Rebus puzzles, visual riddles that encode language through imagery, spatial arrangement, and symbolic substitution, pose a unique challenge to current vision-language models (VLMs). Unlike traditional image captioning or question answering tasks, rebus solving requires multi-modal abstraction, symbolic reasoning, and a grasp of cultural, phonetic and linguistic puns. In this paper, we investigate the capacity of contemporary VLMs to interpret and solve rebus puzzles by constructing a hand-generated and annotated benchmark of diverse English-language rebus puzzles, ranging from simple pictographic substitutions to spatially-dependent cues ("head" over "heels"). We analyze how different VLMs perform, and our findings reveal that while VLMs exhibit some surprising capabilities in decoding simple visual clues, they struggle significantly with tasks requiring abstract reasoning, lateral thinking, and understanding visual metaphors.
VOT remains a fundamental yet challenging task in computer vision due to dynamic appearance changes, occlusions, and background clutter. Traditional trackers, relying primarily on visual cues, often struggle in such complex scenarios. Recent advancements in VLMs have shown promise in semantic understanding for tasks like open-vocabulary detection and image captioning, suggesting their potential for VOT. However, the direct application of VLMs to VOT is hindered by critical limitations: the absence of a rich and comprehensive textual representation that semantically captures the target object's nuances, limiting the effective use of language information; inefficient fusion mechanisms that fail to optimally integrate visual and textual features, preventing a holistic understanding of the target; and a lack of temporal modeling of the target's evolving appearance in the language domain, leading to a disconnect between the initial description and the object's subsequent visual changes. To bridge these gaps and unlock the full potential of VLMs for VOT, we propose CLDTracker, a novel Comprehensive Language Description framework for robust visual Tracking. Our tracker introduces a dual-branch architecture consisting of a textual and a visual branch. In the textual branch, we construct a rich bag of textual descriptions derived by harnessing the powerful VLMs such as CLIP and GPT-4V, enriched with semantic and contextual cues to address the lack of rich textual representation. Experiments on six standard VOT benchmarks demonstrate that CLDTracker achieves SOTA performance, validating the effectiveness of leveraging robust and temporally-adaptive vision-language representations for tracking. Code and models are publicly available at: https://github.com/HamadYA/CLDTracker
Generating informative and knowledge-rich image captions remains a challenge for many existing captioning models, which often produce generic descriptions that lack specificity and contextual depth. To address this limitation, we propose KRCapVLM, a knowledge replay-based novel image captioning framework using vision-language model. We incorporate beam search decoding to generate more diverse and coherent captions. We also integrate attention-based modules into the image encoder to enhance feature representation. Finally, we employ training schedulers to improve stability and ensure smoother convergence during training. These proposals accelerate substantial gains in both caption quality and knowledge recognition. Our proposed model demonstrates clear improvements in both the accuracy of knowledge recognition and the overall quality of generated captions. It shows a stronger ability to generalize to previously unseen knowledge concepts, producing more informative and contextually relevant descriptions. These results indicate the effectiveness of our approach in enhancing the model's capacity to generate meaningful, knowledge-grounded captions across a range of scenarios.
Transformer-based language models, though not explicitly trained to mimic brain recordings, have demonstrated surprising alignment with brain activity. Progress in these models-through increased size, instruction-tuning, and multimodality-has led to better representational alignment with neural data. Recently, a new class of instruction-tuned multimodal LLMs (MLLMs) have emerged, showing remarkable zero-shot capabilities in open-ended multimodal vision tasks. However, it is unknown whether MLLMs, when prompted with natural instructions, lead to better brain alignment and effectively capture instruction-specific representations. To address this, we first investigate brain alignment, i.e., measuring the degree of predictivity of neural visual activity using text output response embeddings from MLLMs as participants engage in watching natural scenes. Experiments with 10 different instructions show that MLLMs exhibit significantly better brain alignment than vision-only models and perform comparably to non-instruction-tuned multimodal models like CLIP. We also find that while these MLLMs are effective at generating high-quality responses suitable to the task-specific instructions, not all instructions are relevant for brain alignment. Further, by varying instructions, we make the MLLMs encode instruction-specific visual concepts related to the input image. This analysis shows that MLLMs effectively capture count-related and recognition-related concepts, demonstrating strong alignment with brain activity. Notably, the majority of the explained variance of the brain encoding models is shared between MLLM embeddings of image captioning and other instructions. These results suggest that enhancing MLLMs' ability to capture task-specific information could lead to better differentiation between various types of instructions, and thereby improving their precision in predicting brain responses.
DeepSeek-R1 has demonstrated powerful reasoning capabilities in the text domain through stable reinforcement learning (RL). Recently, in the multimodal domain, works have begun to directly apply RL to generate R1-like free-form reasoning for Visual Question Answering (VQA) tasks. However, multimodal tasks share an intrinsically different nature from textual tasks, which heavily rely on the understanding of the input image to solve the problem. Therefore, such free-form reasoning faces two critical limitations in the VQA task: (1) Extended reasoning chains diffuse visual focus away from task-critical regions, degrading answer accuracy. (2) Unverifiable intermediate steps amplify policy-gradient variance and computational costs overhead. To address these issues, in this paper, we introduce SATORI ($\textbf{S}patially$ $\textbf{A}nchored$ $\textbf{T}ask$ $\textbf{O}ptimization$ with $\textbf{R}e\textbf{I}nforcement$ Learning), which decomposes VQA into three verifiable stages, including global image captioning, region localization, and answer prediction, each supplying explicit reward signals. Furthermore, we also introduce VQA-Verify, a 12k dataset annotated with answer-aligned captions and bounding-boxes to facilitate training. Experiments demonstrate consistent performance improvements across seven VQA benchmarks, achieving up to $15.7\%$ improvement in accuracy in accuracy compared to the R1-like baseline. Our analysis of the attention map confirms enhanced focus on critical regions, which brings improvements in accuracy. Our code is available at https://github.com/justairr/SATORI-R1.
Vision-language models (VLMs), such as CLIP, have demonstrated strong performance across a range of downstream tasks. However, CLIP is still limited in negation understanding: the ability to recognize the absence or exclusion of a concept. Existing methods address the problem by using a large language model (LLM) to generate large-scale data of image captions containing negation for further fine-tuning CLIP. However, these methods are both time- and compute-intensive, and their evaluations are typically restricted to image-text matching tasks. To expand the horizon, we (1) introduce a training-time negation data generation pipeline such that negation captions are generated during the training stage, which only increases 2.5% extra training time, and (2) we propose the first benchmark, Neg-TtoI, for evaluating text-to-image generation models on prompts containing negation, assessing model's ability to produce semantically accurate images. We show that our proposed method, TNG-CLIP, achieves SOTA performance on diverse negation benchmarks of image-to-text matching, text-to-image retrieval, and image generation.
Evaluating image captions requires cohesive assessment of both visual semantics and language pragmatics, which is often not entirely captured by most metrics. We introduce Redemption Score, a novel hybrid framework that ranks image captions by triangulating three complementary signals: (1) Mutual Information Divergence (MID) for global image-text distributional alignment, (2) DINO-based perceptual similarity of cycle-generated images for visual grounding, and (3) BERTScore for contextual text similarity against human references. A calibrated fusion of these signals allows Redemption Score to offer a more holistic assessment. On the Flickr8k benchmark, Redemption Score achieves a Kendall-$\tau$ of 56.43, outperforming twelve prior methods and demonstrating superior correlation with human judgments without requiring task-specific training. Our framework provides a more robust and nuanced evaluation by effectively redeeming image semantics and linguistic interpretability indicated by strong transfer of knowledge in the Conceptual Captions and MS COCO datasets.
Large vision-language models (LVLMs) have achieved remarkable performance on multimodal tasks such as visual question answering (VQA) and image captioning. However, they still suffer from hallucinations, generating text inconsistent with visual input, posing significant risks in real-world applications. Existing approaches to address this issue focus on incorporating external knowledge bases, alignment training, or decoding strategies, all of which require substantial computational cost and time. Recent works try to explore more efficient alternatives by adjusting LVLMs' internal representations. Although promising, these methods may cause hallucinations to be insufficiently suppressed or lead to excessive interventions that negatively affect normal semantics. In this work, we leverage sparse autoencoders (SAEs) to identify semantic directions closely associated with either hallucinations or actuality, realizing more precise and direct hallucination-related representations. Our analysis demonstrates that interventions along the faithful direction we identified can mitigate hallucinations, while those along the hallucinatory direction can exacerbate them. Building on these insights, we propose Steering LVLMs via SAE Latent Directions (SSL), a training-free method based on SAE-derived latent directions to mitigate hallucinations in LVLMs. Extensive experiments demonstrate that SSL significantly outperforms existing decoding approaches in mitigating hallucinations, while maintaining transferability across different model architectures with negligible additional time overhead.
Despite the dominance of convolutional and transformer-based architectures in image-to-image retrieval, these models are prone to biases arising from low-level visual features, such as color. Recognizing the lack of semantic understanding as a key limitation, we propose a novel scene graph-based retrieval framework that emphasizes semantic content over superficial image characteristics. Prior approaches to scene graph retrieval predominantly rely on supervised Graph Neural Networks (GNNs), which require ground truth graph pairs driven from image captions. However, the inconsistency of caption-based supervision stemming from variable text encodings undermine retrieval reliability. To address these, we present SCENIR, a Graph Autoencoder-based unsupervised retrieval framework, which eliminates the dependence on labeled training data. Our model demonstrates superior performance across metrics and runtime efficiency, outperforming existing vision-based, multimodal, and supervised GNN approaches. We further advocate for Graph Edit Distance (GED) as a deterministic and robust ground truth measure for scene graph similarity, replacing the inconsistent caption-based alternatives for the first time in image-to-image retrieval evaluation. Finally, we validate the generalizability of our method by applying it to unannotated datasets via automated scene graph generation, while substantially contributing in advancing state-of-the-art in counterfactual image retrieval.
Medical image captioning is a challenging task that requires generating clinically accurate and semantically meaningful descriptions of radiology images. While recent vision-language models (VLMs) such as BLIP, BLIP2, Gemini and ViT-GPT2 show strong performance on natural image datasets, they often produce generic or imprecise captions when applied to specialized medical domains. In this project, we explore the effectiveness of fine-tuning the BLIP model on the ROCO dataset for improved radiology captioning. We compare the fine-tuned BLIP against its zero-shot version, BLIP-2 base, BLIP-2 Instruct and a ViT-GPT2 transformer baseline. Our results demonstrate that domain-specific fine-tuning on BLIP significantly improves performance across both quantitative and qualitative evaluation metrics. We also visualize decoder cross-attention maps to assess interpretability and conduct an ablation study to evaluate the contributions of encoder-only and decoder-only fine-tuning. Our findings highlight the importance of targeted adaptation for medical applications and suggest that decoder-only fine-tuning (encoder-frozen) offers a strong performance baseline with 5% lower training time than full fine-tuning, while full model fine-tuning still yields the best results overall.
As vision-language models (VLMs) become increasingly integrated into daily life, the need for accurate visual culture understanding is becoming critical. Yet, these models frequently fall short in interpreting cultural nuances effectively. Prior work has demonstrated the effectiveness of retrieval-augmented generation (RAG) in enhancing cultural understanding in text-only settings, while its application in multimodal scenarios remains underexplored. To bridge this gap, we introduce RAVENEA (Retrieval-Augmented Visual culturE uNdErstAnding), a new benchmark designed to advance visual culture understanding through retrieval, focusing on two tasks: culture-focused visual question answering (cVQA) and culture-informed image captioning (cIC). RAVENEA extends existing datasets by integrating over 10,000 Wikipedia documents curated and ranked by human annotators. With RAVENEA, we train and evaluate seven multimodal retrievers for each image query, and measure the downstream impact of retrieval-augmented inputs across fourteen state-of-the-art VLMs. Our results show that lightweight VLMs, when augmented with culture-aware retrieval, outperform their non-augmented counterparts (by at least 3.2% absolute on cVQA and 6.2% absolute on cIC). This highlights the value of retrieval-augmented methods and culturally inclusive benchmarks for multimodal understanding.
Due to the unidirectional masking mechanism, Decoder-Only models propagate information from left to right. LVLMs (Large Vision-Language Models) follow the same architecture, with visual information gradually integrated into semantic representations during forward propagation. Through systematic analysis, we observe that over 80\% of the visual information is absorbed into the semantic representations. However, the model's attention still predominantly focuses on the visual representations. This misalignment between the attention distribution and the actual information flow undermines the model's visual understanding ability and contributes to hallucinations. To address this issue, we enhance the model's visual understanding by leveraging the core information embedded in semantic representations. Specifically, we identify attention heads that focus on core semantic representations based on their attention distributions. Then, through a two-stage optimization paradigm, we propagate the advantages of these attention heads across the entire model, aligning the attention distribution with the actual information flow. We evaluate our method on three image captioning benchmarks using five different LVLMs, demonstrating its effectiveness in significantly reducing hallucinations. Further experiments reveal a trade-off between reduced hallucinations and richer details. Notably, our method allows for manual adjustment of the model's conservativeness, enabling flexible control to meet diverse real-world requirements. Code will be released once accepted.
In many real-world applications, deployed models encounter inputs that differ from the data seen during training. Out-of-distribution detection identifies whether an input stems from an unseen distribution, while open-world recognition flags such inputs to ensure the system remains robust as ever-emerging, previously $unknown$ categories appear and must be addressed without retraining. Foundation and vision-language models are pre-trained on large and diverse datasets with the expectation of broad generalization across domains, including medical imaging. However, benchmarking these models on test sets with only a few common outlier types silently collapses the evaluation back to a closed-set problem, masking failures on rare or truly novel conditions encountered in clinical use. We therefore present $NOVA$, a challenging, real-life $evaluation-only$ benchmark of $\sim$900 brain MRI scans that span 281 rare pathologies and heterogeneous acquisition protocols. Each case includes rich clinical narratives and double-blinded expert bounding-box annotations. Together, these enable joint assessment of anomaly localisation, visual captioning, and diagnostic reasoning. Because NOVA is never used for training, it serves as an $extreme$ stress-test of out-of-distribution generalisation: models must bridge a distribution gap both in sample appearance and in semantic space. Baseline results with leading vision-language models (GPT-4o, Gemini 2.0 Flash, and Qwen2.5-VL-72B) reveal substantial performance drops across all tasks, establishing NOVA as a rigorous testbed for advancing models that can detect, localize, and reason about truly unknown anomalies.
We present Sat2Sound, a multimodal representation learning framework for soundscape mapping, designed to predict the distribution of sounds at any location on Earth. Existing methods for this task rely on satellite image and paired geotagged audio samples, which often fail to capture the diversity of sound sources at a given location. To address this limitation, we enhance existing datasets by leveraging a Vision-Language Model (VLM) to generate semantically rich soundscape descriptions for locations depicted in satellite images. Our approach incorporates contrastive learning across audio, audio captions, satellite images, and satellite image captions. We hypothesize that there is a fixed set of soundscape concepts shared across modalities. To this end, we learn a shared codebook of soundscape concepts and represent each sample as a weighted average of these concepts. Sat2Sound achieves state-of-the-art performance in cross-modal retrieval between satellite image and audio on two datasets: GeoSound and SoundingEarth. Additionally, building on Sat2Sound's ability to retrieve detailed soundscape captions, we introduce a novel application: location-based soundscape synthesis, which enables immersive acoustic experiences. Our code and models will be publicly available.
Vision-language models (VLMs) have shown remarkable progress in offline tasks such as image captioning and video question answering. However, real-time interactive environments impose new demands on VLMs, requiring them to generate utterances that are not only semantically accurate but also precisely timed. We identify two core capabilities necessary for such settings -- $\textit{perceptual updating}$ and $\textit{contingency awareness}$ -- and propose a new benchmark task, $\textbf{Temporally-Grounded Language Generation (TGLG)}$, to evaluate them. TGLG requires models to generate utterances in response to streaming video such that both content and timing align with dynamic visual input. To support this benchmark, we curate evaluation datasets from sports broadcasting and egocentric human interaction domains, and introduce a new metric, $\textbf{TRACE}$, to evaluate TGLG by jointly measuring semantic similarity and temporal alignment. Finally, we present $\textbf{Vision-Language Model with Time-Synchronized Interleaving (VLM-TSI)}$, a model that interleaves visual and linguistic tokens in a time-synchronized manner, enabling real-time language generation without relying on turn-based assumptions. Experimental results show that VLM-TSI significantly outperforms a strong baseline, yet overall performance remains modest -- highlighting the difficulty of TGLG and motivating further research in real-time VLMs. Code and data available $\href{https://github.com/yukw777/tglg}{here}$.
Language priors are a major cause of hallucinations in Large Vision-Language Models (LVLMs), often leading to text that is linguistically plausible but visually inconsistent. Recent work explores contrastive decoding as a training-free solution, but these methods typically construct negative contexts from the original image, resulting in visual information loss and distorted distribution. Motivated by the observation that language priors stem from the LLM backbone and remain consistent across images, we propose Cross-Images Contrastive Decoding (CICD), a simple yet effective training-free method that uses different images to construct negative contexts. We further analyze the cross-image behavior of language priors and introduce a distinction between essential priors (supporting fluency) and detrimental priors (causing hallucinations). By selectively preserving essential priors and suppressing detrimental ones, our method reduces hallucinations while maintaining coherent and fluent language generation. Experiments on 4 benchmarks and 6 LVLMs across three model families confirm the effectiveness and generalizability of CICD, especially in image captioning, where language priors are particularly pronounced. Code will be released once accepted.
A wide variety of agentic AI applications - ranging from cognitive assistants for dementia patients to robotics - demand a robust memory system grounded in reality. In this paper, we propose such a memory system consisting of three components. First, we combine Vision Language Models for image captioning and entity disambiguation with Large Language Models for consistent information extraction during perception. Second, the extracted information is represented in a memory consisting of a knowledge graph enhanced by vector embeddings to efficiently manage relational information. Third, we combine semantic search and graph query generation for question answering via Retrieval Augmented Generation. We illustrate the system's working and potential using a real-world example.
Localized image captioning has made significant progress with models like the Describe Anything Model (DAM), which can generate detailed region-specific descriptions without explicit region-text supervision. However, such capabilities have yet to be widely applied to specialized domains like medical imaging, where diagnostic interpretation relies on subtle regional findings rather than global understanding. To mitigate this gap, we propose MedDAM, the first comprehensive framework leveraging large vision-language models for region-specific captioning in medical images. MedDAM employs medical expert-designed prompts tailored to specific imaging modalities and establishes a robust evaluation benchmark comprising a customized assessment protocol, data pre-processing pipeline, and specialized QA template library. This benchmark evaluates both MedDAM and other adaptable large vision-language models, focusing on clinical factuality through attribute-level verification tasks, thereby circumventing the absence of ground-truth region-caption pairs in medical datasets. Extensive experiments on the VinDr-CXR, LIDC-IDRI, and SkinCon datasets demonstrate MedDAM's superiority over leading peers (including GPT-4o, Claude 3.7 Sonnet, LLaMA-3.2 Vision, Qwen2.5-VL, GPT-4Rol, and OMG-LLaVA) in the task, revealing the importance of region-level semantic alignment in medical image understanding and establishing MedDAM as a promising foundation for clinical vision-language integration.
Understanding visual art requires reasoning across multiple perspectives -- cultural, historical, and stylistic -- beyond mere object recognition. While recent multimodal large language models (MLLMs) perform well on general image captioning, they often fail to capture the nuanced interpretations that fine art demands. We propose ArtRAG, a novel, training-free framework that combines structured knowledge with retrieval-augmented generation (RAG) for multi-perspective artwork explanation. ArtRAG automatically constructs an Art Context Knowledge Graph (ACKG) from domain-specific textual sources, organizing entities such as artists, movements, themes, and historical events into a rich, interpretable graph. At inference time, a multi-granular structured retriever selects semantically and topologically relevant subgraphs to guide generation. This enables MLLMs to produce contextually grounded, culturally informed art descriptions. Experiments on the SemArt and Artpedia datasets show that ArtRAG outperforms several heavily trained baselines. Human evaluations further confirm that ArtRAG generates coherent, insightful, and culturally enriched interpretations.
Medical image reporting (MIR) aims to generate structured clinical descriptions from radiological images. Existing methods struggle with fine-grained feature extraction, multimodal alignment, and generalization across diverse imaging types, often relying on vanilla transformers and focusing primarily on chest X-rays. We propose MicarVLMoE, a vision-language mixture-of-experts model with gated cross-aligned fusion, designed to address these limitations. Our architecture includes: (i) a multiscale vision encoder (MSVE) for capturing anatomical details at varying resolutions, (ii) a multihead dual-branch latent attention (MDLA) module for vision-language alignment through latent bottleneck representations, and (iii) a modulated mixture-of-experts (MoE) decoder for adaptive expert specialization. We extend MIR to CT scans, retinal imaging, MRI scans, and gross pathology images, reporting state-of-the-art results on COVCTR, MMR, PGROSS, and ROCO datasets. Extensive experiments and ablations confirm improved clinical accuracy, cross-modal alignment, and model interpretability. Code is available at https://github.com/AI-14/micar-vl-moe.