From random passphrases to thematic constraints: a data-driven analysis of techniques to increase LLM output diversity without sacrificing coherence.

An investigation into absorption points in character generation, revealing how training data biases create repetitive outputs and what we can do about it.