LLM Robustness Analysis Shows Varying Responses to Scientific Skepticism

Testing across Llama-3.1-8B, Qwen2.5-7B, and Mistral-7B reveals that models do not sycophantically retreat from scientific consensus. Instead, they adopt distinct behaviors: reactive assertion, surface hedging, or non-response when faced with user doubt.