💬 NLP

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

This paper introduces "Creative Collision," a novel activation steering framework that superimposes opposing directorial persona vectors (Spielberg vs. Scorsese) to reveal that one persona often dominates the other, intermediate mixing paradoxically enhances coherence, and both moral tones localize to the same deep transformer layer, offering new insights into the geometry of competing semantic directions for controllable creative generation.

Original authors: Subramanyam Sahoo, Justin Shenk

Published 2026-06-16

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Subramanyam Sahoo, Justin Shenk

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a very smart, creative robot that writes stories. Usually, if you want the robot to write like a specific person, you might just tell it, "Write like Steven Spielberg" or "Write like Martin Scorsese."

But this paper asks a different, more complex question: What happens if we try to make the robot write like both directors at the exact same time?

The researchers call this "Creative Collision." They wanted to see what happens when two very different "personalities" crash into each other inside the robot's brain.

The Two Directors (The Opposing Forces)

To test this, they picked two famous movie directors who represent opposite ends of the moral spectrum:

Steven Spielberg: Think E.T. or Schindler's List. His stories are about hope, childhood wonder, redemption, and happy endings.
Martin Scorsese: Think Goodfellas or Taxi Driver. His stories are about moral gray areas, betrayal, violence, and characters making bad choices.

The Experiment: Mixing the "Flavors"

The researchers didn't just ask the robot to switch back and forth. Instead, they created a "mixing knob" (called $\alpha$ ).

Turn the knob to 0: The robot writes purely like Spielberg.
Turn the knob to 1: The robot writes purely like Scorsese.
Turn the knob to 0.5: The robot tries to be 50% Spielberg and 50% Scorsese simultaneously.

They turned this knob in small steps and watched what happened to the stories the robot generated.

The Big Surprises

1. Spielberg is the "Heavyweight Champion"
The most surprising finding was that Spielberg's style is incredibly dominant. Even when the researchers tried to mix in a lot of Scorsese (up to 75% Scorsese!), the stories still sounded mostly like Spielberg.

The Analogy: Imagine trying to mix a drop of dark ink into a giant bucket of bright orange paint. No matter how much dark ink you add, the bucket stays orange. The robot's internal "optimism" is so strong that it swallows the "darkness" unless you turn the knob all the way to 100% Scorsese.

2. The "Coherence Valley" (The Sweet Spot)
Usually, when you force a robot to do something it doesn't naturally want to do, it starts making mistakes and writing nonsense. You'd expect that mixing two opposite styles would make the robot confused and the story messy.

The Surprise: When they mixed the two directors in the middle (around 25% to 50% Scorsese), the stories actually became clearer and more coherent than when they tried to force the robot to be purely Scorsese.
The Analogy: Think of it like pushing a heavy box. If you push it straight forward with all your might, it might tip over. But if you push it from two slightly different angles at the same time, the forces cancel each other out just enough to keep the box stable. The "collision" of the two styles actually helped the robot stay on track.

3. The "Moral Amplifier"
When they mixed the styles slightly (about 25% Scorsese), the stories didn't just become a boring mix. They became morally complex.

The Analogy: Pure Spielberg is like a sunny day; pure Scorsese is like a storm. But the "collision" zone was like a dramatic sunset where you see both the beauty of the light and the darkness of the clouds. The robot generated stories that had both hope and tragedy happening at the same time, creating a richer, more interesting moral tone than either director alone.

4. Where in the Brain Does This Happen?
The researchers looked inside the robot's "brain" (which has 40 layers of processing) to see where these personality changes happened.

The Finding: They found a specific "control room" at Layer 28 (about 70% of the way through the brain). This is where the robot decides the "moral tone" of the story. Both the happy Spielberg vibes and the dark Scorsese vibes live in this exact same spot, just pointing in opposite directions.

Why Does This Matter?

The paper suggests that the robot's training (what it learned from the internet) made it naturally lean toward "good" and "hopeful" stories. It takes a huge amount of force to make it write dark, cynical stories.

However, the "Creative Collision" method shows that if you want to create something complex and interesting, you don't always need to force the robot to be 100% one thing. Sometimes, gently mixing two opposing ideas creates a "sweet spot" where the story is stable, coherent, and emotionally deep.

In short: The robot has a strong "good guy" bias that is hard to break. But if you gently push it toward "bad guy" territory without fully letting go of "good guy," you get the most interesting, coherent, and morally complex stories of all.

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

Problem Statement

Activation steering has established itself as a potent mechanism for modifying Large Language Model (LLM) behavior at inference time by injecting learned semantic directions into the residual stream. While prior research has successfully applied this technique to single-direction tasks—such as enhancing truthfulness, safety, or specific personas—the dynamics of multi-vector collision remain unexplored. Specifically, it is unknown what occurs when two semantically opposing steering vectors are superimposed and compete for representational control. This paper addresses this gap by investigating the "Creative Collision" regime, where opposing moral and stylistic directions are interpolated to observe non-linear interactions, coherence degradation, and directional dominance.

Methodology

1. Directorial Contrast Corpus

The authors constructed a paired corpus $D = \{(x^{SPL}_i, x^{SCO}_i)\}$ consisting of 100 excerpts each from Steven Spielberg and Martin Scorsese filmographies.

Spielberg: Drawn from films like E.T., Schindler's List, and Saving Private Ryan, characterized by redemptive arcs, emotional catharsis, and optimistic resolution.
Scorsese: Drawn from films like Goodfellas, Taxi Driver, and Casino, characterized by moral ambiguity, violence, and self-destructive trajectories.
Matching: Pairs were matched on narrative situations (e.g., confrontation, moral choice) to isolate directorial moral tone from plot complexity.

2. Steering Vector Extraction

Using a 14-billion-parameter decoder-only transformer (40 layers), the authors extracted steering vectors via mean-difference contrast against a shared neutral baseline corpus $B$ .

Vectors: $v^{(l)}_{SPL}$ and $v^{(l)}_{SCO}$ were computed as the difference between the mean activation of the respective director's corpus and the baseline at layer $l$ .
Normalization: Vectors were $\ell_2$ -normalized. Notably, they were not anti-parallel; at layer 28, their cosine similarity was $\approx 0.29$ , indicating a shared representational subspace for cinematic emotional content despite divergent moral resolutions.

3. Interpolation and Steering Protocol

The study introduced a parametric family of interpolated vectors:
$v^{(l)}_\alpha = (1 - \alpha)\hat{v}^{(l)}_{SPL} + \alpha\hat{v}^{(l)}_{SCO}, \quad \alpha \in [0, 1]$

Steering: The interpolated vector was added to the residual stream at layers $L = \{20, \dots, 38\}$ with a steering coefficient $\lambda \in \{0.5, 1.0, 1.5, 2.0\}$ .
Evaluation: 50 creative fiction passages were generated per condition. Metrics included:
- Moral Valence: Scored by a fine-tuned classifier ( $[-1, +1]$ ).
- Coherence: Measured via token-level perplexity under the unsteered base model.
- Directional Dominance: Probability assigned by a stylometric classifier to the Spielberg class.
- Surface Style: Metrics including sentence length, dialogue density, and lexical diversity.

Key Results

1. Non-Monotonic Moral Valence

Contrary to the naive expectation that moral tone would shift linearly from Spielberg (positive) to Scorsese (negative) as $\alpha$ increases, the results were non-monotone.

At $\lambda = 1.0$ , moral valence was most negative at $\alpha = 0.25$ ( $MV \approx -0.38$ ) rather than at $\alpha = 1.0$ .
This "moral incoherence" at intermediate points suggests constructive interference where the Scorsese component introduces darkness but fails to dominate, creating a penalized state.
At high steering magnitudes ( $\lambda = 1.5$ ), the moral signal collapsed entirely to near-zero, indicating the residual stream was pushed off the natural text manifold.

2. The Coherence Valley

A counterintuitive finding emerged regarding generation coherence (perplexity).

Pure Steering: Single-director steering ( $\alpha=0$ or $1$) caused the largest perplexity increases at high $\lambda$ .
Intermediate Collision: Conditions with mixed vectors ( $\alpha \in \{0.25, 0.5\}$ ) remained significantly more coherent.
Geometric Explanation: The authors proved that interpolating two non-anti-parallel unit vectors ( $\rho < 1$ ) reduces the $\ell_2$ norm of the resulting vector. The norm is minimized at $\alpha = 0.5$ . Since coherence cost scales with the perturbation magnitude, intermediate collision points apply a weaker effective perturbation, preserving the model's proximity to the natural text manifold.

3. Directional Dominance

Spielberg's representational signature exhibited robust dominance across almost the entire interpolation range.

For $\alpha \in \{0.0, 0.25, 0.50, 0.75\}$ , the stylometric classifier assigned a probability of $\approx 1.0$ to the Spielberg class.
Only at $\alpha = 1.0$ (pure Scorsese) did the classifier approach the decision boundary ( $P \approx 0.49$ ).
Mechanisms: This is attributed to (1) a prior bias in pretraining data toward prosocial narratives and (2) alignment amplification via RLHF, which reinforces positive generation, effectively acting as a persistent Spielberg-direction prior.

4. Layer-wise Localisation

Both directors' moral effects peaked at Layer 28 (approx. 70% depth in the 40-layer model).

The effects were roughly anti-symmetric ( $\Delta MV \approx -0.50$ for Scorsese, $+0.47$ for Spielberg).
This co-localisation suggests Layer 28 functions as a shared "moral-tone hub" where high-level semantic properties are causally encoded, supporting the linear representation hypothesis for moral valence.

5. Moral Amplification

Decomposing moral valence into positive ( $p+$ ) and negative ( $p-$ ) subscores revealed that the collision regime ( $\alpha \approx 0.25$ ) produced text with both high positive and high negative moral scores simultaneously. This "moral amplification" indicates that collision can generate text with greater moral complexity and explicitness than either pure-director extreme.

Significance and Claims

The paper claims to advance the understanding of competing semantic directions in transformer activation spaces. Its primary contributions are:

Safety Implications: The robust directional dominance of prosocial content suggests that models fine-tuned for helpfulness maintain a strong internal prior against morally dark content. Steering toward undesirable content requires substantially larger activation magnitudes than steering toward benign content, raising the empirical bar for adversarial activation-level attacks.
Design Principle for Coherence: The "coherence valley" provides a geometric principle: when high steering magnitude is required, interpolating between two semantically opposed directions can preserve coherence better than applying a single direction at equivalent total magnitude, provided the vectors are non-antipodal.
Creative Generation: The "moral amplification" effect suggests that deliberately injecting a small opposing-direction component can increase narrative tension and moral depth without abandoning the dominant persona, offering a new tool for controllable creative generation.
Mechanistic Interpretability: The identification of Layer 28 as a shared moral-tone substrate and the confirmation of linear moral-valence directions provide concrete targets for future causal tracing and feature decomposition studies.

The authors emphasize that while the capability to steer toward dark content existed previously, their work quantifies the resistance imposed by alignment fine-tuning. They view the finding that such steering is difficult to achieve as a net positive for safety literature, while acknowledging the potential for "moral amplification" to be exploited for high-intensity content generation.