safetyfactbullishVision-language guardrails do not need chain-of-thought reasoning to effectively detect response harmfulnessComputer Vision28 Jul 2026http://arxiv.org/abs/2607.21401v1