Adversarial attacks on face recognition systems (FRSs) pose serious security and privacy threats, especially when these systems are used for identity verification. In this paper, we propose a novel method for generating adversarial faces-synthetic facial images that are visually distinct yet recognized as a target identity by the FRS. Unlike iterative optimization-based approaches (e.g., gradient descent or other iterative solvers), our method leverages the structural characteristics of the FRS feature space. We figure out that individuals sharing the same attribute (e.g., gender or race) form an attributed subsphere. By utilizing such subspheres, our method achieves both non-adaptiveness and a remarkably small number of queries. This eliminates the need for relying on transferability and open-source surrogate models, which have been a typical strategy when repeated adaptive queries to commercial FRSs are impossible. Despite requiring only a single non-adaptive query consisting of 100 face images, our method achieves a high success rate of over 93% against AWS's CompareFaces API at its default threshold. Furthermore, unlike many existing attacks that perturb a given image, our method can deliberately produce adversarial faces that impersonate the target identity while exhibiting high-level attributes chosen by the adversary.
Referring expression counting (REC) algorithms are for more flexible and interactive counting ability across varied fine-grained text expressions. However, the requirement for fine-grained attribute understanding poses challenges for prior arts, as they struggle to accurately align attribute information with correct visual patterns. Given the proven importance of ''visual density'', it is presumed that the limitations of current REC approaches stem from an under-exploration of ''contextual attribute density'' (CAD). In the scope of REC, we define CAD as the measure of the information intensity of one certain fine-grained attribute in visual regions. To model the CAD, we propose a U-shape CAD estimator in which referring expression and multi-scale visual features from GroundingDINO can interact with each other. With additional density supervision, we can effectively encode CAD, which is subsequently decoded via a novel attention procedure with CAD-refined queries. Integrating all these contributions, our framework significantly outperforms state-of-the-art REC methods, achieves $30\%$ error reduction in counting metrics and a $10\%$ improvement in localization accuracy. The surprising results shed light on the significance of contextual attribute density for REC. Code will be at github.com/Xu3XiWang/CAD-GD.
Distributed pre-training of large models at scale often imposes heavy memory demands on individual nodes and incurs significant intra-node communication costs. We propose a novel alternative approach that reduces the memory requirements by training small, structured subnetworks of the model on separate workers. Unlike pipelining, our method avoids inter-node activation communication and maintains bandwidth requirements that are comparable to or lower than standard data parallel communication schemes based on all-reduce. We evaluate two subnetwork construction strategies guided by the principle of ensuring uniform representation of each parameter across the distributed training setup. Our results show that the stochastic block dropping technique consistently outperforms the width-wise subnetwork construction previously explored in federated learning. We empirically attribute this superior performance to stronger gradient alignment in subnetworks that retain blocks having skip connections. Preliminary experiments highlight the promise of our approach, achieving a 20-40% reduction in memory usage without any loss in performance.
Colors are omnipresent in today's world and play a vital role in how humans perceive and interact with their surroundings. However, it is challenging for computers to imitate human color perception. This paper introduces the Human Perception-Based Fuzzy Color Model, COLIBRI (Color Linguistic-Based Representation and Interpretation), designed to bridge the gap between computational color representations and human visual perception. The proposed model uses fuzzy sets and logic to create a framework for color categorization. Using a three-phase experimental approach, the study first identifies distinguishable color stimuli for hue, saturation, and intensity through preliminary experiments, followed by a large-scale human categorization survey involving more than 1000 human subjects. The resulting data are used to extract fuzzy partitions and generate membership functions that reflect real-world perceptual uncertainty. The model incorporates a mechanism for adaptation that allows refinement based on feedback and contextual changes. Comparative evaluations demonstrate the model's alignment with human perception compared to traditional color models, such as RGB, HSV, and LAB. To the best of our knowledge, no previous research has documented the construction of a model for color attribute specification based on a sample of this size or a comparable sample of the human population (n = 2496). Our findings are significant for fields such as design, artificial intelligence, marketing, and human-computer interaction, where perceptually relevant color representation is critical.
Face Recognition (FR) tasks have made significant progress with the advent of Deep Neural Networks, particularly through margin-based triplet losses that embed facial images into high-dimensional feature spaces. During training, these contrastive losses focus exclusively on identity information as labels. However, we observe a multiscale geometric structure emerging in the embedding space, influenced by interpretable facial (e.g., hair color) and image attributes (e.g., contrast). We propose a geometric approach to describe the dependence or invariance of FR models to these attributes and introduce a physics-inspired alignment metric. We evaluate the proposed metric on controlled, simplified models and widely used FR models fine-tuned with synthetic data for targeted attribute augmentation. Our findings reveal that the models exhibit varying degrees of invariance across different attributes, providing insight into their strengths and weaknesses and enabling deeper interpretability. Code available here: https://github.com/mantonios107/attrs-fr-embs}{https://github.com/mantonios107/attrs-fr-embs
Recent studies have utilized visual large language models (VLMs) to answer not only "Is this face a forgery?" but also "Why is the face a forgery?" These studies introduced forgery-related attributes, such as forgery location and type, to construct deepfake VQA datasets and train VLMs, achieving high accuracy while providing human-understandable explanatory text descriptions. However, these methods still have limitations. For example, they do not fully leverage face quality-related attributes, which are often abnormal in forged faces, and they lack effective training strategies for forgery-aware VLMs. In this paper, we extend the VQA dataset to create DD-VQA+, which features a richer set of attributes and a more diverse range of samples. Furthermore, we introduce a novel forgery detection framework, MGFFD-VLM, which integrates an Attribute-Driven Hybrid LoRA Strategy to enhance the capabilities of Visual Large Language Models (VLMs). Additionally, our framework incorporates Multi-Granularity Prompt Learning and a Forgery-Aware Training Strategy. By transforming classification and forgery segmentation results into prompts, our method not only improves forgery classification but also enhances interpretability. To further boost detection performance, we design multiple forgery-related auxiliary losses. Experimental results demonstrate that our approach surpasses existing methods in both text-based forgery judgment and analysis, achieving superior accuracy.
Long-context language models (LCLMs) have exhibited impressive capabilities in long-context understanding tasks. Among these, long-context referencing -- a crucial task that requires LCLMs to attribute items of interest to specific parts of long-context data -- remains underexplored. To bridge this gap, this paper proposes Referencing Evaluation for Long-context Language Models (Ref-Long), a novel benchmark designed to assess the long-context referencing capability of LCLMs. Specifically, Ref-Long requires LCLMs to identify the indexes of documents that reference a specific key, emphasizing contextual relationships between the key and the documents over simple retrieval. Based on the task design, we construct three subsets ranging from synthetic to realistic scenarios to form the Ref-Long benchmark. Experimental results of 13 LCLMs reveal significant shortcomings in long-context referencing, even among advanced models like GPT-4o. To further investigate these challenges, we conduct comprehensive analyses, including human evaluations, task format adjustments, fine-tuning experiments, and error analyses, leading to several key insights. Our data and code can be found in https://github. com/wujunjie1998/Ref-Long.
While large multi-modal models (LMMs) demonstrate promising capabilities in segmentation and comprehension, they still struggle with two limitations: inaccurate segmentation and hallucinated comprehension. These challenges stem primarily from constraints in weak visual comprehension and a lack of fine-grained perception. To alleviate these limitations, we propose LIRA, a framework that capitalizes on the complementary relationship between visual comprehension and segmentation via two key components: (1) Semantic-Enhanced Feature Extractor (SEFE) improves object attribute inference by fusing semantic and pixel-level features, leading to more accurate segmentation; (2) Interleaved Local Visual Coupling (ILVC) autoregressively generates local descriptions after extracting local features based on segmentation masks, offering fine-grained supervision to mitigate hallucinations. Furthermore, we find that the precision of object segmentation is positively correlated with the latent related semantics of the <seg> token. To quantify this relationship and the model's potential semantic inferring ability, we introduce the Attributes Evaluation (AttrEval) dataset. Our experiments show that LIRA achieves state-of-the-art performance in both segmentation and comprehension tasks. Code will be available at https://github.com/echo840/LIRA.
Reward models trained on human preference data have demonstrated strong effectiveness in aligning Large Language Models (LLMs) with human intent under the framework of Reinforcement Learning from Human Feedback (RLHF). However, RLHF remains vulnerable to reward hacking, where the policy exploits imperfections in the reward function rather than genuinely learning the intended behavior. Although significant efforts have been made to mitigate reward hacking, they predominantly focus on and evaluate in-distribution scenarios, where the training and testing data for the reward model share the same distribution. In this paper, we empirically show that state-of-the-art methods struggle in more challenging out-of-distribution (OOD) settings. We further demonstrate that incorporating fine-grained multi-attribute scores helps address this challenge. However, the limited availability of high-quality data often leads to weak performance of multi-objective reward functions, which can negatively impact overall performance and become the bottleneck. To address this issue, we propose a unified reward modeling framework that jointly trains Bradley--Terry (BT) single-objective and multi-objective regression-based reward functions using a shared embedding space. We theoretically establish a connection between the BT loss and the regression objective and highlight their complementary benefits. Specifically, the regression task enhances the single-objective reward function's ability to mitigate reward hacking in challenging OOD settings, while BT-based training improves the scoring capability of the multi-objective reward function, enabling a 7B model to outperform a 70B baseline. Extensive experimental results demonstrate that our framework significantly improves both the robustness and the scoring performance of reward models.
Despite the rapid advances in Text-to-Image (T2I) generation models, their evaluation remains challenging in domains like fashion, involving complex compositional generation. Recent automated T2I evaluation methods leverage pre-trained vision-language models to measure cross-modal alignment. However, our preliminary study reveals that they are still limited in assessing rich entity-attribute semantics, facing challenges in attribute confusion, i.e., when attributes are correctly depicted but associated to the wrong entities. To address this, we build on a Visual Question Answering (VQA) localization strategy targeting one single entity at a time across both visual and textual modalities. We propose a localized human evaluation protocol and introduce a novel automatic metric, Localized VQAScore (L-VQAScore), that combines visual localization with VQA probing both correct (reflection) and miss-localized (leakage) attribute generation. On a newly curated dataset featuring challenging compositional alignment scenarios, L-VQAScore outperforms state-of-the-art T2I evaluation methods in terms of correlation with human judgments, demonstrating its strength in capturing fine-grained entity-attribute associations. We believe L-VQAScore can be a reliable and scalable alternative to subjective evaluations.
Pedestrian attribute recognition (PAR) is a fundamental perception task in intelligent transportation and security. To tackle this fine-grained task, most existing methods focus on extracting regional features to enrich attribute information. However, a regional feature is typically used to predict a fixed set of pre-defined attributes in these methods, which limits the performance and practicality in two aspects: 1) Regional features may compromise fine-grained patterns unique to certain attributes in favor of capturing common characteristics shared across attributes. 2) Regional features cannot generalize to predict unseen attributes in the test time. In this paper, we propose the \textbf{F}ine-grained \textbf{O}ptimization with semanti\textbf{C} g\textbf{U}ided under\textbf{S}tanding (FOCUS) approach for PAR, which adaptively extracts fine-grained attribute-level features for each attribute individually, regardless of whether the attributes are seen or not during training. Specifically, we propose the Multi-Granularity Mix Tokens (MGMT) to capture latent features at varying levels of visual granularity, thereby enriching the diversity of the extracted information. Next, we introduce the Attribute-guided Visual Feature Extraction (AVFE) module, which leverages textual attributes as queries to retrieve their corresponding visual attribute features from the Mix Tokens using a cross-attention mechanism. To ensure that textual attributes focus on the appropriate Mix Tokens, we further incorporate a Region-Aware Contrastive Learning (RACL) method, encouraging attributes within the same region to share consistent attention maps. Extensive experiments on PA100K, PETA, and RAPv1 datasets demonstrate the effectiveness and strong generalization ability of our method.
Image-conditioned generation methods, such as depth- and canny-conditioned approaches, have demonstrated remarkable abilities for precise image synthesis. However, existing models still struggle to accurately control the content of multiple instances (or regions). Even state-of-the-art models like FLUX and 3DIS face challenges, such as attribute leakage between instances, which limits user control. To address these issues, we introduce DreamRenderer, a training-free approach built upon the FLUX model. DreamRenderer enables users to control the content of each instance via bounding boxes or masks, while ensuring overall visual harmony. We propose two key innovations: 1) Bridge Image Tokens for Hard Text Attribute Binding, which uses replicated image tokens as bridge tokens to ensure that T5 text embeddings, pre-trained solely on text data, bind the correct visual attributes for each instance during Joint Attention; 2) Hard Image Attribute Binding applied only to vital layers. Through our analysis of FLUX, we identify the critical layers responsible for instance attribute rendering and apply Hard Image Attribute Binding only in these layers, using soft binding in the others. This approach ensures precise control while preserving image quality. Evaluations on the COCO-POS and COCO-MIG benchmarks demonstrate that DreamRenderer improves the Image Success Ratio by 17.7% over FLUX and enhances the performance of layout-to-image models like GLIGEN and 3DIS by up to 26.8%. Project Page: https://limuloo.github.io/DreamRenderer/.
Effective cross-modal retrieval is essential for applications like information retrieval and recommendation systems, particularly in specialized domains such as manufacturing, where product information often consists of visual samples paired with a textual description. This paper investigates the use of Vision Language Models(VLMs) for zero-shot text-to-image retrieval on fabric samples. We address the lack of publicly available datasets by introducing an automated annotation pipeline that uses Multimodal Large Language Models (MLLMs) to generate two types of textual descriptions: freeform natural language and structured attribute-based descriptions. We produce these descriptions to evaluate retrieval performance across three Vision-Language Models: CLIP, LAION-CLIP, and Meta's Perception Encoder. Our experiments demonstrate that structured, attribute-rich descriptions significantly enhance retrieval accuracy, particularly for visually complex fabric classes, with the Perception Encoder outperforming other models due to its robust feature alignment capabilities. However, zero-shot retrieval remains challenging in this fine-grained domain, underscoring the need for domain-adapted approaches. Our findings highlight the importance of combining technical textual descriptions with advanced VLMs to optimize cross-modal retrieval in industrial applications.
Vision-Language Models (VLMs) like CLIP achieve cross-modal semantic alignment through contrastive learning, exhibiting robust zero-shot generalization. Traditional prompt engineering, however, predominantly relies on coarse-grained category labels, neglecting fine-grained local semantics. Existing approaches assume that VLMs inherently recognize localized visual details and attempt to enhance classification by augmenting text prompts with attribute descriptors generated by large language models. However, our systematic experiments reveal critical limitations: CLIP's strong bias toward global image patterns hinders its ability to process localized visual descriptors. To address this fundamental constraint, we propose a simple, effective, and plug-and-play solution that enables CLIP to ``See Both the Forest and the Trees." Specifically, we employ stochastic multi-crop augmentation to activate CLIP's latent capacity for localized feature analysis. By cropping only partial regions, the approach effectively constrains the model's receptive field and recalibrates its attention mechanism, thereby mitigating its inherent bias. We evaluate the proposed method under zero-shot, few-shot, and test-time adaptation settings, and extensive experiments demonstrate that D&D achieves promising performance.
Understanding whether large language models (LLMs) possess a world model-a structured understanding of the world that supports generalization beyond surface-level patterns-is central to assessing their reliability, especially in high-stakes applications. We propose a formal framework for evaluating whether an LLM exhibits a sufficiently robust world model, defined as producing consistent outputs across semantically equivalent prompts while distinguishing between prompts that express different intents. We introduce a new evaluation approach to measure this that decomposes model response variability into three components: variability due to user purpose, user articulation, and model instability. An LLM with a strong world model should attribute most of the variability in its responses to changes in foundational purpose rather than superficial changes in articulation. This approach allows us to quantify how much of a model's behavior is semantically grounded rather than driven by model instability or alternative wording. We apply this framework to evaluate LLMs across diverse domains. Our results show how larger models attribute a greater share of output variability to changes in user purpose, indicating a more robust world model. This improvement is not uniform, however: larger models do not consistently outperform smaller ones across all domains, and their advantage in robustness is often modest. These findings highlight the importance of moving beyond accuracy-based benchmarks toward semantic diagnostics that more directly assess the structure and stability of a model's internal understanding of the world.
Multimodal large language models (MLLMs) have achieved strong performance on vision-language tasks but still struggle with fine-grained visual differences, leading to hallucinations or missed semantic shifts. We attribute this to limitations in both training data and learning objectives. To address these issues, we propose a controlled data generation pipeline that produces minimally edited image pairs with semantically aligned captions. Using this pipeline, we construct the Micro Edit Dataset (MED), containing over 50K image-text pairs spanning 11 fine-grained edit categories, including attribute, count, position, and object presence changes. Building on MED, we introduce a supervised fine-tuning (SFT) framework with a feature-level consistency loss that promotes stable visual embeddings under small edits. We evaluate our approach on the Micro Edit Detection benchmark, which includes carefully balanced evaluation pairs designed to test sensitivity to subtle visual variations across the same edit categories. Our method improves difference detection accuracy and reduces hallucinations compared to strong baselines, including GPT-4o. Moreover, it yields consistent gains on standard vision-language tasks such as image captioning and visual question answering. These results demonstrate the effectiveness of combining targeted data and alignment objectives for enhancing fine-grained visual reasoning in MLLMs.
Understanding the locus of semantic representation in large language models (LLMs) is crucial for interpretability and architectural innovation. The dominant paradigm posits that trainable input embeddings serve as foundational "meaning vectors." This paper challenges that view. We construct Transformer models where the embedding layer is entirely frozen, with vectors derived not from data, but from the visual structure of Unicode glyphs. These non-semantic, precomputed visual embeddings are fixed throughout training. Our method is compatible with any tokenizer, including a novel Unicode-centric tokenizer we introduce to ensure universal text coverage. Despite the absence of trainable, semantically initialized embeddings, our models converge, generate coherent text, and, critically, outperform architecturally identical models with trainable embeddings on the MMLU reasoning benchmark. We attribute this to "representational interference" in conventional models, where the embedding layer is burdened with learning both structural and semantic features. Our results indicate that high-level semantics are not inherent to input embeddings but are an emergent property of the Transformer's compositional architecture and data scale. This reframes the role of embeddings from meaning containers to structural primitives. We release all code and models to foster further research.
We present the first work demonstrating that a pure Mamba block can achieve efficient Dense Global Fusion, meanwhile guaranteeing top performance for camera-LiDAR multi-modal 3D object detection. Our motivation stems from the observation that existing fusion strategies are constrained by their inability to simultaneously achieve efficiency, long-range modeling, and retaining complete scene information. Inspired by recent advances in state-space models (SSMs) and linear attention, we leverage their linear complexity and long-range modeling capabilities to address these challenges. However, this is non-trivial since our experiments reveal that simply adopting efficient linear-complexity methods does not necessarily yield improvements and may even degrade performance. We attribute this degradation to the loss of height information during multi-modal alignment, leading to deviations in sequence order. To resolve this, we propose height-fidelity LiDAR encoding that preserves precise height information through voxel compression in continuous space, thereby enhancing camera-LiDAR alignment. Subsequently, we introduce the Hybrid Mamba Block, which leverages the enriched height-informed features to conduct local and global contextual learning. By integrating these components, our method achieves state-of-the-art performance with the top-tire NDS score of 75.0 on the nuScenes validation benchmark, even surpassing methods that utilize high-resolution inputs. Meanwhile, our method maintains efficiency, achieving faster inference speed than most recent state-of-the-art methods.
Legal NLP remains underdeveloped in regions like India due to the scarcity of structured datasets. We introduce IndianBailJudgments-1200, a new benchmark dataset comprising 1200 Indian court judgments on bail decisions, annotated across 20+ attributes including bail outcome, IPC sections, crime type, and legal reasoning. Annotations were generated using a prompt-engineered GPT-4o pipeline and verified for consistency. This resource supports a wide range of legal NLP tasks such as outcome prediction, summarization, and fairness analysis, and is the first publicly available dataset focused specifically on Indian bail jurisprudence.
We often attribute human characteristics to large language models (LLMs) and claim that they "know" certain things. LLMs have an internal probabilistic knowledge that represents information retained during training. How can we assess the veracity of this knowledge? We examine two common methods for probing the veracity of LLMs and discover several assumptions that are flawed. To address these flawed assumptions, we introduce sAwMIL (short for Sparse Aware Multiple-Instance Learning), a probing method that utilizes the internal activations of LLMs to separate statements into true, false, and neither. sAwMIL is based on multiple-instance learning and conformal prediction. We evaluate sAwMIL on 5 validity criteria across 16 open-source LLMs, including both default and chat-based variants, as well as on 3 new datasets. Among the insights we provide are: (1) the veracity signal is often concentrated in the third quarter of an LLM's depth; (2) truth and falsehood signals are not always symmetric; (3) linear probes perform better on chat models than on default models; (4) nonlinear probes may be required to capture veracity signals for some LLMs with reinforcement learning from human feedback or knowledge distillation; and (5) LLMs capture a third type of signal that is distinct from true and false and is neither true nor false. These findings provide a reliable method for verifying what LLMs "know" and how certain they are of their probabilistic internal knowledge.