Existing methods like ensemble still struggle to aggregate these models with the desired higher-order communications. In this work, we propose Cola, a novel paradigm that coordinates multiple VLMs for visual reasoning.

Sparse mixture-of-experts are domain generalizable learners.

Existing methods like ensemble still struggle to aggregate these models with the desired higher-order communications. In this work, we propose Cola, a novel paradigm that coordinates multiple VLMs for visual reasoning.

Vision-Language Models (VLMs) are advancing multimodal AI, yet their performance consistency across tasks is underexamined. We benchmark CLIP, BLIP, and LXMERT across diverse datasets spanning retrieval, captioning, and reasoning. Our evaluation includes task accuracy, generation quality, efficiency, and a novel Cross-Dataset Consistency (CDC) metric. CLIP shows strongest generalization (CDC: 0.92), BLIP excels on curated data, and LXMERT leads in structured reasoning. These results expose trade-offs between generalization and specialization, informing industrial deployment of VLMs and guiding development toward robust, task-flexible architectures.

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.

Vision language models (VLMs) are designed to extract relevant visuospatial information from images. Some research suggests that VLMs can exhibit humanlike scene understanding, while other investigations reveal difficulties in their ability to process relational information. To achieve widespread applicability, VLMs must perform reliably, yielding comparable competence across a wide variety of related tasks. We sought to test how reliable these architectures are at engaging in trivial spatial cognition, e.g., recognizing whether one object is left of another in an uncluttered scene. We developed a benchmark dataset -- TableTest -- whose images depict 3D scenes of objects arranged on a table, and used it to evaluate state-of-the-art VLMs. Results show that performance could be degraded by minor variations of prompts that use logically equivalent descriptions. These analyses suggest limitations in how VLMs may reason about spatial relations in real-world applications. They also reveal novel opportunities for bolstering image caption corpora for more efficient training and testing.

Image Captioning for state-of-the-art VLMs has significantly improved over time; however, this comes at the cost of increased computational complexity, making them less accessible for resource-constrained applications such as mobile devices and assistive technologies. Alternatively, smaller VLMs prioritize high-level scene descriptions, overlooking finer details that contribute to a richer understanding of an image. In this paper, we introduce a training-free framework that enhances caption diversity and informativeness by explicitly attending to distinct image regions using a comparably small VLM, BLIP, as the backbone. Our approach leverages structured segmentation to produce hierarchical representations that capture both global and localized semantics. Without requiring additional model training, we demonstrate that our method allows smaller VLMs to achieve performance comparable to larger models in terms of image-caption alignment, semantic integrity, and diversity. We evaluate our framework on MSCOCO, Flickr30k, and Nocaps test datasets, achieving a Div-2 score of 0.735, 0.750, and 0.748 for each dataset respectively, while maintaining strong image-caption relevancy and semantic integrity with the human-annotated captions.

Generative models are nowadays widely used to generate graphical content used for multiple purposes, e.g. web, art, advertisement. However, it has been shown that the images generated by these models could reinforce societal biases already existing in specific contexts. In this paper, we focus on understanding if this is the case when one generates images related to various software engineering tasks. In fact, the Software Engineering (SE) community is not immune from gender and ethnicity disparities, which could be amplified by the use of these models. Hence, if used without consciousness, artificially generated images could reinforce these biases in the SE domain. Specifically, we perform an extensive empirical evaluation of the gender and ethnicity bias exposed by three versions of the Stable Diffusion (SD) model (a very popular open-source text-to-image model) - SD 2, SD XL, and SD 3 - towards SE tasks. We obtain 6,720 images by feeding each model with two sets of prompts describing different software-related tasks: one set includes the Software Engineer keyword, and one set does not include any specification of the person performing the task. Next, we evaluate the gender and ethnicity disparities in the generated images. Results show how all models are significantly biased towards male figures when representing software engineers. On the contrary, while SD 2 and SD XL are strongly biased towards White figures, SD 3 is slightly more biased towards Asian figures. Nevertheless, all models significantly under-represent Black and Arab figures, regardless of the prompt style used. The results of our analysis highlight severe concerns about adopting those models to generate content for SE tasks and open the field for future research on bias mitigation in this context.

Incorporating automatically predicted human feedback into the process of training generative models has attracted substantial recent interest, while feedback at inference time has received less attention. The typical feedback at training time, i.e., preferences of choice given two samples, does not naturally transfer to the inference phase. We introduce a novel type of feedback -- caption reformulations -- and train models to mimic reformulation feedback based on human annotations. Our method does not require training the image captioning model itself, thereby demanding substantially less computational effort. We experiment with two types of reformulation feedback: first, we collect a dataset of human reformulations that correct errors in the generated captions. We find that incorporating reformulation models trained on this data into the inference phase of existing image captioning models results in improved captions, especially when the original captions are of low quality. We apply our method to non-English image captioning, a domain where robust models are less prevalent, and gain substantial improvement. Second, we apply reformulations to style transfer. Quantitative evaluations reveal state-of-the-art performance on German image captioning and English style transfer, while human validation with a detailed comparative framework exposes the specific axes of improvement.