Creative Collision: Directorial Persona Steering and Competition in Large Language Models

A study posted to arXiv examines what happens when two opposing directorial personas are injected simultaneously into a large language model, finding that one vector can dominate the other across nearly the full range of mixing parameters. The paper, titled “Creative Collision: Directorial Persona Steering and Competition in Large Language Models,” was submitted by Subramanyam Sahoo on 15 June 2026 [1]. The researchers constructed persona vectors for filmmakers Steven Spielberg, characterized by an optimistic and redemptive moral valence, and Martin Scorsese, associated with a dark and morally ambiguous tone, using mean-difference activation contrast on curated screenplay-derived corpora [2]. They then interpolated between the two vectors with a scalar mixing parameter α ∈ [0,1] and a steering coefficient λ [2]. Across five evaluation axes, the study reports that Spielberg’s representational signature exhibits robust directional dominance, suppressing Scorsese’s moral influence across almost the entire interpolation range [2]. A second finding shows that intermediate collision points can paradoxically improve generation coherence relative to pure single-director steering at high steering coefficients [2]. The work also identifies that both personas localise maximally to layer 28 of a 40-layer decoder-only transformer, which the authors describe as a shared moral-tone substrate [2]. The concept of activation steering builds on a broader body of research into controlling model outputs at inference time. Anthropic, co-founded by theoretical physicist Jared Kaplan, has explored related techniques for aligning model behavior with human intent [8]. The preprint server arXiv, where the paper appears, was itself shaped in part by astrophysicist Joanne Cohn, who played a role in its creation as an open-access repository for early research [7]. The submission file for the study is 1,039 KB [1]. The authors argue the results have direct implications for controllable creative generation and value-aligned narrative synthesis, illuminating the geometry of competing semantic directions in transformer residual streams [2].