arXiv Artificial Intelligence

CASA: Classification Augmented with Safety Attention for Robust Multimodal Alignment

3 hours ago 10.08.2026 07:00 👁 2 views
CASA: Classification Augmented with Safety Attention for Robust Multimodal Alignment

Quick summary

arXiv:2604.00310v2 Announce Type: replace-cross Abstract: Multimodal large-language models (MLLMs) often experience degraded safety alignment when harmful queries exploit cross-modal interactions. Models aligned on text alone show a higher rate of successful attacks when extended to two or more modalities. We propose a simple conditional decoding strategy, CASA (Classification Augmented with Safety Attention) that uses internal representations of MLLMs to predict a binary safety token before response generation. The classification is guided by a safety attention mechanism that computes attenti

Key takeaways

  • arXiv:2604.00310v2 Announce Type: replace-cross Abstract: Multimodal large-language models (MLLMs) often experience degraded safety alignment when harmful queries exploit cross-modal interactions.
  • Models aligned on text alone show a higher rate of successful attacks when extended to two or more modalities.
  • We propose a simple conditional decoding strategy, CASA (Classification Augmented with Safety Attention) that uses internal representations of MLLMs to predict a binary safety token before response generation.

Why it matters

This development is a reminder to test misuse and data-leak scenarios alongside speed and quality. Trust should come from testable controls and clear failure reporting, not protection claims alone.

Kaynak sitede devamını oku: arXiv Artificial Intelligence ↗