity information via the proposed novel Context and Affinity Transformer (CATrans) in a hierarchical architecture. Specifically, the Relation-guided Context Transformer (RCT) propagates context information from support to query images conditioned on more informative support features. Based on the observation that a huge feature distinction between support and query pairs brings barriers for context knowledge transfer, the Relation-guided Affinity Transformer (RAT) measures attention-aware affinity as auxiliary information for FSS, in which the self-affinity is responsible for more reliable cross-affinity. We conduct experiments to demonstrate the effectiveness of the proposed model, outperforming the state-of-the-art methods.

Precise and accurate predictions over boundary areas are essential for semantic segmentation. However, the commonly used convolutional operators tend to smooth and blur local detail cues, making it difficult for deep models to generate accurate boundary predictions. In this paper, we introduce an operator-level approach to enhance semantic boundary awareness, so as to improve the prediction of the deep semantic segmentation model. Specifically, we formulate the boundary feature enhancement process as an anisotropic diffusion process. We propose a novel learnable approach called semantic diffusion network (SDN) for approximating the diffusion process, which contains a parameterized semantic difference convolution operator followed by a feature fusion module and constructs a differentiable mapping from original backbone features to advanced boundary-aware features. The proposed SDN is an efficient and flexible module that can be plugged into existing encoder-decoder segmentation models. Extensive experiments show that our approach can achieve consistent improvements over several typical state-of-the-art segmentation baseline models on challenging public benchmarks.

Estimating feature point correspondence is a common technique in computer vision. A line of recent data-driven approaches utilizing the graph neural networks improved the matching accuracy by a large margin. However, these learning-based methods require a lot of labeled training data, which are expensive to collect. Moreover, we find most methods are sensitive to global transforms, for example, a random rotation. On the contrary, classical geometric approaches are immune to rotational transformation though their performance is generally inferior. To tackle these issues, we propose a new learning-based matching framework, which is designed to be rotationally invariant. The model only takes geometric information as input. It consists of three parts: a graph neural network to generate a high-level local feature, an attention-based module to normalize the rotational transform, and a global feature matching module based on proximal optimization. To justify our approach, we provide a convergence guarantee for the proximal method for graph matching. The overall performance is validated by numerical experiments. In particular, our approach is trained on the synthetic random graphs and then applied to several real-world datasets. The experimental results demonstrate that our method is robust to rotational transform and highlights its strong performance of matching accuracy.