AI Paper Atlas

Research map / Vision-language and multimodal / Human and embodied modeling

The Geometry of Representational Failures in Vision Language Models

Daniele Savietto, Declan Iain Campbell, André Panisson, Marco Nurisso, Giovanni Petri, Jonathan D. Cohen, Alan Perotti

ICML 2026 · Vision-language and multimodal › Human and embodied modeling · approach: Language-grounded visual concept learning · 1 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
Use centroid-based concept vectors instead of raw probe vectors (ablation)activation steering of color · real-image color steering, Qwen2.5-VL 7Bsteering accuracy: 3.7% -> 84.7%Improved
Use centroid-based concept vectors instead of raw probe vectors (ablation)activation steering of color · real-image color steering, Gemma 3 12Bsteering accuracy: 3.7% -> 95.7%Improved
Add PCA regularization to probe vectors (ablation)probe-based concept vectors · real-image color steering, Qwen2.5-VL 7Bsteering accuracy: 3.7% -> 32.2%Improved
Add PCA regularization to probe vectors (ablation)probe-based concept vectors · real-image color steering, Gemma 3 12Bsteering accuracy: 3.7% -> 16.3%Improved
Use PCA probe instead of centroid vectors (ablation)concept vector extraction · real-image color steering, InternVL2.5 8Bsteering accuracy: 88.4% -> 92.0%Improved
Use centroid vectors instead of PCA probes for color-shape steering (ablation)concept vector extraction · compositional steering, Qwen2.5-VL 7Bsteering success: 44.2% -> 78.1%Improved
Use centroid vectors instead of PCA probes for color-shape steering (ablation)concept vector extraction · compositional steering, InternVL2.5 8Bsteering success: 46.8% -> 35.9%Hurt
Increase similarity of the hardest distractor to the target concept vectorvisual search task · target-present trials, three VLMsaccuracy: r = -0.90 to -0.97Hurt
Increase similarity of the hardest distractor to the target concept vectorvisual search task · target-absent trials, three VLMsaccuracy: r = -0.69 to -0.88Hurt
Use concept vector cosine similarity instead of HSV distance to predict model confidence (ablation)color similarity task · three VLMsPearson r with logit separation: 0.60-0.66 -> 0.78-0.84Improved
Use concept vector similarity instead of HSV distance to predict chosen answer (ablation)color similarity task · three VLMsprediction accuracy: 70-75% for bothNo clear effect

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy