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

Trait induction and control2026arXivApproved editorial review

Authors: Subramanyam Sahoo, Justin Shenk

Keywords: Activation steering, Directorial style, Qwen3-14B, Moral valence measurement, Vector interpolation, Construct validity, Corpus confounding, Reproducibility

Source: Open primary source (opens in a new tab)

2
Authors
10
Findings
22
Limitations
4
Evidence

Editorial summary

English

This paper, accepted at the ICML 2026 Workshop on Human-AI Co-Creativity, asks whether two corpus-derived activation directions can be mixed to control an LLM's narrative tone. It casts Steven Spielberg as an optimistic, redemptive pole and Martin Scorsese as a dark, morally ambiguous pole. The declared protocol uses an open 14-billion-parameter, 40-layer transformer; the repository identifies Qwen3-14B. For each director it computes mean corpus activations, subtracts a neutral mean, normalizes both directions, and interpolates them with alpha from zero to one. The mixture is added to layers 20-38 at four lambda coefficients. According to the PDF, each of 20 conditions produces 50 moral-choice scenes capped at 200 tokens and is evaluated with an ETHICS-trained moral classifier, perplexity, spaCy metrics, a stylometric classifier, and vector geometry. The conceptual proposal, studying interactions among multiple activation directions rather than one direction, is relevant to style control and causal analysis of representations. However, the public implementation does not match this protocol on central design choices, so its results must be read as a different exploratory experiment. The released corpus contains 80 entries per class, not 100. The Scorsese set consists of 40 split sections from Taxi Driver and 40 from Raging Bull; all 80 Spielberg entries come from Schindler's List; all 80 neutral entries come from Lost in Translation. The notebook downloads IMSDb screenplays, splits them sequentially on INT./EXT. markers, and truncates at 80. There is no narrative-situation pairing, five-film-per-director sampling, diverse neutral sample, or per-entry provenance as claimed. Director, screenwriter, film, characters, formatting, era, topic, violence, and length are therefore confounded with the label. The actual experiment does not generate 1,000 scenes either. It uses five fixed prompts, five outputs per condition, and 100 outputs total, with max_new_tokens=512. This explains published values of 225-500 words and as many as 500 words per sentence, which are incompatible with the declared 200-token cap. The largest discrepancy concerns moral valence. The repository includes no ETHICS classifier or fine-tuning code; it deterministically counts 15 positive and 15 negative substrings. A zero means none of those words occurred, not that a learned model found the text neutral or that steering left the natural activation manifold. Across 100 records the score ranges from -1 to only +.2. At lambda=1, alpha means are -.20, -.40, -.20, 0, and 0; even the pure Spielberg endpoint is not positive. The repository's own analysis reports F=.1628 and p=.956692 for the effect of alpha on valence, a null result that does not support moral interpolation or amplification. Collapse to zero at lambda=1.5 reflects absence of the 30 marker strings, not demonstrated classifier saturation. The published geometry also does not describe the executed code. The paper says both directions are normalized and derives a 20% norm reduction for the middle mixture. The notebook never normalizes them: it reports average norms of 94.4638 for Scorsese and 72.7357 for Spielberg and mixes the raw vectors. For two unit directions, alpha=.5 must have equal cosine to both endpoints and their equality crossing must occur at .5. The CSV instead gives .8868 similarity to Scorsese and .6814 to Spielberg at alpha=.5, while the PDF places the crossing near .4, the signature of unequal raw norms. The proposition may be algebraically correct for unit vectors, but it does not explain this experiment. Even with normalization, comparing lower-norm mixtures at the same nominal lambda confounds a collision advantage with a weaker effective intervention; norm-matched mixtures, random directions, and a lambda=0 control are missing. Layer-28 localization is also much weaker than described. The code uses lambda=.5 rather than 1 and plots raw lexicon valence rather than a shift from an unsteered control. It selects the largest absolute mean among ten layers with only five outputs per director-layer. At L28, Scorsese has mean -.52 with SD .8672 and Spielberg +.441 with SD .4364. Without a baseline, intervals, selection correction, confirmation set, or replication, this does not establish a causal moral-tone hub. Directional dominance comes from logistic regression over Qwen layer-30 activations, not surface stylometry. It cross-validates on fragments and then fits all 160 screenplay sections, rather than training on 70% and retaining 30%. Because sections from the same three screenplays enter different folds, the .975 accuracy can reflect characters, films, and formatting. The pure Scorsese endpoint is also unstable: P(Spielberg) is .5673, .5909, .1757, and .7628 as lambda changes. Classifying pure Scorsese output mostly as Spielberg in three of four configurations indicates a validity problem, not proof that RLHF imposes a Spielberg prior. No aligned-versus-unaligned model comparison is performed. The artifact has meaningful openness: it releases an executed notebook, corpora, activations, a 100-row metric file, layer and dominance results, figures, and an Apache-2.0 license. This enables auditing and descriptive recomputation. But there is no README, locked environment, model revision, seed, tests, CI, data card, or manifest; seven of 20 raw-generation files are missing and the last cell ends in NameError. Corpus and classifier mean pooling includes padded positions because attention_mask is ignored. The license also does not establish rights to redistribute third-party screenplays. No human evaluation validates director resemblance, morality, coherence, or narrative richness. The defensible contribution is an exploratory notebook and a useful hypothesis about mixing corpus-derived directions. It is not validated evidence of opposing moral personas, a layer-28 moral substrate, alignment-caused prosocial dominance, or coherence improvement caused by semantic collision.

Español

Este trabajo, aceptado en el workshop de Human-AI Co-Creativity de ICML 2026, estudia si dos direcciones de activación derivadas de corpus cinematográficos pueden mezclarse para controlar el tono narrativo de un LLM. El artículo presenta a Steven Spielberg como polo optimista y redentor y a Martin Scorsese como polo oscuro y moralmente ambiguo. Su protocolo declarado usa un transformer abierto de 14.000 millones de parámetros y 40 capas; el repositorio identifica Qwen3-14B. Para cada director calcula la activación media de un corpus, resta una media neutral, normaliza las dos direcciones y las interpola con alpha entre 0 y 1. La mezcla se suma a las capas 20-38 con cuatro coeficientes lambda. Según el PDF, cada una de las 20 condiciones genera 50 escenas de hasta 200 tokens a partir de un prompt de elección moral, y se evalúa con un clasificador moral entrenado en ETHICS, perplejidad, métricas spaCy, un clasificador estilométrico y geometría vectorial. La propuesta conceptual, estudiar la interacción de varias direcciones en vez de una sola, es relevante para control de estilo y análisis causal de representaciones. Sin embargo, la implementación pública no corresponde a ese protocolo en decisiones centrales, por lo que los resultados deben leerse como un experimento exploratorio distinto. El corpus publicado contiene 80 entradas por clase, no 100. Las 80 entradas Scorsese proceden de 40 secciones de Taxi Driver y 40 de Raging Bull; las 80 Spielberg son de Schindler's List; las 80 neutrales, de Lost in Translation. El notebook descarga guiones de IMSDb, los divide secuencialmente por INT./EXT. y corta en 80. No hay emparejamiento por situación narrativa, cinco películas por director, muestra neutral diversa ni trazabilidad entrada a entrada como afirma el artículo. Por ello, director, guionista, película, personajes, formato, época, tema, violencia y longitud quedan confundidos con la etiqueta. El experimento real tampoco genera 1.000 escenas. Usa cinco prompts fijos, cinco salidas por condición y 100 salidas totales, con max_new_tokens=512. Esto explica que las métricas publicadas alcancen 225-500 palabras y hasta 500 palabras por frase, valores incompatibles con el límite declarado de 200 tokens. La discrepancia más importante afecta a la valencia moral. El repositorio no incluye un clasificador ETHICS ni código de fine-tuning: cuenta determinísticamente 15 subcadenas positivas y 15 negativas. Un cero significa que no apareció ninguna palabra de esa lista, no que un clasificador aprendió neutralidad ni que la activación salió del manifold. En los 100 registros la puntuación va de -1 a solo +0,2. Con lambda=1, las medias para alpha 0, 0,25, 0,5, 0,75 y 1 son -0,20, -0,40, -0,20, 0 y 0; incluso el extremo Spielberg no es positivo. El propio análisis guardado obtiene F=0,1628 y p=0,956692 para el efecto de alpha sobre la valencia, un resultado nulo que no respalda la interpolación moral ni la amplificación descritas. El colapso a cero con lambda=1,5 refleja ausencia de las 30 palabras, no saturación demostrada de un clasificador. La geometría publicada tampoco describe el código ejecutado. El artículo dice que normaliza ambas direcciones y deriva una reducción de norma del 20% para la mezcla intermedia. El notebook nunca normaliza: informa normas medias 94,4638 para Scorsese y 72,7357 para Spielberg y mezcla esos vectores crudos. Para dos direcciones unitarias, alpha=0,5 debe tener igual coseno con ambos extremos y el cruce debe ocurrir en 0,5. El CSV da en alpha=0,5 cosenos 0,8868 con Scorsese y 0,6814 con Spielberg, y el PDF sitúa el cruce cerca de 0,4: es la firma de las normas desiguales. Por tanto, la proposición puede ser algebraicamente correcta para vectores unitarios, pero no explica el experimento. Incluso con normalización, comparar mezclas de menor norma al mismo lambda confunde la supuesta ventaja de colisión con una intervención efectiva más débil; faltan comparaciones con norma igualada, direcciones aleatorias y lambda=0. La localización en capa 28 también es más frágil de lo descrito. El código usa lambda=0,5, no 1, y representa valencia léxica bruta, no un desplazamiento respecto a un control sin steering. Selecciona el máximo absoluto entre diez capas con solo cinco salidas por capa y director. En L28 la media Scorsese es -0,52 con desviación 0,8672 y la Spielberg +0,441 con desviación 0,4364. Sin baseline, intervalos, corrección por selección, conjunto de confirmación o réplica, eso no demuestra un hub causal de tono moral. La dominancia direccional procede de una regresión logística sobre activaciones Qwen de capa 30, no de estilometría superficial. Se valida por folds de fragmentos y luego se entrena con las 160 secciones, no con el 70% y un 30% retenido. Como fragmentos de los mismos tres guiones aparecen en los folds, la precisión 0,975 puede capturar personajes, película y formato. Además, el extremo Scorsese es inestable: P(Spielberg) vale 0,5673, 0,5909, 0,1757 y 0,7628 al variar lambda. Ser clasificado mayoritariamente como Spielberg en tres de cuatro configuraciones puras muestra un problema de validez del clasificador, no una prueba de que RLHF imponga un prior Spielberg. No se compara ningún modelo alineado con uno no alineado. Hay aspectos positivos de apertura: el repositorio ofrece notebook ejecutado, corpus, activaciones, 100 filas de métricas, resultados de capas y dominancia, figuras y licencia Apache-2.0. Eso permite auditar las afirmaciones y recomputar descriptivos. Pero no hay README, entorno bloqueado, revisión de modelo, semilla, tests, CI, data card o manifiesto; faltan siete de los veinte archivos de generaciones crudas y la última celda termina con NameError. El mean pooling de corpus y clasificador incluye posiciones de padding porque ignora attention_mask. La licencia tampoco acredita derechos para redistribuir guiones de terceros. No existe evaluación humana de semejanza directorial, moralidad, coherencia o riqueza narrativa. La contribución defendible es un notebook exploratorio y una hipótesis útil sobre mezcla de direcciones derivadas de corpus. No es evidencia validada de personalidades morales opuestas, de un sustrato moral en L28, de dominancia prosocial causada por alineamiento ni de una mejora de coherencia causada por la colisión semántica.

Research question

How do two steering directions derived from director-labeled corpora interact when interpolated and injected into Qwen3-14B, and interpretablely change the moral tone, fluency, and style of generation?

Method

The PDF declares normalized director-minus-neutral vectors, 100 passages per director, 100 neutral, 50 generations per combination of five alpha and four lambda, one moral prompt, 200 token limit, ETHICS classifier, spaCy metrics, and a classifier with 30% held out. The executed artifact uses Qwen3-14B quantized to 4 bits, three IMSDb corpora of 80 sections, raw unnormalized vectors, five prompts, five generations per condition, 512 tokens, a deterministic list of moral words, metrics by split/regex, and logistic regression over Qwen activations.

Sample: Artifact: 80 Scorsese sections, 80 Spielberg sections, and 80 neutral; 20 conditions with five prompts and one output per prompt, 100 total generations. Localization: ten even layers by two directions and five prompts, 100 outputs. The PDF declares 100 passages per partition and 1,000 generations, figures not implemented in the repository.

Findings

  • The artifact executes a complete grid of 20 conditions and publishes 100 rows of auditable metrics.
  • Perplexity increases with lambda and is lower in some mixtures than at the extremes, but without matching norm or including lambda=0 the collision effect is not identified.
  • The ANOVA saved for valence by alpha is null: F=0.1628, p=0.956692.
  • The Spielberg extreme does not produce positive valence with lambda=1; its mean is -0.20.
  • The supposed moral collapse at lambda=1.5 is absence of words from a fixed lexicon, not saturation of an ETHICS classifier.
  • The vectors are mixed without normalization and have mean norms 94.4638 and 72.7357.
  • The geometric asymmetry of alpha=0.5 contradicts the two unit vector model used in the proposition.
  • L28 maximizes absolute score in a small exploratory sweep, with large deviations and no non-directed control.
  • The dominance classifier reaches 0.975 by cross-validation of fragments, but confuses movie and format and fails non-monotonically at the Scorsese extreme.
  • The repository facilitates audit, but does not reproduce the central protocol described in the PDF.

Limitations

  • The public corpus does not match in size, movies, neutrality, pairing, or provenance with the described corpus.
  • Director, screenwriter, movie, characters, format, and theme are confounded with the label.
  • There are only five observations per condition and the same five prompts are repeated across the grid.
  • No seed is fixed and stochastic sampling is not repeated.
  • There is no lambda=0 condition or baseline generation without steering.
  • The real limit is 512 tokens, not 200.
  • The moral metric is a deterministic unvalidated lexicon, not an ETHICS classifier.
  • The effect of alpha on that metric is not significant in the analysis itself.
  • Style metrics do not use spaCy and have fragile heuristics.
  • Perplexity does not equal narrative coherence.
  • The vectors are not normalized as required by the published theory.
  • The effective magnitude of interventions is not matched when comparing mixtures and extremes.
  • The mean activation includes padding by ignoring attention_mask.
  • The localization uses lambda=0.5 and raw valence, not lambda=1 and delta against control.
  • L28 is selected as the maximum among ten layers without correction or confirmation.
  • The classifier validation separates fragments from the same scripts, not independent movies.
  • Pure Scorsese classification is unstable and non-monotonic with lambda.
  • There is no human evaluation of style, morality, or coherence.
  • An aligned model is not compared with an unaligned one to attribute results to RLHF.
  • Missing README, locked environment, model review, tests, CI, data card, and seven raw output files.
  • The notebook ends with NameError in the final optional cell.
  • The provenance and rights of the redistributed scripts are not documented on input.

What the study does not establish

  • It does not demonstrate that Spielberg and Scorsese are opposite poles of a latent moral axis.
  • It does not validate a stable psychological or moral personality of the model.
  • It does not demonstrate that semantic mixing causes an improvement in coherence.
  • It does not demonstrate a moral tone substrate or hub in layer 28.
  • It does not demonstrate robust Spielberg dominance over the Scorsese extreme.
  • It does not demonstrate that RLHF or alignment cause a prosocial prior.
  • It does not demonstrate moral amplification or complexity using a validated classifier.
  • It does not generalize to other models, corpora, directors, languages, or domains.
  • It does not publicly implement the protocol of 1,000 generations, unit vectors, and ETHICS classifier described.
  • It does not resolve rights, consent, or governance of imitation of styles of living persons.

Traceability

Scope: Full text

Version: arXiv:2606.16240v1

Consulted source: https://arxiv.org/abs/2606.16240v1

Review: Codex ten-page full-text visual, complete TeX, public notebook, corpus, metric, geometry, construct-validity and reproducibility audit, 2026-07-16

Approval: Codex fidelity pass, 2026-07-16

English translation: approved, 2026-07-18

Models evaluated

  • Qwen/Qwen3-14B in 4-bit NF4
  • Layer-30 Qwen activation logistic-regression classifier
  • Unsteered Qwen3-14B perplexity evaluator

Instruments and metrics

  • Mean-difference residual-stream activation vectors
  • Five-by-four alpha and steering-coefficient grid
  • Thirty-substring moral lexicon in the released notebook
  • Base-model token perplexity
  • Whitespace and regex surface-style heuristics
  • Layer-wise single-layer intervention sweep
  • Cosine similarity of collision vectors
  • Five-fold cross-validated logistic regression on Qwen activations

Data used

  • Released Taxi Driver screenplay sections
  • Released Raging Bull screenplay sections
  • Released Schindler's List screenplay sections
  • Released Lost in Translation screenplay sections
  • ETHICS only as an unsupported paper claim; no ETHICS artifact released

Evidence and location

  • Metadata, workshop acceptance, and version: Official arXiv record 2606.16240v1, checked 2026-07-16
  • Declared method and results: arXiv v1, all ten PDF pages, complete TeX and appendix
  • Code, corpus, real protocol, metrics, and notebook errors: Public Creative-Collision repository at commit 23b49c35a3f46a1eae4064deec32c7a8eeeaf2a1
  • Discrepancies, recalculations, validity, and reproducibility: reports/verification/article-298-creative-collision-paper-artifact-corpus-normalisation-moral-metric-and-reproducibility-audit.json