Abstract:Recent progress in Multimodal Large Language Models (MLLMs) has highlighted the critical roles of both the visual backbone and the underlying language model. While prior work has primarily focused on scaling these components to billions of parameters, the trade-offs between model size, architecture, and performance remain underexplored. Additionally, inconsistencies in training data and evaluation protocols have hindered direct comparisons, making it difficult to derive optimal design choices. In this paper, we introduce LLaVA-MORE, a new family of MLLMs that integrates recent language models with diverse visual backbones. To ensure fair comparisons, we employ a unified training protocol applied consistently across all architectures. Our analysis systematically explores both small- and medium-scale LLMs -- including Phi-4, LLaMA-3.1, and Gemma-2 -- to evaluate multimodal reasoning, generation, and instruction following, while examining the relationship between model size and performance. Beyond evaluating the LLM impact on final results, we conduct a comprehensive study of various visual encoders, ranging from CLIP-based architectures to alternatives such as DINOv2, SigLIP, and SigLIP2. Additional experiments investigate the effects of increased image resolution and variations in pre-training datasets. Overall, our results provide insights into the design of more effective MLLMs, offering a reproducible evaluation framework that facilitates direct comparisons and can guide future model development. Our source code and trained models are publicly available at: https://github.com/aimagelab/LLaVA-MORE.
Abstract:Addressing the retrieval of unsafe content from vision-language models such as CLIP is an important step towards real-world integration. Current efforts have relied on unlearning techniques that try to erase the model's knowledge of unsafe concepts. While effective in reducing unwanted outputs, unlearning limits the model's capacity to discern between safe and unsafe content. In this work, we introduce a novel approach that shifts from unlearning to an awareness paradigm by leveraging the inherent hierarchical properties of the hyperbolic space. We propose to encode safe and unsafe content as an entailment hierarchy, where both are placed in different regions of hyperbolic space. Our HySAC, Hyperbolic Safety-Aware CLIP, employs entailment loss functions to model the hierarchical and asymmetrical relations between safe and unsafe image-text pairs. This modelling, ineffective in standard vision-language models due to their reliance on Euclidean embeddings, endows the model with awareness of unsafe content, enabling it to serve as both a multimodal unsafe classifier and a flexible content retriever, with the option to dynamically redirect unsafe queries toward safer alternatives or retain the original output. Extensive experiments show that our approach not only enhances safety recognition but also establishes a more adaptable and interpretable framework for content moderation in vision-language models. Our source code is available at https://github.com/aimagelab/HySAC.
Abstract:Cross-modal retrieval is gaining increasing efficacy and interest from the research community, thanks to large-scale training, novel architectural and learning designs, and its application in LLMs and multimodal LLMs. In this paper, we move a step forward and design an approach that allows for multimodal queries, composed of both an image and a text, and can search within collections of multimodal documents, where images and text are interleaved. Our model, ReT, employs multi-level representations extracted from different layers of both visual and textual backbones, both at the query and document side. To allow for multi-level and cross-modal understanding and feature extraction, ReT employs a novel Transformer-based recurrent cell that integrates both textual and visual features at different layers, and leverages sigmoidal gates inspired by the classical design of LSTMs. Extensive experiments on M2KR and M-BEIR benchmarks show that ReT achieves state-of-the-art performance across diverse settings. Our source code and trained models are publicly available at https://github.com/aimagelab/ReT.
Abstract:Video Question Answering (Video QA) is a challenging video understanding task that requires models to comprehend entire videos, identify the most relevant information based on contextual cues from a given question, and reason accurately to provide answers. Recent advancements in Multimodal Large Language Models (MLLMs) have transformed video QA by leveraging their exceptional commonsense reasoning capabilities. This progress is largely driven by the effective alignment between visual data and the language space of MLLMs. However, for video QA, an additional space-time alignment poses a considerable challenge for extracting question-relevant information across frames. In this work, we investigate diverse temporal modeling techniques to integrate with MLLMs, aiming to achieve question-guided temporal modeling that leverages pre-trained visual and textual alignment in MLLMs. We propose T-Former, a novel temporal modeling method that creates a question-guided temporal bridge between frame-wise visual perception and the reasoning capabilities of LLMs. Our evaluation across multiple video QA benchmarks demonstrates that T-Former competes favorably with existing temporal modeling approaches and aligns with recent advancements in video QA.
Abstract:Recent work has empirically shown that Vision-Language Models (VLMs) struggle to fully understand the compositional properties of the human language, usually modeling an image caption as a "bag of words". As a result, they perform poorly on compositional tasks, which require a deeper understanding of the different entities of a sentence (subject, verb, etc.) jointly with their mutual relationships in order to be solved. In this paper, we model the dependency relations among textual and visual tokens using a Causal Graphical Model (CGM), built using a dependency parser, and we train a decoder conditioned by the VLM visual encoder. Differently from standard autoregressive or parallel predictions, our decoder's generative process is partially-ordered following the CGM structure. This structure encourages the decoder to learn only the main causal dependencies in a sentence discarding spurious correlations. Using extensive experiments on five compositional benchmarks, we show that our method significantly outperforms all the state-of-the-art compositional approaches by a large margin, and it also improves over methods trained using much larger datasets.
Abstract:The task of image captioning demands an algorithm to generate natural language descriptions of visual inputs. Recent advancements have seen a convergence between image captioning research and the development of Large Language Models (LLMs) and Multimodal LLMs -- like GPT-4V and Gemini -- which extend the capabilities of text-only LLMs to multiple modalities. This paper investigates whether Multimodal LLMs can supplant traditional image captioning networks by evaluating their performance on various image description benchmarks. We explore both the zero-shot capabilities of these models and their adaptability to different semantic domains through fine-tuning methods, including prompt learning, prefix tuning, and low-rank adaptation. Our results demonstrate that while Multimodal LLMs achieve impressive zero-shot performance, fine-tuning for specific domains while maintaining their generalization capabilities intact remains challenging. We discuss the implications of these findings for future research in image captioning and the development of more adaptable Multimodal LLMs.
Abstract:Open-Vocabulary Segmentation (OVS) aims at segmenting images from free-form textual concepts without predefined training classes. While existing vision-language models such as CLIP can generate segmentation masks by leveraging coarse spatial information from Vision Transformers, they face challenges in spatial localization due to their global alignment of image and text features. Conversely, self-supervised visual models like DINO excel in fine-grained visual encoding but lack integration with language. To bridge this gap, we present Talk2DINO, a novel hybrid approach that combines the spatial accuracy of DINOv2 with the language understanding of CLIP. Our approach aligns the textual embeddings of CLIP to the patch-level features of DINOv2 through a learned mapping function without the need to fine-tune the underlying backbones. At training time, we exploit the attention maps of DINOv2 to selectively align local visual patches with textual embeddings. We show that the powerful semantic and localization abilities of Talk2DINO can enhance the segmentation process, resulting in more natural and less noisy segmentations, and that our approach can also effectively distinguish foreground objects from the background. Experimental results demonstrate that Talk2DINO achieves state-of-the-art performance across several unsupervised OVS benchmarks. Source code and models are publicly available at: https://lorebianchi98.github.io/Talk2DINO/.
Abstract:Multimodal LLMs (MLLMs) are the natural extension of large language models to handle multimodal inputs, combining text and image data. They have recently garnered attention due to their capability to address complex tasks involving both modalities. However, their effectiveness is limited to the knowledge acquired during training, which restricts their practical utility. In this work, we introduce a novel method to enhance the adaptability of MLLMs by integrating external knowledge sources. Our proposed model, Reflective LLaVA (ReflectiVA), utilizes reflective tokens to dynamically determine the need for external knowledge and predict the relevance of information retrieved from an external database. Tokens are trained following a two-stage two-model training recipe. This ultimately enables the MLLM to manage external knowledge while preserving fluency and performance on tasks where external knowledge is not needed. Through our experiments, we demonstrate the efficacy of ReflectiVA for knowledge-based visual question answering, highlighting its superior performance compared to existing methods. Source code and trained models are publicly available at https://github.com/aimagelab/ReflectiVA.
Abstract:In the last years, the research interest in visual navigation towards objects in indoor environments has grown significantly. This growth can be attributed to the recent availability of large navigation datasets in photo-realistic simulated environments, like Gibson and Matterport3D. However, the navigation tasks supported by these datasets are often restricted to the objects present in the environment at acquisition time. Also, they fail to account for the realistic scenario in which the target object is a user-specific instance that can be easily confused with similar objects and may be found in multiple locations within the environment. To address these limitations, we propose a new task denominated Personalized Instance-based Navigation (PIN), in which an embodied agent is tasked with locating and reaching a specific personal object by distinguishing it among multiple instances of the same category. The task is accompanied by PInNED, a dedicated new dataset composed of photo-realistic scenes augmented with additional 3D objects. In each episode, the target object is presented to the agent using two modalities: a set of visual reference images on a neutral background and manually annotated textual descriptions. Through comprehensive evaluations and analyses, we showcase the challenges of the PIN task as well as the performance and shortcomings of currently available methods designed for object-driven navigation, considering modular and end-to-end agents.
Abstract:Despite significant advancements in caption generation, existing evaluation metrics often fail to capture the full quality or fine-grained details of captions. This is mainly due to their reliance on non-specific human-written references or noisy pre-training data. Still, finding an effective metric is crucial not only for captions evaluation but also for the generation phase. Metrics can indeed play a key role in the fine-tuning stage of captioning models, ultimately enhancing the quality of the generated captions. In this paper, we propose PAC-S++, a learnable metric that leverages the CLIP model, pre-trained on both web-collected and cleaned data and regularized through additional pairs of generated visual and textual positive samples. Exploiting this stronger and curated pre-training, we also apply PAC-S++ as a reward in the Self-Critical Sequence Training (SCST) stage typically employed to fine-tune captioning models. Extensive experiments on different image and video datasets highlight the effectiveness of PAC-S++ compared to popular metrics for the task, including its sensitivity to object hallucinations. Furthermore, we show that integrating PAC-S++ into the fine-tuning stage of a captioning model results in semantically richer captions with fewer repetitions and grammatical errors. Evaluations on out-of-domain benchmarks further demonstrate the efficacy of our fine-tuning approach in enhancing model capabilities. Source code and trained models are publicly available at: https://github.com/aimagelab/pacscore.