Symbolic logical reasoning is a critical yet underexplored capability of large language models (LLMs), providing reliable and verifiable decision-making in high-stakes domains such as mathematical reasoning and legal judgment.In this study, we present a systematic analysis of logical reasoning under controlled increases in logical complexity, and reveal a previously unrecognized phenomenon, which we term Logical Phase Transitions: rather than degrading smoothly, logical reasoning performance remains stable within a regime but collapses abruptly beyond a critical logical depth, mirroring physical phase transitions such as water freezing beyond a critical temperature threshold.Building on this insight, we propose Neuro-Symbolic Curriculum Tuning, a principled framework that adaptively aligns natural language with logical symbols to establish a shared representation, and reshapes training dynamics around phase-transition boundaries to progressively strengthen reasoning at increasing logical depths.Experiments on five benchmarks show that our approach effectively mitigates logical reasoning collapse at high complexity, yielding average accuracy gains of +1.26 in naive prompting and +3.95 in CoT, while improving generalization to unseen logical compositions.Code and data are available at: https://github.com/ AI4SS/Logical-Phase-Transitions.

Yunyao Zhang, Xinglang Zhang, Junxi Sheng, Wenbing Li, Junqing Yu, Yi-Ping Phoebe Chen, Wei Yang, Zikai Song. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

A recent endeavor in one class of video anomaly detection is to leverage diffusion models and posit the task as a generation problem, where the diffusion model is trained to recover normal patterns exclusively, thus reporting abnormal patterns as outliers. Yet, existing attempts neglect the various formations of anomaly and predict normal samples at the feature level regardless that abnormal objects in surveillance videos are often relatively small. To address this, a novel patch-based diffusion model is proposed, specifically engineered to capture fine-grained local information. We further observe that anomalies in videos manifest themselves as deviations in both appearance and motion. Therefore, we argue that a comprehensive solution must consider both of these aspects simultaneously to achieve accurate frame prediction. To address this, we introduce innovative motion and appearance conditions that are seamlessly integrated into our patch diffusion model. These conditions are designed to guide the model in generating coherent and contextually appropriate predictions for both semantic content and motion relations. Experimental results on four challenging video anomaly detection datasets empirically substantiate the efficacy of our proposed approach, demonstrating that it consistently outperforms most existing methods in detecting abnormal behaviors.

Multi-object tracking is a challenging vision task that requires simultaneous reasoning about object detection and object association. Conventional solutions use frame as the basic unit and typically rely on a motion predictor that exploits the appearance features to associate detected candidates, leading to insufficient adaptability to long-term associations. In this study, we propose a section-based multi-object tracking approach that integrates a temporal coherent Object Flow Tracker (OFTrack), capable of achieving simultaneous multi-frame tracking by treating multiple consecutive frames as the basic processing unit, denoted as a “section”. Our OFTrack boosts the optical flow to the object flow by employing object perception and section-based motion estimation strategies. Object perception adopts object-aware sampling and scale-aware correlation to enable precise target discrimination. Motion estimation models the correlation of different objects in multi-frames via specialized temporal-spatial attention to achieve robust association in very long videos. Additionally, to address the oscillation of unpredictable trajectories in multi-frame estimation, we have designed temporal coherent enhancement including the trajectory masking pre-training and the smoothing constraint on trajectory curves. Comprehensive experiments on several widely used benchmarks demonstrate the superior performance of our approach.

Video-based Large Language Models (Video-LLMs) have witnessed substantial advancements in recent years, propelled by the advancement in multi-modal LLMs. Although these models have demonstrated proficiency in providing the overall description of videos, they struggle with fine-grained understanding, particularly in aspects such as visual dynamics and video details inquiries. To tackle these shortcomings, we find that fine-tuning Video-LLMs on self-supervised fragment tasks, greatly improve their fine-grained video understanding abilities. Hence we propose two key contributions: (1) Self-Supervised Fragment Fine-Tuning (SF2T), a novel effortless fine-tuning method, employs the rich inherent characteristics of videos for training, while unlocking more fine-grained understanding ability of Video-LLMs. Moreover, it relieves researchers from labor-intensive annotations and smartly circumvents the limitations of natural language, which often fails to capture the complex spatiotemporal variations in videos; (2) A novel benchmark dataset, namely FineVidBench, for rigorously assessing Video-LLMs’ performance at both the scene and fragment levels, offering a comprehensive evaluation of their capabilities. We assessed multiple models and validated the effectiveness of SF2T on them. Experimental results reveal that our approach improves their ability to capture and interpret spatiotemporal details.

In this paper, we introduce Ref-GS, a novel approach for directional light factorization in 2D Gaussian splatting [8], which enables photorealistic view-dependent appearance rendering and precise geometry recovery. Ref-GS builds upon the deferred rendering of Gaussian splatting and applies directional encoding to the deferred-rendered surface, effectively reducing the ambiguity between orientation and viewing angle. Next, we introduce a spherical Mip-grid to capture varying levels of surface roughness, enabling roughness-aware Gaussian shading. Additionally, we propose a simple yet efficient geometry-lighting factorization that connects geometry and lighting via the vector outer product, significantly reducing renderer overhead when integrating volumetric attributes. Our method achieves superior photorealistic rendering for a range of open-world scenes while also accurately recovering geometry. See our interactive project page.

The development of large language models (LLMs) has significantly advanced the emergence of large multimodal models (LMMs). While LMMs have achieved tremendous success by promoting the synergy between multimodal comprehension and creation, they often face challenges when confronted with out-of-distribution data, such as which can hardly distinguish orientation, quantity, color, structure, etc. This is primarily due to their reliance on image encoders trained to encode images into task-relevant features, which may lead them to disregard irrelevant details. Delving into the modeling capabilities of diffusion models for images naturally prompts the question: Can diffusion models serve as the eyes of large language models for image perception? In this paper, we propose DEEM, a simple but effective approach that utilizes the generative feedback of diffusion models to align the semantic distributions of the image encoder. This addresses the drawbacks of previous methods that solely relied on image encoders like CLIP-ViT, thereby enhancing the model's resilience against out-of-distribution samples and reducing visual hallucinations. Importantly, this is achieved without requiring additional training modules and with fewer training parameters. We extensively evaluated DEEM on both our newly constructed RobustVQA benchmark and other well-known benchmarks, POPE and MMVP, for visual hallucination and perception. In particular, DEEM improves LMM's visual perception performance to a large extent (e.g., 4\% ↑ on RobustVQA, 6.5\% ↑ on MMVP and 12.8 \% ↑ on POPE ). Compared to the state-of-the-art interleaved content generation models, DEEM exhibits enhanced robustness and a superior capacity to alleviate model hallucinations while utilizing fewer trainable parameters, less pre-training data (10\%), and a smaller base model size. Extensive experiments demonstrate that DEEM enhances the performance of LMMs on various downstream tasks without inferior performance in the long term, including visual question answering, image captioning, and text-conditioned image synthesis.

Segmentation of brain structures from MRI is crucial for evaluating brain morphology, yet existing CNN and transformer-based methods struggle to delineate complex structures accurately. While current diffusion models have shown promise in image segmentation, they are inadequate when applied directly to brain MRI due to neglecting anatomical information. To address this, we propose Collaborative Anatomy Diffusion (CA-Diff), a framework integrating spatial anatomical features to enhance segmentation accuracy of the diffusion model. Specifically, we introduce distance field as an auxiliary anatomical condition to provide global spatial context, alongside a collaborative diffusion process to model its joint distribution with anatomical structures, enabling effective utilization of anatomical features for segmentation. Furthermore, we introduce a consistency loss to refine relationships between the distance field and anatomical structures and design a time adapted channel attention module to enhance the U-Net feature fusion procedure. Extensive experiments show that CA-Diff outperforms state-of-the-art (SOTA) methods.

Generating multi-view images from a single input view using image-conditioned diffusion models is a recent advancement and has shown considerable potential. However, issues such as the lack of consistency in synthesized views and over-smoothing in extracted geometry persist. Previous methods integrate multi-view consistency modules or impose additional supervisory to enhance view consistency while compromising on the flexibility of camera positioning and limiting the versatility of view synthesis. In this study, we consider the radiance field optimized during geometry extraction as a more rigid consistency prior, compared to volume and ray aggregation used in previous works. We further identify and rectify a critical bias in the traditional radiance field optimization process through score distillation from a multi-view diffuser. We introduce an Unbiased Score Distillation (USD) that utilizes unconditioned noises from a 2D diffusion model, greatly refining the radiance field fidelity. We leverage the rendered views from the optimized radiance field as the basis and develop a two-step specialization process of a 2D diffusion model, which is adept at conducting object-specific denoising and generating high-quality multi-view images. Finally, we recover faithful geometry and texture directly from the refined multi-view images. Empirical evaluations demonstrate that our optimized geometry and view distillation technique generates comparable results to the state-of-the-art models trained on extensive datasets, all while maintaining freedom in camera positioning. Source code of our work is publicly available at: https://youjiazhang.github.io/USD/.

Social media platforms serve as central hubs for content dissemination, opinion expression, and public engagement across diverse modalities. Accurately predicting the popularity of social media videos enables valuable applications in content recommendation, trend detection, and audience engagement. In this paper, we present Multimodal Video Predictor (MVP), our winning solution to the Video Track of the SMP Challenge 2025. MVP constructs expressive post representations by integrating deep video features extracted from pretrained models with user metadata and contextual information. The framework applies systematic preprocessing techniques, including log-transformations and outlier removal, to improve model robustness. A gradient-boosted regression model is trained to capture complex patterns across modalities. Our approach ranked first in the official evaluation of the Video Track, demonstrating its effectiveness and reliability for multimodal video popularity prediction on social platforms. The source code is available at https://github.com/yllhwa/SMPDVideo.

Video large language models (video-LLMs) have demonstrated impressive capabilities in multimodal understanding, but their potential as zero-shot evaluators for temporal consistency in video captions remains underexplored. Existing methods notably underperform in detecting critical temporal errors, such as missing, hallucinated, or misordered actions. To address this gap, we introduce two key contributions. (1) TimeJudge: a novel zero-shot framework that recasts temporal error detection as answering calibrated binary question pairs. It incorporates modality-sensitive confidence calibration and uses consistency-weighted voting for robust prediction aggregation. (2) TEDBench: a rigorously constructed benchmark featuring videos across four distinct complexity levels, specifically designed with fine-grained temporal error annotations to evaluate video-LLM performance on this task. Through a comprehensive evaluation of multiple state-of-the-art video-LLMs on TEDBench, we demonstrate that TimeJudge consistently yields substantial gains in terms of recall and F1-score without requiring any task-specific fine-tuning. Our approach provides a generalizable, scalable, and training-free solution for enhancing the temporal error detection capabilities of video-LLMs.

Generating realistic human motion sequences from text descriptions is a challenging task that requires capturing the rich expressiveness of both natural language and human motion. Recent advances in diffusion models have enabled significant progress in human motion synthesis. However, existing methods struggle to handle text inputs that describe complex or long motions. In this paper, we propose the Adaptable Motion Diffusion (AMD) model, which leverages a Large Language Model (LLM) to parse the input text into a sequence of concise and interpretable anatomical scripts that correspond to the target motion. This process exploits the LLM’s ability to provide anatomical guidance for complex motion synthesis. We then devise a two-branch fusion scheme that balances the influence of the input text and the anatomical scripts on the inverse diffusion process, which adaptively ensures the semantic fidelity and diversity of the synthesized motion. Our method can effectively handle texts with complex or long motion descriptions, where existing methods often fail. Experiments on datasets with relatively more complex motions, such as CLCD1 and CLCD2, demonstrate that our AMD significantly outperforms existing state-of-the-art models.

Multi-object tracking (MOT) is a challenging vision task that aims to detect individual objects within a single frame and associate them across multiple frames. Recent MOT approaches can be categorized into two-stage tracking-by-detection (TBD) methods and one-stage joint detection and tracking (JDT) methods. Despite the success of these approaches, they also suffer from common problems, such as harmful global or local inconsistency, poor trade-off between robustness and model complexity, and lack of flexibility in different scenes within the same video. In this paper we propose a simple but robust framework that formulates object detection and association jointly as a consistent denoising diffusion process from paired noise boxes to paired ground-truth boxes. This novel progressive denoising diffusion strategy substantially augments the tracker's effectiveness, enabling it to discriminate between various objects. During the training stage, paired object boxes diffuse from paired ground-truth boxes to random distribution, and the model learns detection and tracking simultaneously by reversing this noising process. In inference, the model refines a set of paired randomly generated boxes to the detection and tracking results in a flexible one-step or multi-step denoising diffusion process. Extensive experiments on three widely used MOT benchmarks, including MOT17, MOT20, and DanceTrack, demonstrate that our approach achieves competitive performance compared to the current state-of-the-art methods. Code is available at https://github.com/RainBowLuoCS/DiffusionTrack.

In spite of the rapidly evolving landscape of text-to-image generation, the synthesis and manipulation of multiple entities while adhering to specific relational constraints pose enduring challenges. This paper introduces an innovative progressive synthesis and editing operation that systematically incorporates entities into the target image, ensuring their adherence to spatial and relational constraints at each sequential step. Our key insight stems from the observation that while a pre-trained text-to-image diffusion model adeptly handles one or two entities, it often falters when dealing with a greater number. To address this limitation, we propose harnessing the capabilities of a Large Language Model (LLM) to decompose intricate and protracted text descriptions into coherent directives adhering to stringent formats. To facilitate the execution of directives involving distinct semantic operations—namely insertion, editing, and erasing—we formulate the Stimulus, Response, and Fusion (SRF) framework. Within this framework, latent regions are gently stimulated in alignment with each operation, followed by the fusion of the responsive latent components to achieve cohesive entity manipulation. Our proposed framework yields notable advancements in object synthesis, particularly when confronted with intricate and lengthy textual inputs. Consequently, it establishes a new benchmark for text-to-image generation tasks, further elevating the field's performance standards.

Distributing the computational workload of neural networks across cloud servers and local devices is an effective strategy for deploying resource-intensive deep models to edge devices. Therefore, compressing the deep features without compromising performance is crucial for saving server storage and transmission bandwidth. However, existing feature compression approaches are model or task specific and require training from scratch. In this paper, we propose a general and efficient framework for compressing deep features without requiring any prior knowledge of the semantics or task of the features. Our key observation is that different parts of the the feature map have different importance levels for a specific task. We can apply compression operation to a deeper degree for less irrelevant parts to achieve a high compression rate, while preserving the performance by applying a lower compression ratio to the more important parts. Focusing on this idea, we use the activation map generated by GradCAM [1] to classify each deep feature channel into essential and peripheral categories. To improve classification accuracy, we utilise semantic segmentations to provide natural boundaries for scoring each semantic channel. Peripheral channels are compressed using binary compression to achieve a high compression rate, while essential channels are compressed using mask compression. To effectively separate the essential and peripheral channels for a given input feature, we adopt a data-driven approach to identify the essential channels from datasets. Experimental results demonstrate that our method outperforms the state-of-the-art feature compression methods and is generalizable for various deep models.

Phonocardiogram classification methods using deep neural networks have been widely applied to the early detection of cardiovascular diseases recently.Despite their excellent recognition rate, the sizeable computational complexity limits their further development.Nowadays, knowledge distillation (KD) is an established paradigm for model compression.While current research on multi-teacher KD has shown potential to impart more comprehensive knowledge to the student than single-teacher KD, this approach is not suitable for all scenarios.This paper proposes a novel KD strategy to realise an adaptive multi-teacher instruction mechanism.We design a teacher selection strategy called voting network to tell the contribution of different teachers on each distillation points, so that the student can choose the useful information and renounce the redundant one.An evaluation demonstrates that our method reaches excellent accuracy (92.8 %) while maintaining a low computational complexity (0.7 M).

Point tracking is a challenging task in computer vision, aiming to establish point-wise correspondence across long video sequences. Recent advancements have primarily focused on temporal modeling techniques to improve local feature similarity, often overlooking the valuable semantic consistency inherent in tracked points. In this paper, we introduce a novel approach leveraging language embeddings to enhance the coherence of frame-wise visual features related to the same object. Our proposed method, termed autogenic language embedding for visual feature enhancement, strengthens point correspondence in long-term sequences. Unlike existing visual-language schemes, our approach learns text embeddings from visual features through a dedicated mapping network, enabling seamless adaptation to various tracking tasks without explicit text annotations. Additionally, we introduce a consistency decoder that efficiently integrates text tokens into visual features with minimal computational overhead. Through enhanced visual consistency, our approach significantly improves tracking trajectories in lengthy videos with substantial appearance variations. Extensive experiments on widely-used tracking benchmarks demonstrate the superior performance of our method, showcasing notable enhancements compared to trackers relying solely on visual cues. The code will be available at https://github.com/SkyeSong38/ALTrack.

The essence of multi-modal fusion lies in exploiting the complementary information inherent in diverse modalities.However, most prevalent fusion methods rely on traditional neural architectures and are inadequately equipped to capture the dynamics of interactions across modalities, particularly in presence of complex intra- and inter-modality correlations.Recent advancements in State Space Models (SSMs), notably exemplified by the Mamba model, have emerged as promising contenders. Particularly, its state evolving process implies stronger modality fusion paradigm, making multi-modal fusion on SSMs an appealing direction. However, fusing multiple modalities is challenging for SSMs due to its hardware-aware parallelism designs. To this end, this paper proposes the Coupled SSM model, for coupling state chains of multiple modalities while maintaining independence of intra-modality state processes. Specifically, in our coupled scheme, we devise an inter-modal hidden states transition scheme, in which the current state is dependent on the states of its own chain and that of the neighbouring chains at the previous time-step. To fully comply with the hardware-aware parallelism, we obtain the global convolution kernel by deriving the state equation while introducing the historical state.Extensive experiments on CMU-MOSEI, CH-SIMS, CH-SIMSV2 through multi-domain input verify the effectiveness of our model compared to current state-of-the-art methods, improved F1-Score by 0.4%, 0.9%, and 2.3% on the three datasets respectively, 49% faster inference and 83.7% GPU memory save. The results demonstrate that Coupled Mamba model is capable of enhanced multi-modal fusion.

Transformer framework has been showing superior performances in visual object tracking for its great strength in information aggregation across the template and search image with the well-known attention mechanism. Most recent advances focus on exploring attention mechanism variants for better information aggregation. We find these schemes are equivalent to or even just a subset of the basic self-attention mechanism. In this paper, we prove that the vanilla self-attention structure is sufficient for information aggregation, and structural adaption is unnecessary. The key is not the attention structure, but how to extract the discriminative feature for tracking and enhance the communication between the target and search image. Based on this finding, we adopt the basic vision transformer (ViT) architecture as our main tracker and concatenate the template and search image for feature embedding. To guide the encoder to capture the invariant feature for tracking, we attach a lightweight correlative masked decoder which reconstructs the original template and search image from the corresponding masked tokens. The correlative masked decoder serves as a plugin for the compact transformer tracker and is skipped in inference. Our compact tracker uses the most simple structure which only consists of a ViT backbone and a box head, and can run at 40 fps. Extensive experiments show the proposed compact transform tracker outperforms existing approaches, including advanced attention variants, and demonstrates the sufficiency of self-attention in tracking tasks. Our method achieves state-of-the-art performance on five challenging datasets, along with the VOT2020, UAV123, LaSOT, TrackingNet, and GOT-10k benchmarks. Our project is available at https://github.com/HUSTDML/CTTrack.

Transformer architecture has been showing its great strength in visual object tracking, for its effective attention mechanism. Existing transformer-based approaches adopt the pixel-to-pixel attention strategy on flattened image features and unavoidably ignore the integrity of ob-jects. In this paper, we propose a new transformer ar-chitecture with multi-scale cyclic shifting window attention for visual object tracking, elevating the attention from pixel to window level. The cross-window multi-scale at-tention has the advantage of aggregating attention at dif-ferent scales and generates the best fine-scale match for the target object. Furthermore, the cyclic shifting strat-egy brings greater accuracy by expanding the window sam-ples with positional information, and at the same time saves huge amounts of computational power by removing redun-dant calculations. Extensive experiments demonstrate the superior performance of our method, which also sets the new state-of-the-art records on five challenging datasets, along with the VOT2020, UAV123, LaSOT, TrackingNet, and GOT-lOk benchmarks. Our project is available at https://github.com/SkyeSong38/CSWinTT.

Player tracking in broadcast soccer videos has received widespread attention in the field of sports video analysis, however, we note that there is not a suitable tracking algorithm specifically for soccer video, and the existing benchmarks used for soccer player tracking cover few scenarios with low difficulties. From the observation of the soccer scene that interference and occlusion are knotty problems because the distractors are extremely similar to the targets, a distractor-aware player tracking algorithm and a high-quality benchmark for soccer play tracking (BSPT) have been presented. The distractor-aware player tracking algorithm is able to perceive semantic information about distracting players in the background by similarity judgment, the semantic distractor-aware information is encoded into a context vector and is constantly updated as the objects move through a video sequence. Distractor-aware information is then appended to the tracking result of the baseline tracker to improve the intra-class discriminative power. BSPT contains a total of 120 sequences with rich annotations. Each sequence covers 8 specialized frame-level attributes from soccer scenarios and the player occlusion situations are finely divided into 4 categories for a more comprehensive comparison. In the experimental section, the performance of our algorithm and the other 14 compared trackers are evaluated on BSPT with detailed analysis. Experimental results reveal the effectiveness of the proposed distractor-aware model especially under the attribute of occlusion. The BSPT benchmark and raw experimental results are available on the project page at http://media.hust.edu.cn/BSPT.htm.