Fictional concepts survive and spread best when they violate only a couple of intuitive physical expectations. Many people assume that wildly bizarre creations or completely realistic figures attract the largest audiences over time. Feeding human training examples into language model prompts allowed software to measure these rule breaks across massive collections of fictional monsters.
Standard language models initially assigned too many rule violations to ordinary concept descriptions during zero-shot testing. Much like calibrating a scale with known test weights, researchers inserted previously verified human examples directly into each prompt. This calibration preserved sensitivity while nearly doubling the ability of the models to reject false violations. The calibrated software then processed hundreds of second-generation and later Pokémon entries to catalog their counterintuitive traits.
Researchers evaluated the Claude and Gemini models against human baselines before applying the calibrated pipeline to an expanded monster dataset. The team merged the automated trait scores with an independent public popularity ranking. A fitted parabola revealed a statistically significant curve with its peak sitting at 2.04 rule violations, matching theoretical predictions.
The authors state that a modest human-coded seed set can calibrate language models to test classic cultural transmission theories across large populations. This approach allows researchers to analyze massive cultural archives without facing the high financial cost of manual content coding.