TRACE is a new framework that reorganizes 168 tasks for text‑vision embedding models into General, Adaptability, and Challenge categories. Experiments on two models show that adding general and adaptation data boosts Adaptability scores by 7.67 and 7.09 points while keeping General performance largely unchanged, yet Challenge performance stays below 5.0 with minimal gains from more data. The study highlights that data adaptation alone cannot overcome weaknesses in multilingual and visual‑document tasks, indicating a need for methodological advances.
arXiv:2607. 07179v1 Announce Type: cross Abstract: Document Visual Question Answering (DocVQA) presents a complex multimodal challenge, requiring models to exploit visual, textual, and layout information from documents.
Document Visual Question Answering (DocVQA) presents a complex multimodal challenge, requiring models to exploit visual, textual, and layout information from documents. Although Vision-Language Models (VLMs) have shown remarkable performance in text-vision tasks, their robustness and transferability to different document domains remains underexplored.
arXiv:2605.28190v2 Announce Type: replace Abstract: Embedding benchmarks like MTEB report a single score per model, implicitly treating robustness as a static, scalar property. We argue that embeddin...
The paper introduces Vision-Free Adaptation (VFA), a method that separates multilingual language enhancement from visual alignment in multimodal large language models. VFA fine‑tunes a base LLM on multilingual text to create a multilingual task vector, which is then merged with the vision‑aligned task vector of an existing MLLM. Experiments on five MLLMs and six multilingual benchmarks show consistent gains while preserving multimodal and text‑only performance, and using less than 2% of text data narrows the performance gap to fully multimodal‑trained models.
arXiv:2608. 05785v1 Announce Type: cross Abstract: Multilingual text embedding models are commonly adapted using a single training objective across diverse tasks, despite different tasks requiring fundamentally different optimization strategies.
The paper introduces the Generative Embedding Benchmark (GEB), which evaluates how much content from an embedding can be recovered by a decoder that only has access to the frozen embedding and a question, without the original image or intermediate features. GEB uses a curated visual‑question‑answering dataset with 1,800 development and 900 test items covering natural images, scene text, and visual documents. Experiments on seven public embedding models show that visual‑only scores range from 28.25 to 33.21, while joint image‑question encoding boosts scores up to 65.56, revealing that generative readout uncovers information bottlenecks not captured by traditional separability‑based benchmarks.
arXiv:2604. 02327v2 Announce Type: replace-cross Abstract: Pretrained Vision Transformers (ViTs) such as DINOv2 and MAE provide generic image features that can be applied to a variety of downstream tasks such as retrieval, classification, and segmentation.
The paper introduces PM4Bench, a multimodal, multilingual, multi-task benchmark built on a strictly parallel 10‑language corpus, allowing fair cross‑lingual comparison of Large Vision‑Language Models (LVLMs). It also proposes a vision setting that embeds textual inputs directly into images to better mimic real deployment scenarios. Experiments show OCR performance drives cross‑lingual gaps, leading to an OCR‑centric GRPO training strategy that improves multilingual VQA and reduces disparities without costly supervision.
Vision-language models exhibit remarkable zero-shot capabilities but suffer significant performance degradation under distribution shifts. While test-time adaptation (TTA) via Low-Rank Adaptation offers a parameter-efficient solution, we identify a fundamental bottleneck in current methods: the reliance on static rank configurations.
The study investigates how unified vision‑language models (VLMs) can simultaneously support visual understanding and generation. Using controlled benchmarks (SmartWatch and modified CelebA) that pair VQA, captioning, and text‑to‑image tasks, the authors evaluate several LLM‑based architectures built on SigLIP and VQ‑VAE visual spaces. Results show that mixed training can improve both understanding and generation, but the gains depend on how well the visual input and output spaces are aligned; misaligned or distorted visual spaces can weaken or reverse these benefits. The paper also demonstrates that balancing data across tasks and controlling attribute frequencies can help recover underrepresented visual concepts, and that the transfer is driven more by the base language model’s learned relationships than by visual adapters.
arXiv:2606. 06943v1 Announce Type: cross Abstract: Vision-language models (VLMs) such as CLIP achieve strong zero-shot recognition but remain highly fragile under adversarial perturbations.
High benchmark accuracy does not guarantee genuine use of visual evidence. We study this problem in traffic accident Video Question Answering (VideoQA), where correct answers should depend on scene-specific visual evidence but may instead be inferred from textual shortcuts.