Modern LLMs have tiny GPTs hidden inside them
Short summary
A researcher tested whether Qwen (a large language model) has internalized representations of GPT-2's text generation patterns by having Qwen complete text that GPT-2 started. Results suggest Qwen can mimic GPT-2's style and infer that GPT-2-generated text is from an earlier era, implying modern LLMs may develop internal models of other models during training.
- When Qwen completed GPT-2-generated text prefixes, its continuations overlapped more with GPT-2's actual completions than with Qwen's own natural completions of the same prompts, across multiple overlap metrics.
- Qwen inferred significantly earlier years (matching GPT-2's pre-2019 training data) when asked to complete years in GPT-2-generated articles versus its own generations, even after removing articles with explicit year references.
- The experiment used 12 internet headlines from September 2026 (outside both models' training cutoffs) with varying completion lengths (96-512 tokens), and results held up when tested on Qwen's 14B parameter version.
Full summary
The author tested whether large language models contain internal models of other LLMs by having Qwen (a 4 billion parameter model) complete text that GPT-2 had started. The setup was straightforward: feed Qwen a headline plus some GPT-2-generated text, then measure whether Qwen's continuation looked more like how GPT-2 would finish its own text or how Qwen naturally completes things. Using overlap metrics across different token lengths (96, 256, 512 initial tokens), Qwen's completions consistently resembled GPT-2's hidden continuations more than Qwen's own natural responses. This held up in a secondary test too: when the author appended "Year: 20" to full articles and asked Qwen to complete the year, Qwen inferred earlier dates for GPT-2 text than for its own, suggesting it had learned that GPT-2's style belongs to an earlier era (GPT-2 was trained on pre-2019 data).
The concrete example drives home the pattern. Given a headline about air quality and toddler development plus GPT-2's prefix about recruiting ADHD students, Qwen wrote: "The children with ADHD performed significantly better than the control group." That mirrors GPT-2's actual hidden continuation about ADHD kids having better IQ scores. When given only the headline, Qwen wrote something entirely different: "Children in cities with cleaner air performed better on tests." The author controlled for confounds—removing articles with explicit year references—and the effect persisted even at the larger 14 billion parameter scale.
The author frames this as exploratory rather than definitive, but the implications are worth taking seriously. If LLMs genuinely develop internal approximations of how other models behave, they could use those proxies to simulate their own outputs before committing to them, estimate uncertainty more accurately through internal sampling, or develop something resembling self-awareness. Larger models would likely build sharper representations of predecessor models, making this effect stronger at scale. The limitation: instruction-tuned models don't work well for this test since they're optimized for following orders, not raw text completion.