safetyfactneutralHigh-temperature sampling increases diversity in LLM outputs but weakens model guardrails by reducing refusal responses to harmful promptsComputation and Language28 Jul 2026http://arxiv.org/abs/2607.20791v1