Recent advances in large language models have enabled task prompting for open-ended text generation. In the vision domain, a longstanding goal is developing models capable of general visual learning, encompassing tasks such as image generation, editing, low-level processing, and dense perception. Although recent efforts have aimed at building vision foundation models that support prompting, significant challenges remain, particularly in accurately comprehending visual prompts and addressing the ambiguity inherent in textual prompts. To address this, we introduce X-Prompt, a purely auto-regressive large vision-language model designed for generalizable visual learning via in-context prompting. X-Prompt can process visual and textual prompts as context, enabling precise task interpretation and accurate execution. A novel prompt-token fusion mechanism effectively extracts relevant task information from complex prompts while significantly reducing the token length. Additionally, a unified training strategy for text and image prediction enhances task awareness, enabling seamless adaptation to open-ended prompts. Extensive experiments demonstrate that X-Prompt effectively interprets in-context prompts and exhibits generalization across both in-domain and out-of-domain visual tasks, paving the way for future advancements in general visual learning.
论文检索
输入标题、作者或关键词，从 7,537 篇学术成果中精准定位
Pretraining a Multiple Instance Learning (MIL) aggregator enables the derivation of Whole Slide Image (WSI)-level embeddings from patch-level representations without supervision. While recent multimodal MIL pretraining approaches leveraging auxiliary modalities have demonstrated performance gains over unimodal WSI pretraining, the acquisition of these additional modalities necessitates extensive clinical profiling. This requirement increases costs and limits scalability in existing WSI datasets lacking such paired modalities. To address this, we propose Gigapixel Vision-Concept Knowledge Contrastive pretraining (GECKO), which aligns WSIs with a Concept Prior derived from the available WSIs. First, we derive an inherently interpretable concept prior by computing the similarity between each WSI patch and textual descriptions of predefined pathology concepts. GECKO then employs a dual-branch MIL network: one branch aggregates patch embeddings into a WSI-level deep embedding, while the other aggregates the concept prior to a corresponding WSI-level concept embedding. Both aggregated embeddings are aligned using a contrastive objective, thereby pretraining the entire dual-branch MIL model. Moreover, when auxiliary modalities such as transcriptomics data are available, GECKO seamlessly integrates them. Across five diverse tasks, GECKO consistently outperforms prior unimodal and multimodal pretraining approaches while also delivering clinically meaningful interpretability that bridges the gap between computational models and pathology expertise. Code is made available at github.com/bmi-imaginelab/GECKO
Reliable integration of prior information is crucial for self-verifying and self-updating HD maps. However, no public dataset includes the required triplet of prior maps, current maps, and sensor data. As a result, existing methods must rely on synthetic priors, which create inconsistencies and lead to a significant sim2real gap. To address this, we introduce ArgoTweak, the first dataset to complete the triplet with realistic map priors. At its core, ArgoTweak employs a bijective mapping framework, breaking down large-scale modifications into fine-grained atomic changes at the map element level, thus ensuring interpretability. This paradigm shift enables accurate change detection and integration while preserving unchanged elements with high fidelity. Experiments show that training models on ArgoTweak significantly reduces the sim2real gap compared to synthetic priors. Extensive ablations further highlight the impact of structured priors and detailed change annotations. By establishing a benchmark for explainable, prior-aided HD mapping, ArgoTweak advances scalable, self-improving mapping solutions. The dataset, baselines, map modification toolbox, and further resources are available at https://KTH-RPL.github.io/ArgoTweak/.
Recent years have witnessed remarkable advances in Large Vision-Language Models (LVLMs), which have achieved human-level performance across various complex vision-language tasks. Following LLaVA's paradigm, mainstream LVLMs typically employ a shallow MLP for visual-language alignment through a two-stage training process: pretraining for cross-modal alignment followed by instruction tuning. While this approach has proven effective, the underlying mechanisms of how MLPs bridge the modality gap remain poorly understood. Although some research has explored how LLMs process transformed visual tokens, few studies have investigated the fundamental alignment mechanism. Furthermore, the MLP adapter requires retraining whenever switching LLM backbones. To address these limitations, we first investigate the working principles of MLP adapters and discover that they learn to project visual embeddings into subspaces spanned by corresponding text embeddings progressively. Based on this insight, we propose LangBridge, a novel adapter that explicitly maps visual tokens to linear combinations of LLM vocabulary embeddings. This innovative design enables pretraining-free adapter transfer across different LLMs while maintaining performance. Our experimental results demonstrate that a LangBridge adapter pre-trained on Qwen2-0.5B can be directly applied to larger models such as LLaMA3-8B or Qwen2.5-14B while maintaining competitive performance. Overall, LangBridge enables interpretable vision-language alignment by grounding visual representations in LLM vocab embedding, while its plug-and-play design ensures efficient reuse across multiple LLMs with nearly no performance degradation. See our project page at https://curryx-001.github.io/LangBridge.github.io/.
This paper explains training-time out-of-distribution (OOD) detection from a novel view, i.e., interactions between different input variables of deep neural networks (DNNs). Specifically, we provide a unified understanding of the effectiveness of current training-time OOD detection methods, i.e., DNNs trained with these methods all encode more complex interactions for inference than those trained without training-time methods, which contributes to their superior OOD detection performance. We further conduct thorough empirical analyses and verify that complex interactions play a primary role in OOD detection, by developing a simple-yet-efficient method to force the DNN to learn interactions of specific complexities and evaluate the change of OOD detection performances. Besides, we also use interactions to investigate why near-OOD samples are more difficult to distinguish from in-distribution (ID) samples than far-OOD samples, mainly because compared to far-OOD samples, the distribution of interactions in near-OOD samples is more similar to that of ID samples. Moreover, we discover that training-time OOD detection methods can effectively decrease such similarities.
It has been observed that deep neural networks (DNNs) often use both genuine as well as spurious features.In this work, we propose "Amending Inherent Interpretability via Self-Supervised Masking" (AIM), a simple yet surprisingly effective method that promotes the network's utilization of genuine features over spurious alternatives without requiring additional annotations.In particular, AIM uses features at multiple encoding stages to guide a self-supervised, sample-specific feature-masking process. As a result, AIM allows training well-performing and inherently interpretable models that faithfully summarize the decision process.When tested on challenging datasets designed to assess reliance on spurious features and out-of-domain generalization, AIM networks demonstrate significant dual benefits: Evaluations show that AIM improves interpretability, as measured by the Energy Pointing Game (EPG) score, by ~6-37%, while simultaneously enhancing accuracy by ~10-40%. These impressive performance gains are further validated on the standard in-domain CUB-200 dataset for fine-grained classification. The results provide compelling evidence supporting our hypothesis that AIM finds genuine and meaningful features that directly contribute to its improved human interpretability.
Understanding the dynamics of a physical scene involves reasoning about the diverse ways it can potentially change, especially as a result of local interactions. We present the Flow Poke Transformer (FPT), a novel framework for directly predicting the distribution of local motion, conditioned on sparse interactions termed "pokes". Unlike traditional methods that typically only enable dense sampling of a single realization of scene dynamics, FPT provides an interpretable directly accessible representation of multi-modal scene motion, its dependency on physical interactions and the inherent uncertainties of scene dynamics. We also evaluate our model on several downstream tasks to enable comparisons with prior methods and highlight the flexibility of our approach. On dense face motion generation, our generic pre-trained model surpasses specialized baselines. FPT can be fine-tuned in strongly out-of-distribution tasks such as synthetic datasets to enable significant improvements over in-domain methods in articulated object motion estimation. Additionally, predicting explicit motion distributions directly enables our method to achieve competitive performance on tasks like moving part segmentation from pokes which further demonstrates the versatility of our FPT.
Text-to-image generation models often struggle to interpret spatially aware text prompts effectively. To overcome this, existing approaches typically require millions of high-quality semantic layout annotations consisting of bounding boxes and regional prompts. This paper shows that the large amounts of regional prompts are non-necessary for the latest diffusion transformers like SD3 or FLUX.In this paper, we propose an efficient hybrid layout framework for diffusion transformers. Our approach drastically reduces need for extensive layout annotations and minimizes reliance on regional prompt annotations--incurring only minimal additional computational cost during inference--while maintaining high-quality layout adherence. Our key insight is to break the layout-control task into two sequential stages: first, generating the target objects within the designated regions specified by an anonymous layout, and second, refining these outputs to ensure they strictly adhere to the regional prompts in the semantic layout. Building on this insight, we propose a hybrid layout control scheme that first fine-tunes the DiTs (e.g., SD3) to follow an anonymous layout, then continues fine-tuning the DiTs to follow the semantic layout, and finally includes a quality-tuning stage to enhance visual aesthetics. We show that this hybrid design is highly data-efficient, as we find only using a small amount of semantic layout annotations is sufficient, thereby significantly reducing dependency on regional prompts. In addition, we propose an efficient regional diffusion transformer to encode the spatial layout information using just a set of lower-resolution regional tokens instead of various carefully designed layout tokens. The region-wise diffusion loss over these regional tokens can guide the diffusion transformer learn to follow the given layout implicitly. We empirically validate the effectiveness of our approach by comparing it with the latest version of SiamLayout and show that our method achieves better results while being more than 10 times more data efficient and ensuring superior aesthetics.
Large-scale vision-language models (VLMs), such as CLIP, have achieved remarkable success in zero-shot learning (ZSL) by leveraging large-scale visual-text pair datasets. However, these methods often lack interpretability, as they compute the similarity between an entire query image and the embedded category words, making it difficult to explain their predictions. One approach to address this issue is to develop interpretable models by integrating language, where classifiers are built using discrete attributes, similar to human perception. This introduces a new challenge: how to effectively align local visual features with corresponding attributes based on pre-trained VLMs. To tackle this, we propose LaZSL, a locally-aligned vision-language model for interpretable ZSL. LaZSL employs local visual-semantic alignment via optimal transport to perform interaction between visual regions and their associated attributes, facilitating effective alignment and providing interpretable similarity without the need for additional training. Extensive experiments demonstrate that our method offers several advantages, including enhanced interpretability, improved accuracy, and strong domain generalization. Codes available at: https://github.com/shiming-chen/LaZSL.
Depth images captured by Time-of-Flight (ToF) sensors are prone to noise, requiring denoising for reliable downstream applications. Previous works either focus on single-frame processing, or perform multi-frame processing without considering depth variations at corresponding pixels across frames, leading to undesirable temporal inconsistency and spatial ambiguity. In this paper, we propose a novel ToF depth denoising network leveraging motion-invariant graph fusion to simultaneously enhance temporal stability and spatial sharpness. Specifically, despite depth shifts across frames, graph structures exhibit temporal self-similarity, enabling cross-frame geometric attention for graph fusion. Then, by incorporating an image smoothness prior on the fused graph and data fidelity term derived from ToF noise distribution, we formulate a maximum a posterior problem for ToF denoising. Finally, the solution is unrolled into iterative filters whose weights are adaptively learned from the graph-informed geometric attention, producing a high-performance yet interpretable network. Experimental results demonstrate that the proposed scheme achieves state-of-the-art performance in terms of accuracy and consistency on synthetic DVToF dataset and exhibits robust generalization on the real Kinectv2 dataset. Source code is available at https://github.com/davidweidawang/GIGA-ToF.
Bootstrapping Grounded Chain-of-Thought in Multimodal LLMs for Data-Efficient Model Adaptation
PDF ↗Multimodal Large Language Models (MLLMs) have demonstrated remarkable capabilities in interpreting images using natural language. However, without using large-scale datasets for retraining, these models are difficult to adapt to specialized vision tasks, e.g., chart understanding. This problem is caused by a mismatch between pre-training and downstream datasets: pre-training datasets primarily concentrate on scenes and objects but contain limited information about specialized, non-object images, such as charts and tables. In this paper, we share an interesting finding that training an MLLM with chain-of-thought (CoT) reasoning data can facilitate model adaptation in specialized vision tasks, especially under data-limited regimes. However, we identify a critical issue within CoT data distilled from pre-trained MLLMs, i.e., the data often contains multiple factual errors in the reasoning steps. To address the problem, we propose Grounded Chain-of-Thought (GCoT), a simple bootstrapping-based approach that aims to inject grounding information (i.e., bounding boxes) into CoT data, essentially making the reasoning steps more faithful to input images. We evaluate our approach on five specialized vision tasks, which cover a variety of visual formats including charts, tables, receipts, and reports. The results demonstrate that under data-limited regimes our approach significantly improves upon fine-tuning and distillation.
Information Pursuit (IP) is a recently introduced learning framework to construct classifiers that are interpretable-by-design. Given a set of task-relevant and interpretable data queries, IP selects a small subset of the most informative queries and makes predictions based on the gathered query-answer pairs. However, a key limitation of IP is its dependency on task-relevant interpretable queries, which typically require considerable data annotation and curation efforts. While previous approaches have explored using general-purpose large language models to generate these query sets, they rely on prompt engineering heuristics and often yield suboptimal query sets, resulting in a performance gap between IP and non-interpretable black-box predictors. In this work, we propose parameterizing IP queries as a learnable dictionary defined in the latent space of vision-language models such as CLIP. We formulate an optimization objective to learn IP queries and propose an alternating optimization algorithm that shares appealing connections with classic sparse dictionary learning algorithms. Our learned dictionary outperforms baseline methods based on handcrafted or prompted dictionaries across several image classification benchmarks.
Know Your Attention Maps: Class-specific Token Masking for Weakly Supervised Semantic Segmentation
PDF ↗Weakly Supervised Semantic Segmentation (WSSS) is a challenging problem that has been extensively studied in recent years. Traditional approaches often rely on external modules like Class Activation Maps to highlight regions of interest and generate pseudo segmentation masks. In this work, we propose an end-to-end method that directly utilizes the attention maps learned by a Vision Transformer (ViT) for WSSS. We propose training a sparse ViT with multiple [CLS] tokens (one for each class), using a random masking strategy to promote [CLS] token - class assignment. At inference time, we aggregate the different self-attention maps of each [CLS] token corresponding to the predicted labels to generate pseudo segmentation masks. Our proposed approach enhances the interpretability of self-attention maps and ensures accurate class assignments. Extensive experiments on two standard benchmarks and three specialized datasets demonstrate that our method generates accurate pseudo-masks, outperforming related works. Those pseudo-masks can be used to train a segmentation model which achieves results comparable to fully-supervised models, significantly reducing the need for fine-grained labeled data.
Are VLMs Ready for Autonomous Driving? An Empirical Study from the Reliability, Data and Metric Perspectives
PDF ↗Recent advancements in Vision-Language Models (VLMs) have fueled interest in autonomous driving applications, particularly for interpretable decision-making. However, the assumption that VLMs provide visually grounded and reliable driving explanations remains unexamined. To address this, we introduce DriveBench, a benchmark evaluating 12 VLMs across 17 settings, covering 19,200 images, 20,498 QA pairs, and four key driving tasks. Our findings reveal that VLMs often generate plausible responses from general knowledge or textual cues rather than true visual grounding, especially under degraded or missing visual inputs. This behavior, concealed by dataset imbalances and insufficient evaluation metrics, poses significant risks in safety-critical scenarios like autonomous driving. We further observe that VLMs possess inherent corruption-awareness but only explicitly acknowledge these issues when directly prompted. Given the challenges and inspired by the inherent corruption awareness, we propose Robust Agentic Utilization (RAU), leveraging VLMs' corruption awareness and agentic planning with external tools to enhance perception reliability for downstream tasks. Our study challenges existing evaluation paradigms and provides a roadmap toward more robust and interpretable autonomous driving systems.
Recent advances in Large Multimodal Models (LMMs) lead to significant breakthroughs in both academia and industry. One question that arises is how we, as humans, can understand their internal neural representations. This paper takes an initial step towards addressing this question by presenting a versatile framework to identify and interpret the semantics within LMMs. Specifically, 1) we first apply a Sparse Autoencoder(SAE) to disentangle the representations into human understandable features. 2) We then present an automatic interpretation framework to interpreted the open-semantic features learned in SAE by the LMMs themselves. We employ this framework to analyze the LLaVA-NeXT-8B model using the LLaVA-OV-72B model, demonstrating that these features can effectively steer the model's behavior. Our results contribute to a deeper understanding of why LMMs excel in specific tasks, including EQ tests, and illuminate the nature of their mistakes along with potential strategies for their rectification. These findings offer new insights into the internal mechanisms of LMMs and suggest parallels with the cognitive processes of the human brain.
Adversarial attack plays a critical role in evaluating the robustness of deep learning models. Jacobian-based Saliency Map Attack (JSMA) is an interpretable adversarial method that offers excellent pixel-level control and provides valuable insights into model vulnerabilities. However, its quadratic computational complexity O(M^2 xN) renders it impractical for large-scale datasets, limiting its application despite its inherent value. This paper proposes FastJSMA, an efficient attack method that addresses these computational limitations. Our approach introduces a gradient decoupling mechanism that decomposes the Jacobian calculation into complementary class suppression (g^-) and class excitation (g^+) gradients, reducing complexity to O(M\sqrt N ). Additionally, we implement a class probing mechanism and an adaptive saliency threshold to further optimize the process. Experimental results across multiple datasets demonstrate that FastJSMA maintains comparable attack success rates while dramatically reducing computation time--requiring only 2.9% of JSMA's processing time on CIFAR-10 and 1.2% on CIFAR-100, and successfully operating on ImageNet where traditional JSMA fails due to memory constraints. This advancement enables the practical application of interpretable saliency map-based attacks on large-scale datasets, balancing effectiveness with computational efficiency.
C2MIL: Synchronizing Semantic and Topological Causalities in Multiple Instance Learning for Robust and Interpretable Survival Analysis
PDF ↗Graph-based Multiple Instance Learning (MIL) is widely used in survival analysis with Hematoxylin and Eosin (H&E)-stained whole slide images (WSIs) due to its ability to capture topological information. However, variations in staining and scanning can introduce semantic bias, while topological subgraphs that are not relevant to the causal relationships can create noise, resulting in biased slide-level representations. These issues can hinder both the interpretability and generalization of the analysis. To tackle this, we introduce a dual structural causal model as the theoretical foundation and propose a novel and interpretable dual causal graph-based MIL model, C2MIL. C2MIL incorporates a novel cross-scale adaptive feature disentangling module for semantic causal intervention and a new Bernoulli differentiable causal subgraph sampling method for topological causal discovery. A joint optimization strategy combining disentangling supervision and contrastive learning enables simultaneous refinement of both semantic and topological causalities. Experiments demonstrate that C2MIL consistently improves generalization and interpretability over existing methods and can serve as a causal enhancement for diverse MIL baselines. The code is available at https://github.com/mimic0127/C2MIL.
Most frame-based learned video codecs can be interpreted as recurrent neural networks (RNNs) propagating reference information along the temporal dimension. This work revisits the limitations of the current approaches from an RNN perspective. The output-recurrence methods, which propagate decoded frames, are intuitive but impose dual constraints on the output decoded frames, leading to suboptimal rate-distortion performance. In contrast, the hidden-to-hidden connection approaches, which propagate latent features within the RNN, offer greater flexibility but require large buffer sizes. To address these issues, we propose HyTIP, a learned video coding framework that combines both mechanisms. Our hybrid buffering strategy uses explicit decoded frames and a small number of implicit latent features to achieve competitive coding performance. Experimental results show that our HyTIP outperforms the sole use of either output-recurrence or hidden-to-hidden approaches. Furthermore, it achieves comparable performance to state-of-the-art methods but with a much smaller buffer size, and outperforms VTM 17.0 (Low-delay B) in terms of PSNR-RGB and MS-SSIM-RGB. The source code of HyTIP is available at https://github.com/NYCU-MAPL/HyTIP.
CoTMR: Chain-of-Thought Multi-Scale Reasoning for Training-Free Zero-Shot Composed Image Retrieval
PDF ↗Zero-Shot Composed Image Retrieval (ZS-CIR) aims to retrieve target images by integrating information from a composed query (reference image and modification text) without training samples. Existing methods primarily combine caption models and Large Language Models (LLMs) to generate target captions from composed queries but face various issues such as incompatibility, visual information loss, and insufficient reasoning. In this work, we propose CoTMR, a training-free framework with novel Chain-of-thought (CoT) and Multi-scale Reasoning. Instead of relying on caption models for modality transformation, CoTMR directly employs the Large Vision-Language Model (LVLM) to achieve unified understanding and reasoning of composed queries. To enhance reasoning reliability, we devise CIRCoT, which guides the LVLM to perform step-by-step reasoning by following predefined subtasks. Additionally, while most existing approaches focus solely on global-level reasoning, CoTMR introduces fine-grained predictions about the presence or absence of key elements at the object scale for more comprehensive reasoning. Furthermore, we design a Multi-Grained Scoring (MGS) mechanism, which integrates CLIP similarity scores of the above reasoning outputs with candidate images to realize precise retrieval. Extensive experiments demonstrate that our CoTMR not only drastically outperforms previous methods across four prominent benchmarks but also offers appealing interpretability.
The rapid advancements in generative technology have emerged as a double-edged sword. While offering powerful tools that enhance convenience, they also pose significant social concerns. As defenders, current synthetic image detection methods often lack artifact-level textual interpretability and are overly focused on image manipulation detection, and current datasets usually suffer from outdated generators and a lack of fine-grained annotations. In this paper, we introduce SynthScars, a high-quality and diverse dataset consisting of 12,236 fully synthetic images with human-expert annotations. It features 4 distinct image content types, 3 categories of artifacts, and fine-grained annotations covering pixel-level segmentation, detailed textual explanations, and artifact category labels. Furthermore, we propose LEGION (LEarning to Ground and explain for Synthetic Image detectiON), a multimodal large language model (MLLM)-based image forgery analysis framework that integrates artifact detection, segmentation, and explanation. Building upon this capability, we further explore LEGION as a controller, integrating it into image refinement pipelines to guide the generation of higher-quality and more realistic images. Extensive experiments show that LEGION outperforms existing methods across multiple benchmarks, particularly surpassing the second-best traditional expert on SynthScars by 3.31% in mIoU and 7.75% in F1 score. Moreover, the refined images generated under its guidance exhibit stronger alignment with human preferences. More information about LEGION can be found at https://opendatalab.github.io/LEGION.