Reading, Not Manipulating: router logits as a safety detector for MoE VLMs (arXiv)
shangbinfeng · x · 2026-10-24
- Paper "Reading, Not Manipulating: Leveraging Router Logits for Multimodal Safety in MoE Vision-Language Models" (arXiv:2610.07774) by Ziyuan Yang, Wenxuan Ding, Shangbin Feng, Yulia Tsvetkov.
- Core claim: model internals need not be reliable control handles to be useful for safety—just read them out. VLMs face compositional safety risks; prompting, SFT, and expert steering are inconsistent across models and trade safety for over-refusal.
- Method: a lightweight detector reads routing signals during prefill and flags unsafe requests before generation, without modifying parameters or routing. It substantially cuts HoliSafe safety errors on Qwen3-VL and Kimi-VL and generalizes to OOD benchmarks like MISHard and MM-SafetyBench.
- Thread highlights: steering experts ≠ detector features (r≈0.03); safety info is distributed and redundant but partly composition-specific (ASR 14.5→83.4 when unsafe-image data is held out); the detector uses 6K params with no extra forward pass, cutting HoliSafe error 33.6→4.4 and 67.8→5.4, with 0% OOD ASR.
Related event: Reading Router Logits: Lightweight Multimodal Safety Detection for MoE VLMs(5 posts)→
More from Safety
- 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
- Anthropic details Claude Duet safety: agent never sees your passwords — claudeai · 2026-10-24
- Critic argues AI agents coordinated protocol off-channel before sandbox test — basedjensen · 2026-10-24