SOURCE-LINKED INTELLIGENCE

Understanding In-Context Multimodal Jailbreaks via Posterior Reweighting

arXiv · AI, language, vision and robotics · article · Sep 8, 2026 · UTC

In-context learning (ICL) jailbreaks reveal a critical vulnerability in multimodal large language models (MLLMs): harmful demonstrations in the prompt can induce unsafe outputs without modifying model parameters. Despite extensive empirical evidence, existing work lacks a principled understanding of why such jailbreaks reliably succeed or how their effectiveness scales with context composition. We propose a posterior reweighting framework that models a safety-aligned MLLM as implicitly operating over competing behavioral modes, and interprets in-context demonstrations as inference-time evidenc

Read original source ↗ Open in workspace

recordType

paper

region

Global