vision-language models

2026-10-27

RESEARCH · 1 SOURCE · arXiv cs.AI

IRON-Meme benchmark released for structured understanding of identities and roles in political memes

The authors introduce IRON-Meme, a human-annotated benchmark of 1,110 political memes built on a social-science grounded framework that labels social identities, viewpoints (ingroup/outgroup), and narrative roles (moral character, intelligence, status, power). They evaluate nine vision-language models, find frontier-scale and reasoning models score highest, note poorer accuracy on infrequent labels, and report that adding template background knowledge does not improve narrative role classification; the dataset and code are publicly released.

6.0

2026-10-15

RESEARCH · 1 SOURCE · arXiv cs.AI

MOSAIC: sparse multimodal latent attribution for vision–language models (arXiv:2610.14190v1)

The paper introduces MOSAIC (Multimodal Orthogonalized Sparse Autoencoders for Influence Characterization), which learns a shared sparse image–text latent space with contiguous, phrase-aligned visual concepts and applies EK-FAC to decompose training-example influence through those concepts to specific phrases and image regions. The authors evaluate MOSAIC on Qwen and Gemma for captioning, chest X‑ray understanding, and visual question answering, reporting more localized and coherent features and that ablating influence-ranked latents alters outputs more than ablating latents ranked by activation or frequency.

7.0

2026-10-07

RESEARCH · 1 SOURCE · arXiv cs.AI

MOTIVE: multi-view self-verification framework for vision-language models

The arXiv preprint (arXiv:2610.07018v1) introduces MOTIVE, a Multi-View Self-Verification framework that evaluates candidate answers from complementary verification perspectives and learns a correctness-aligned reliability score to decide whether to accept an answer or trigger history-guided rethinking. Experiments across diverse multimodal benchmarks and VLM backbones show MOTIVE improves over strong self-verification and self-correction baselines, reducing unnecessary reasoning turns while increasing answer reliability.

6.0

2026-08-12

MODELS · 1 SOURCE · Microsoft Research

Microsoft Research publishes MindTopo benchmark for VLM spatial/topological reasoning

Microsoft Research introduced MindTopo, a new benchmark designed to test how vision-language models (VLMs) understand topological and spatial relationships. The announcement frames MindTopo as a tool to evaluate and guide improvements in spatial reasoning and planning capabilities of VLMs.

6.0

2026-08-11

RESEARCH · 1 SOURCE · Microsoft Research

Introducing CARE-X: a unified approach for clinically useful radiology VLMs

Microsoft Research published CARE-X, a research proposal for radiology vision-language models (VLMs) that combines auxiliary supervision, reward-aligned learning, and tool-augmented measurement to enable flexible reasoning, calibrated predictions, and measurement-based tools for chest X-ray interpretation.

7.0