Decodable but Misrouted: Sparse Features Uncover a Readout Gap in Vision-Language Models for Harmful Meme Detection

Paper recorded by Signals 4 on 2026-09-16 in cs.AI. Abstract reproduced from arXiv; link to the original below.

Published 2026-09-16 on arXiv · recorded by Signals 4 on 2026-09-17

Category: cs.AI · 人工智能 · first seen 2026-09-17

Abstract

When a large vision-language model misclassifies a harmful meme, the failure may reflect missing internal evidence or an inability to route represented evidence to its output. We distinguish these cases in Gemma-3 and Qwen3.5 using sparse autoencoders, role-conditioned probes, causal interventions, and recovery experiments across six harmful content benchmarks, with additional Spanish and Hindi-En

Read on arXiv →