MoE router logits are a free predictive safety signal for vision-language models
shangbinfeng · x · 2026-10-24
- Researchers found that router logits in MoE VLMs are highly predictive safety signals: a linear probe on routing states during prefill flags unsafe image+text requests before generation, without touching weights or routing.
- Motivated by compositional risk—individually benign images and questions can be harmful together. The figure contrasts this with prompting, SFT, and expert steering, which are inconsistent across models or trade safety for over-refusal.
Related event: Reading Router Logits: Lightweight Multimodal Safety Detection for MoE VLMs(5 posts)→
More from Safety
- Python packages vulnerable to pre-auth RCE via node2shell CVE-2026-111497 through Brotli — cyb3rops · 2026-10-24
- Zvi's 'The Hackening': experts predict AI-driven internet hostility within 6-12 months — TheZvi · 2026-10-24
- OpenAI fired-researchers saga: METR terms in flux, leak accusations denied — sjgadler · 2026-10-24
- Security teams increasingly run open models via Ollama for red teaming — mchiang0610 · 2026-10-24
- France rebuilt its tax code as MLang, a formally verified open-source compiler — fkasummer · 2026-10-24
- AI cyberattacks cost $25-50 per attempt, recovery costs millions — OmarUFlorez · 2026-10-24