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

2026-06-15 · Source: Computation and Language · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Natural Language Processing · Depth: Expert, quick

Summary

A study introduces "Creative Collision," a novel approach to large language model behavior shaping that superimposes two semantically opposing activation steering vectors. Researchers constructed directorial persona vectors for Steven Spielberg (optimistic, redemptive) and Martin Scorsese (dark, morally ambiguous) using mean-difference activation contrast on curated screenplay corpora. By interpolating these vectors with a scalar mixing parameter α∈ [0,1] and a steering coefficient λ, the study evaluated five axes: moral valence, generation coherence, surface style, directional dominance, and vector geometry. Key findings reveal Spielberg's persona exhibits robust directional dominance, suppressing Scorsese's moral influence across most of the interpolation range. Paradoxically, intermediate collision points improved generation coherence compared to pure single-director steering at high λ. Both personas localized maximally to layer 28 of a 40-layer decoder-only transformer, suggesting a shared moral-tone substrate. These results offer insights into competing semantic directions within transformer residual streams.

Key takeaway

For NLP Engineers developing controllable creative generation systems, understanding "Creative Collision" is crucial. You should consider superimposing opposing persona steering vectors to achieve nuanced moral valences and potentially improve generation coherence. This method offers a pathway to fine-tune narrative synthesis, allowing you to balance conflicting stylistic influences. Explore intermediate collision points, as they can yield more coherent outputs than single-direction steering, enhancing your model's creative control.

Key insights

The superposition of opposing steering vectors reveals complex interactions and shared representational substrates in LLMs.

Principles

• Opposing steering vectors can exhibit directional dominance.
• Intermediate semantic collisions may enhance coherence.
• Persona representations can localize to specific transformer layers.

Method

Directorial persona vectors are constructed via mean-difference activation contrast on curated corpora, then interpolated with scalar mixing and steering coefficients.

In practice

• Steer LLM output towards specific moral valences.
• Improve generation coherence using mixed steering.
• Identify shared representational substrates in transformers.

Topics

• Activation Steering
• Large Language Models
• Persona Steering
• Creative Generation
• Transformer Architecture
• Semantic Control

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• Steering at the Source: Style Modulation Heads for Robust Persona Control — Targeting specific "Style Modulation Heads" in LLMs enables robust persona control with minimal coherency degradation.
• The Pragmatic Persona: Discovering LLM Persona through Bridging Inference — LLM personas are deeply encoded in discourse coherence via implicit conceptual links, not just surface-level lexical patterns.
• Modeling Pathology-Like Behavioral Patterns in Language Models Through Behavioral Fine-Tuning — Consistent behavioral fine-tuning can induce stable, pathology-like biases in LLM generative distributions and action selection.
• Multi-Trait Subspace Steering to Reveal the Dark Side of Human-AI Interaction — MultiTraitsss systematically creates "Dark models" to simulate and mitigate harmful human-AI interactions by steering LLMs towards crisis-associated traits.
• Staying with the Uncertainty: Uncertainty-Scaffolding Strategies for Artificial Moral Advisors in LLM-to-LLM Simulated Conversations — Uncertainty-scaffolding strategies enhance engagement quality in LLM-based Artificial Moral Advisors during ethical discussions.
• Playing Devil's Advocate: Off-the-Shelf Persona Vectors Rival Targeted Steering for Sycophancy — Off-the-shelf persona vectors effectively mitigate model sycophancy, rivaling targeted steering methods while preserving accuracy.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computation and Language.