Creative Collision: Directorial Persona Steering and Competition in Large Language Models
Researchers have developed a method called "Creative Collision" to steer large language models using two opposing directorial personas, Steven Spielberg and Martin Scorsese. By interpolating between vectors representing their distinct styles, the study found that Spielberg's persona exhibits directional dominance, largely suppressing Scorsese's influence. Interestingly, intermediate collision points paradoxically improved generation coherence, and both personas were found to localize maximally to a specific layer in the transformer model, suggesting a shared "moral-tone substrate." AI
IMPACT This research offers new insights into controlling LLM behavior and could lead to more nuanced creative generation.