MATH · IN · MODELS

InternVL

Shanghai AI Lab / OpenGVLab

Structures found in this family (4)

Linear Separability, Anisotropy, Linear Direction, Linear Subspace

By model (4)

InternVL3-8B · 8B
Linear Separability, Anisotropy, Linear Direction, Linear Subspace
InternVL2.5 8B · 8B
Linear Direction
InternVL3.5-2B · 2B
Linear Separability
InternVL3.5-8B-Instruct · 8B
Linear Subspace

Observations (8)

A per-head Fisher-ratio score selects steering that nearly matches fine-tuning
MLLM visual tokens split into sink, dead, alive clusters
A contrastive spatial-ID direction is causally bound into object tokens across 11 real VLMs
VLMs encode chart signal linearly but do not route it to decisions
Lightweight VLMs linearly encode object presence but never orientation
Centroid-based concept vectors for colors and color-shape conjunctions causally steer perception in three open-weight VLMs, while naive probe directions largely fail
Six Ekman emotion vectors in vision-language models have a low-dimensional geometry (PC1+PC2 explain 55-75% of variance), and a Procrustes-aligned consensus emotion subspace across three VLMs retains selective steering and transfers to held-out Gemma-3
SVD-based cross-scene averaging isolates a linear 3D-position subspace in real VLMs

Papers

Steering and Rectifying Latent Representation Manifolds in Frozen Multi-modal LLMs for Video Anomaly Detection (2026), What Do Visual Tokens Really Encode? Uncovering Sparsity and Redundancy in Multimodal Large Language Models (2026), Linear Mechanisms for Spatiotemporal Reasoning in Vision Language Models (2026), Encoded but Not Routed: Explaining the Table-Chart Gap in Scientific Claim Verification (2026), Probing Visual Concepts in Lightweight Vision-Language Models for Automated Driving (2026), The Geometry of Representational Failures in Vision Language Models (2026), Do Emotion Concepts Generalize across Sources, Modalities, and Architectures in Vision-Language Models? (2026), Uncovering and Shaping the Latent Representation of 3D Scene Topology in Vision-Language Models (2026)