Hostile prompts cut LLM performance across scales

Scaling isn't the silver bullet for model robustness; if your user is mean, your model is likely to fail.

• –The effect is universal across architecture (Dense vs MoE) and quantization (FP16 vs Q4), indicating it is a fundamental property of current LLM training paradigms.
• –Larger models like Mistral Large 123B show attenuation but remain significantly vulnerable, debunking the idea that simply adding parameters cures sensitivity.
• –Instruction tuning actually amplifies hostility sensitivity in models like Llama 3.1, raising questions about how RLHF and safety training impact behavioral stability.
• –The emergence of extreme position bias in specific configurations (like Mistral 7B Q4) under hostile framing suggests quantization can cause unpredictable distributional collapses.