The paper investigates how evaluation protocols and ground‑truth quality affect the classification performance of Multimodal Large Language Models (MLLMs). It identifies and corrects key issues such as discarded out‑of‑list outputs, inflated distractor choices, and poor open‑world mapping, and shows that design choices like batch size, image ordering, and text encoder selection significantly influence accuracy. Using a multilabel reannotation of 625 ImageNet‑1k classes (ReGT), the study finds that corrected labels can boost MLLM performance by up to 10.8%, narrowing the gap with supervised models, and demonstrates that MLLMs can assist human annotators in about half of difficult cases.
arXiv:2606. 28551v1 Announce Type: cross Abstract: Building performant Vision-Language Models (VLMs) requires carefully curating large-scale training datasets, yet the community lacks systematic benchmarks for evaluating such curation strategies.
arXiv:2609.06245v1 Announce Type: cross Abstract: Multimodal Large Language Models (MLLMs) perform strongly on general visual understanding tasks such as visual question answering, yet they often str...
The paper introduces FindIt, the first comprehensive benchmark for evaluating the promptable localization abilities of generalist multimodal large language models (MLLMs). It covers four core task categories—object detection, referring expression detection, instance-level detection, and video-based detection—and provides a unified framework that standardizes inputs, enforces parsable bounding box outputs, and defines transparent evaluation protocols. Using this benchmark, the authors assess a range of open-source and proprietary MLLMs, revealing both their strengths and limitations, particularly their sensitivity to formatting constraints and difficulty generalizing to minor variations.
SAFIRE is a large-scale benchmark for fire and smoke understanding in multimodal large language models (MLLMs), featuring 83,000 captioned images across 20 scenarios and 193,000 multiple-choice VQA questions derived from a 9.7K-image subset. The benchmark evaluates 10 dimensions of performance, from basic perception to higher-order reasoning, and employs a GPT‑5.4-assisted verification pipeline to ensure annotation quality. Experiments on ten open-source MLLMs (8B–38B) reveal an average accuracy of 61.9%, highlighting significant gaps in safety-critical reasoning, while fine-tuning vision encoders on just 7% of SAFIRE data boosts fire-scene classification accuracy from 20.1% to 64.5%. All resources are publicly available at https://risys-lab.github.io/SAFIRE/.
The paper introduces MOSAIC, a method for personalized federated vision‑language models that addresses domain heterogeneity by focusing on class‑specific cross‑domain residuals. It constructs a decision‑aware harmfulness score to identify residuals that hurt image‑text decision margins, then uses a low‑rank residual adapter with shared class factors and private domain factors, an image‑conditioned gate, and harmful‑pair‑aware reweighting to refine local updates. Experiments on Office31, OfficeHome, and DomainNet100 show consistent improvements in macro‑client top‑1 accuracy across various domain‑shift scenarios.
The paper introduces Percept-V, a dataset of 6,000 program-generated images across 30 domains designed to test simple visual perception skills from the TVPS-4 framework. Experiments show that state‑of‑the‑art multimodal large language models perform poorly compared to humans, especially as image complexity increases, and that fine‑tuning yields only limited generalization to related datasets. The study highlights specific perception skills that remain challenging for current models.
arXiv:2605.26380v2 Announce Type: replace-cross Abstract: Frontier multimodal large language models (MLLMs) have been reported to achieve over 90\% accuracy on fine-grained perception benchmarks. How...
arXiv:2608.28696v1 Announce Type: new Abstract: Visual in-context learning (ICL) with multimodal large language models (MLLMs) is effective for fine-grained visual classification, but each retrieved...
Vision-Language Models (VLMs) such as CLIP are now foundational to multimodal systems, yet their robustness to spurious correlations remains poorly understood at scale. We present the first large-scale empirical study of 194 publicly available VLMs, including 16 model families, covering a wide range of model sizes, 24 training datasets, and three evaluation benchmarks, namely ImageNet (overall performance), CelebA (typical single-attribute bias), and UrbanCars (complex multi-attribute biases).
arXiv:2607. 18615v1 Announce Type: cross Abstract: Machine unlearning for vision-language models (VLMs) remains underexplored.
arXiv:2606. 08970v1 Announce Type: new Abstract: Vision-language models (VLMs) with varying performance and resource requirements are widely deployed, making it difficult for users to select the most appropriate one among numerous VLM candidates.
arXiv:2609.01027v1 Announce Type: new Abstract: Out-of-distribution (OOD) detection predicts whether a test image belongs to none of the predefined classes. To evaluate this task, benchmarks need ima...