What do LLMs think when you don't tell them what to think about?
This title could be clearer and more informative.Try out Clickbait Shieldfor free (5 uses left this month).
Researchers from Together.ai studied what LLMs generate when given minimal, topic-neutral prompts with no chat templates or system instructions. The findings reveal that different model families have distinct 'knowledge priors': GPT-OSS defaults heavily to programming (27.1%) and mathematics (24.6%), Llama favors literary and narrative content, DeepSeek generates religious content at elevated rates, and Qwen frequently outputs multiple-choice exam questions. These topical biases are consistent across different prompts, embeddings, and labelers, suggesting they reflect population-level fingerprints rather than noise. Additionally, depth differs by family — GPT-OSS produces expert-level content 68.2% of the time, while Llama and Qwen skew toward basic material. Degenerate outputs (repetitive or meaningless text) also carry model-specific signatures, including Llama emitting real personal social media URLs, raising safety and privacy concerns. The research argues that near-unconstrained generation exposes model defaults that standard benchmarks systematically miss.
