What do LLMs think when you don't tell them what to think about?

Together AI · 5 months ago

Researchers studied what large language models generate when given minimal, topic-neutral prompts without explicit instructions, finding that different model families exhibit distinct topical preferences independent of task specification. GPT-OSS defaults to programming (27.1%) and mathematics (24.6%), Llama produces more literary content, DeepSeek generates religious material at higher rates, and Qwen frequently outputs multiple-choice exam questions. The study reveals that model behaviors including depth of technical content and degenerate text patterns are systematic and model-specific rather than random, with implications for auditing, safety, and understanding unconstrained model behavior.