Experiment suggests modern LLMs like Qwen hide tiny GPT2 self-models inside

paraschopra · x · 2026-09-22

Paras Chopra ran an exploratory study testing whether modern LLMs contain self-models of LLMs. Using 12 post-cutoff Sept 2026 headlines, he had GPT2-medium generate continuations, then compared GPT2's own completions against Qwen3-base (4B) finishing the same text, with Qwen's natural completions as control.

Key finding: Qwen's completions of GPT2-started text matched GPT2's style more than its own natural style, suggesting Qwen can simulate GPT2's distribution internally — consistent with the hypothesis that LLMs pretrain on so much AI-generated text that they implicitly model LLMs themselves. He plans a follow-up asking models to judge which model produced a completion. Caveat: quick-and-dirty exploratory work.

Original post →

More from Models

Models channel →