Skill composition is the ability to combine previously learned skills to solve new tasks. As neural networks acquire increasingly complex skills during their pretraining, it is not clear how successfully they can compose them. In this paper, we focus on Multimodal Large Language Models (MLLM), and study their ability to compose skills across modalities. To this end, we design three evaluation tasks which can be solved sequentially composing two modality-dependent skills, and evaluate several open MLLMs under two main settings: i) prompting the model to directly solve the task, and ii) using a two-step cascaded inference approach, which manually enforces the composition of the two skills for a given task. Even with these straightforward compositions, we find that all evaluated MLLMs exhibit a significant cross-modality skill composition gap. To mitigate the aforementioned gap, we explore two alternatives: i) use chain-of-thought prompting to explicitly instruct MLLMs for skill composition and ii) a specific fine-tuning recipe to promote skill composition. Although those strategies improve model performance, they still exhibit significant skill composition gaps, suggesting that more research is needed to improve cross-modal skill composition in MLLMs.

Large Vision-Language Models (LVLMs) have achieved significant progress in combining visual comprehension with language generation. Despite this success, the training data of LVLMs still suffers from Long-Tail (LT) problems, where the data distribution is highly imbalanced. Previous works have mainly focused on traditional VLM architectures, i.e., CLIP or ViT, and specific tasks such as recognition and classification. Nevertheless, the exploration of LVLM (e.g. LLaVA) and more general tasks (e.g. Visual Question Answering and Visual Reasoning) remains under-explored. In this paper, we first conduct an in-depth analysis of the LT issues in LVLMs and identify two core causes: the overrepresentation of head concepts and the underrepresentation of tail concepts. Based on the above observation, we propose an $\textbf{A}$daptive $\textbf{D}$ata $\textbf{R}$efinement Framework ($\textbf{ADR}$), which consists of two stages: $\textbf{D}$ata $\textbf{R}$ebalancing ($\textbf{DR}$) and $\textbf{D}$ata $\textbf{S}$ynthesis ($\textbf{DS}$). In the DR stage, we adaptively rebalance the redundant data based on entity distributions, while in the DS stage, we leverage Denoising Diffusion Probabilistic Models (DDPMs) and scarce images to supplement underrepresented portions. Through comprehensive evaluations across eleven benchmarks, our proposed ADR effectively mitigates the long-tail problem in the training data, improving the average performance of LLaVA 1.5 relatively by 4.36%, without increasing the training data volume.

Vision-Language Models (VLMs) have advanced multi-modal tasks like image captioning, visual question answering, and reasoning. However, they often generate hallucinated outputs inconsistent with the visual context or prompt, limiting reliability in critical applications like autonomous driving and medical imaging. Existing studies link hallucination to statistical biases, language priors, and biased feature learning but lack a structured causal understanding. In this work, we introduce a causal perspective to analyze and mitigate hallucination in VLMs. We hypothesize that hallucination arises from unintended direct influences of either the vision or text modality, bypassing proper multi-modal fusion. To address this, we construct a causal graph for VLMs and employ counterfactual analysis to estimate the Natural Direct Effect (NDE) of vision, text, and their cross-modal interaction on the output. We systematically identify and mitigate these unintended direct effects to ensure that responses are primarily driven by genuine multi-modal fusion. Our approach consists of three steps: (1) designing structural causal graphs to distinguish correct fusion pathways from spurious modality shortcuts, (2) estimating modality-specific and cross-modal NDE using perturbed image representations, hallucinated text embeddings, and degraded visual inputs, and (3) implementing a test-time intervention module to dynamically adjust the model's dependence on each modality. Experimental results demonstrate that our method significantly reduces hallucination while preserving task performance, providing a robust and interpretable framework for improving VLM reliability. To enhance accessibility and reproducibility, our code is publicly available at https://github.com/TREE985/Treble-Counterfactual-VLMs.

Understanding how urban environments are perceived in terms of safety is crucial for urban planning and policymaking. Traditional methods like surveys are limited by high cost, required time, and scalability issues. To overcome these challenges, this study introduces Large Multimodal Models (LMMs), specifically Llava 1.6 7B, as a novel approach to assess safety perceptions of urban spaces using street-view images. In addition, the research investigated how this task is affected by different socio-demographic perspectives, simulated by the model through Persona-based prompts. Without additional fine-tuning, the model achieved an average F1-score of 59.21% in classifying urban scenarios as safe or unsafe, identifying three key drivers of perceived unsafety: isolation, physical decay, and urban infrastructural challenges. Moreover, incorporating Persona-based prompts revealed significant variations in safety perceptions across the socio-demographic groups of age, gender, and nationality. Elder and female Personas consistently perceive higher levels of unsafety than younger or male Personas. Similarly, nationality-specific differences were evident in the proportion of unsafe classifications ranging from 19.71% in Singapore to 40.15% in Botswana. Notably, the model's default configuration aligned most closely with a middle-aged, male Persona. These findings highlight the potential of LMMs as a scalable and cost-effective alternative to traditional methods for urban safety perceptions. While the sensitivity of these models to socio-demographic factors underscores the need for thoughtful deployment, their ability to provide nuanced perspectives makes them a promising tool for AI-driven urban planning.

Multimodal Large Language Models (MLLMs) are renowned for their superior instruction-following and reasoning capabilities across diverse problem domains. However, existing benchmarks primarily focus on assessing factual and logical correctness in downstream tasks, with limited emphasis on evaluating MLLMs' ability to interpret pragmatic cues and intermodal relationships. To address this gap, we assess the competency of MLLMs in performing Multimodal Discourse Analysis (MDA) using Coherence Relations. Our benchmark, CORDIAL, encompasses a broad spectrum of Coherence Relations across 3 different discourse domains at varying levels of granularity. Through our experiments on 10+ MLLMs employing different prompting strategies, we show that even top models like Gemini 1.5 Pro and GPT-4o fail to match the performance of simple classifier-based baselines. This study emphasizes the need to move beyond similarity-based metrics and adopt a discourse-driven framework for evaluating MLLMs, providing a more nuanced assessment of their capabilities. The benchmark and code are available at: https://github.com/aashish2000/CORDIAL.

We present Florence-VL, a new family of multimodal large language models (MLLMs) with enriched visual representations produced by Florence-2, a generative vision foundation model. Unlike the widely used CLIP-style vision transformer trained by contrastive learning, Florence-2 can capture different levels and aspects of visual features, which are more versatile to be adapted to diverse downstream tasks. We propose a novel feature-fusion architecture and an innovative training recipe that effectively integrates Florence-2's visual features into pretrained LLMs, such as Phi 3.5 and LLama 3. In particular, we propose "depth-breath fusion (DBFusion)" to fuse the visual features extracted from different depths and under multiple prompts. Our model training is composed of end-to-end pretraining of the whole model followed by finetuning of the projection layer and the LLM, on a carefully designed recipe of diverse open-source datasets that include high-quality image captions and instruction-tuning pairs. Our quantitative analysis and visualization of Florence-VL's visual features show its advantages over popular vision encoders on vision-language alignment, where the enriched depth and breath play important roles. Florence-VL achieves significant improvements over existing state-of-the-art MLLMs across various multi-modal and vision-centric benchmarks covering general VQA, perception, hallucination, OCR, Chart, knowledge-intensive understanding, etc. To facilitate future research, our models and the complete training recipe are open-sourced. https://github.com/JiuhaiChen/Florence-VL

Multimodal models typically combine a powerful large language model (LLM) with a vision encoder and are then trained on multimodal data via instruction tuning. While this process adapts LLMs to multimodal settings, it remains unclear whether this adaptation compromises their original language reasoning capabilities. In this work, we explore the effects of multimodal instruction tuning on language reasoning performance. We focus on LLaVA, a leading multimodal framework that integrates LLMs such as Vicuna or Mistral with the CLIP vision encoder. We compare the performance of the original LLMs with their multimodal-adapted counterparts across eight language reasoning tasks. Our experiments yield several key insights. First, the impact of multimodal learning varies between Vicuna and Mistral: we observe a degradation in language reasoning for Mistral but improvements for Vicuna across most tasks. Second, while multimodal instruction learning consistently degrades performance on mathematical reasoning tasks (e.g., GSM8K), it enhances performance on commonsense reasoning tasks (e.g., CommonsenseQA). Finally, we demonstrate that a training-free model merging technique can effectively mitigate the language reasoning degradation observed in multimodal-adapted Mistral and even improve performance on visual tasks.

Self-improvement in multimodal large language models (MLLMs) is crucial for enhancing their reliability and robustness. However, current methods often rely heavily on MLLMs themselves as judges, leading to high computational costs and potential pitfalls like reward hacking and model collapse. This paper introduces a novel, model-level judge-free self-improvement framework. Our approach employs a controlled feedback mechanism while eliminating the need for MLLMs in the verification loop. We generate preference learning pairs using a controllable hallucination mechanism and optimize data quality by leveraging lightweight, contrastive language-image encoders to evaluate and reverse pairs when necessary. Evaluations across public benchmarks and our newly introduced IC dataset designed to challenge hallucination control demonstrate that our model outperforms conventional techniques. We achieve superior precision and recall with significantly lower computational demands. This method offers an efficient pathway to scalable self-improvement in MLLMs, balancing performance gains with reduced resource requirements.

Large vision-language models (VLLMs) exhibit promising capabilities for processing multi-modal tasks across various application scenarios. However, their emergence also raises significant data security concerns, given the potential inclusion of sensitive information, such as private photos and medical records, in their training datasets. Detecting inappropriately used data in VLLMs remains a critical and unresolved issue, mainly due to the lack of standardized datasets and suitable methodologies. In this study, we introduce the first membership inference attack (MIA) benchmark tailored for various VLLMs to facilitate training data detection. Then, we propose a novel MIA pipeline specifically designed for token-level image detection. Lastly, we present a new metric called MaxR\'enyi-K%, which is based on the confidence of the model output and applies to both text and image data. We believe that our work can deepen the understanding and methodology of MIAs in the context of VLLMs. Our code and datasets are available at https://github.com/LIONS-EPFL/VL-MIA.

The conventional targeted adversarial attacks add a small perturbation to an image to make neural network models estimate the image as a predefined target class, even if it is not the correct target class. Recently, for visual-language models (VLMs), the focus of targeted adversarial attacks is to generate a perturbation that makes VLMs answer intended target text outputs. For example, they aim to make a small perturbation on an image to make VLMs' answers change from "there is an apple" to "there is a baseball." However, answering just intended text outputs is insufficient for tricky questions like "if there is a baseball, tell me what is below it." This is because the target of the adversarial attacks does not consider the overall integrity of the original image, thereby leading to a lack of visual reasoning. In this work, we focus on generating targeted adversarial examples with visual reasoning against VLMs. To this end, we propose 1) a novel adversarial attack procedure -- namely, Replace-then-Perturb and 2) a contrastive learning-based adversarial loss -- namely, Contrastive-Adv. In Replace-then-Perturb, we first leverage a text-guided segmentation model to find the target object in the image. Then, we get rid of the target object and inpaint the empty space with the desired prompt. By doing this, we can generate a target image corresponding to the desired prompt, while maintaining the overall integrity of the original image. Furthermore, in Contrastive-Adv, we design a novel loss function to obtain better adversarial examples. Our extensive benchmark results demonstrate that Replace-then-Perturb and Contrastive-Adv outperform the baseline adversarial attack algorithms. We note that the source code to reproduce the results will be available.