BILLY: Steering Large Language Models via Merging Persona Vectors for Creative Generation

Trait induction and control2026ACL AnthologyApproved editorial review

Authors: Tsung-Min Pai, Jui-I Wang, Li-Chun Lu, Shao-Hua Sun, Hung-Yi Lee, Kai-Wei Chang

Keywords: Behavioral guardrails, Dynamic LLM personas, ORPO LoRA adapters, Synthetic preference data, LLM-as-judge

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

6
Authors
6
Findings
15
Limitations
6
Evidence

Editorial summary

English

BILLY replaces an online multi-agent conversation with an activation intervention in a single LLM. For each professional role, Claude Sonnet 4 generates five positive system prompts; the model answers 20 role-eliciting questions under those prompts and a neutral prompt. GPT-4o-mini scores role expression from 0 to 100. In the written method, positive-prompt responses above 50 and neutral responses below 50 are retained, token-averaged residual-stream activations are computed, and the neutral mean is subtracted from the positive mean at every layer. BILLY then takes the arithmetic mean of several vectors and adds alpha times that vector during generation. Main experiments use layer 20 and alpha 2.0. The nominal four-vector default is Creative Professional, Environmentalist, Futurist and Futurist: Futurist is duplicated and receives half the average weight, so this is not four distinct perspectives.

Evaluation uses Qwen2.5-7B-Instruct, Llama-3.1-8B-Instruct and Gemma-3-4B-it on four GPT-4-augmented creativity tasks: Alternative Uses, Instances, Similarities and Scientific Creativity, with 100-question files in the repository. Models produce five ideas per prompt. Baselines are one agent at temperature 0.7, one at 1.0, a multi-role prompt, and four-agent five-round LLM Discussion. GPT-4o-mini rates originality and elaboration from 1 to 5. In Table 2, BILLY has the highest originality point estimate in 10 of 12 model-task cells: all eight Qwen/Llama cells and Gemma AUT/Instances. It does not lead Gemma Scientific, 4.94 versus 4.96 for SA and SA-MRP, or Gemma Similarities, 4.94 versus 4.97 for SA. Elaboration is competitive but loses several rows. Gemma scores show ceiling effects, while Gemma LLM Discussion has unstable, high-variance outputs. Evidence therefore favors this intervention under the chosen rubric and tasks, but is narrower than an unqualified claim of surpassing traditional approaches.

The statistical claim is auditable. Released scripts run two-sided paired t-tests between BILLY and four baselines for each available task, model and metric; LLM Discussion scores are averaged in consecutive groups of four agents. Retrieving the 55 official Git LFS JSON files reproduces 88 tests: 65 have p<0.05 (73.86%) and 58 have p<0.001 (65.91%), matching the rounded 74% and 66% statements. AUT uses n=98 because two answers are missing; other cells use n=100. There is no correction for 88 comparisons, preregistered family, interval or effect size. A global Bonferroni 0.05/88 threshold retains 57 tests (64.77%). The published percentages are numerically real, but significance counts do not communicate magnitude or show universal advantage; 23 comparisons are nonsignificant at 0.05.

Human evaluation compares SA, LLM Discussion and BILLY across all four tasks. Three outputs per task are selected and eleven volunteers rate originality and elaboration. The paper's 132 evaluation scores correspond to 11 raters by 12 outputs; once three methods and two metrics are represented, the CSV contains 11×72=792 scalar values. BILLY has the highest mean in all eight task-metric rows. Correlations between aggregate human and LLM scores are Spearman 0.73/Pearson 0.66 for originality and 0.43/0.40 for elaboration, based on only twelve aggregate points per metric. The decisive limitation is released but omitted from the main paper: all six Krippendorff alpha values range from 0.1253 to 0.2279, far below the repository's own 0.667 acceptability threshold. Pairwise Kendall means range 0.083-0.231 for originality and 0.374-0.582 for elaboration. Human means favor BILLY, but rater consensus is too weak to support stable human validation.

Efficiency is compelling against multi-agent discussion, not against every alternative. Table 5 reports per query: SA 23.9 s, 22.3/407.1 tokens and $0.25 per 10,000; SA-MRP 36.8 s, 221.2/861.1 and $0.56; LLM Discussion 513 s, 88,853/12,922.1 and $25.50; BILLY 19 s, 62.2/475.6 and $0.30. The greater-than-95% saving is relative to LLM Discussion, with roughly 25-fold lower latency. BILLY is nevertheless 20% more expensive than SA by the paper's own token prices, while reporting 20% lower latency. Pricing uses Nebius token rates and amortizes persona-generation inputs, but excludes GPU cost, extraction time, model loading and activation-hook overhead. It does not establish BILLY as the absolute cheapest method.

Composition studies show interaction rather than clean additive control. The one-to-seven-vector curve changes identity and count together: every condition from two upward includes Creative Professional, and the default duplicates Futurist. It cannot identify a causal effect of persona count. On twelve neutral prompts, Gemini 2.5 Pro, Gemini 2 Flash and GPT-4o-mini distribute exactly 100 points among five ad hoc roles. Creative-only raises Creative to 43.2 and Environmentalist-only raises Environmental to 30.8. But Creative+Environmentalist scores 43.9 Creative and only 12.3 Environmental, below both Environmentalist-only and SA's 14.9. The authors acknowledge that creative wording dominates judging. This contradicts reliable simultaneous output control even when internal projections are positive on both directions. Projection is also not independent validation: the activation change created by adding an average of persona vectors is projected onto those same vectors. It confirms the injected direction, not psychological meaning, creative-quality causation or expressed semantics.

The official repository is unusually useful: it publishes 92 syntactically valid Python files, four datasets, 50 serialized vectors, human ratings, prompts, projection tables and enough results to reproduce the significance percentages. It is not a clean end-to-end reproduction. The README expands BILLY incorrectly, claims MIT without a LICENSE, and its quick start names six missing paths. A normal clone leaves 140 JSON files as Git LFS pointers, but the README omits git-lfs pull. There are no tests, CI, container, paper-locked environment or seeds, and several author-machine paths remain. The public generator uses adaptive median thresholds, coherence>=50, a three-sample minimum and first-20 truncation, materially differing from the paper's >50/<50 rule. The API generates its response before adding the optional multi-role prompt, /reset calls a commented-out method, persona modes fail with four vectors, and advertised fusion variants are incomplete or unsafe. The faithful conclusion is that BILLY provides promising evidence that one activation intervention improves judged originality on these benchmarks and radically lowers cost against four-agent discussion. It does not establish coherent personality, true creativity, independent interpretability, reliable human consensus or broad generalization; the artifacts enable partial audit, not yet a documented production-grade reproduction.

Español

BILLY sustituye una conversación multiagente por una intervención en las activaciones de un único LLM. Para cada rol profesional, Claude Sonnet 4 genera cinco system prompts positivos; el modelo responde a 20 preguntas que pretenden elicitar ese rol bajo esos prompts y bajo un prompt neutral. GPT-4o-mini puntúa la expresión del rol de 0 a 100. Según el método escrito, se conservan respuestas positivas con score mayor de 50 y neutrales con score menor de 50; se promedia la activación del residual stream a través de tokens y se resta la media neutral de la positiva en cada capa. BILLY toma después la media aritmética de varios vectores y la suma, multiplicada por alfa, a las activaciones durante la generación. Los experimentos principales usan capa 20 y alfa 2,0. El default denominado de cuatro vectores es Creative Professional, Environmentalist, Futurist y Futurist: Futurist está duplicado y recibe la mitad del peso, por lo que no son cuatro perspectivas distintas.

La evaluación usa Qwen2.5-7B-Instruct, Llama-3.1-8B-Instruct y Gemma-3-4B-it en cuatro tareas de creatividad ampliadas previamente con GPT-4: Alternative Uses, Instances, Similarities y Scientific Creativity, con ficheros de 100 preguntas en el repositorio. Los modelos producen cinco ideas por pregunta. Los baselines son un agente a temperatura 0,7, el mismo a 1,0, un agente con prompt multirrol y LLM Discussion, que emplea cuatro agentes durante cinco rondas. GPT-4o-mini puntúa originalidad y elaboración de 1 a 5. En Table 2, BILLY obtiene el mayor punto de originalidad en 10 de 12 cruces modelo-tarea: los ocho de Qwen/Llama y AUT/Instances con Gemma. No lidera Gemma Scientific, 4,94 frente a 4,96 de SA y SA-MRP, ni Gemma Similarities, 4,94 frente a 4,97 de SA. La elaboración es competitiva, pero pierde varias filas. En Gemma hay fuerte efecto techo y LLM Discussion sufre salidas inestables y gran varianza, algo que los propios autores atribuyen al contexto largo del modelo pequeño. Por tanto, la evidencia favorece la intervención para esta rúbrica y estas tareas, pero es más estrecha que «supera los enfoques tradicionales» sin matices.

Los tests estadísticos sí pueden auditarse. Los scripts hacen t-tests pareados bilaterales entre BILLY y cuatro baselines en cada combinación disponible de tarea, modelo y métrica; para LLM Discussion promedian grupos consecutivos de cuatro agentes. Al recuperar los 55 JSON oficiales almacenados con Git LFS se reproducen 88 contrastes: 65 tienen p<0,05 (73,86 %) y 58 p<0,001 (65,91 %), que redondean a los 74 % y 66 % del texto. AUT tiene n=98 por dos respuestas ausentes; el resto n=100. No hay corrección por 88 comparaciones, familia preregistrada, intervalos ni tamaños de efecto. Un Bonferroni global 0,05/88 conserva 57 tests (64,77 %). La cifra publicada es numéricamente real, pero contar significaciones no informa de magnitud ni demuestra ventaja en todos los casos; 23 comparaciones no alcanzan 0,05.

La evaluación humana compara SA, LLM Discussion y BILLY en las cuatro tareas. Se seleccionan tres salidas por tarea y once voluntarios asignan originalidad y elaboración. El paper llama a esto 132 evaluation scores: corresponde a 11 evaluadores por 12 salidas; al desplegar tres métodos y dos métricas, el CSV contiene 11×72=792 valores escalares. BILLY logra la media más alta en las ocho filas tarea-métrica. Las correlaciones entre medias humanas y juez LLM son Spearman 0,73/Pearson 0,66 para originalidad y 0,43/0,40 para elaboración, calculadas sobre solo doce puntos agregados por métrica. La limitación decisiva aparece en los propios artefactos pero no en el cuerpo principal: los seis alfas de Krippendorff están entre 0,1253 y 0,2279, muy por debajo del umbral 0,667 que el script define como aceptable. Las medias de Kendall entre parejas son 0,083-0,231 en originalidad y 0,374-0,582 en elaboración. Así, las medias favorecen a BILLY, pero no existe consenso humano fiable suficiente para hablar de una validación estable.

El análisis de eficiencia es fuerte frente al baseline multiagente, pero no frente a cualquier alternativa. Table 5 informa por consulta: SA 23,9 s, 22,3/407,1 tokens y 0,25 dólares por 10.000; SA-MRP 36,8 s, 221,2/861,1 y 0,56; LLM Discussion 513 s, 88.853/12.922,1 y 25,50; BILLY 19 s, 62,2/475,6 y 0,30. La reducción mayor del 95 % se refiere a LLM Discussion y la latencia es aproximadamente 25 veces menor. Sin embargo, BILLY cuesta un 20 % más que SA con la propia tarifa del paper, aunque su latencia reportada es un 20 % menor. El coste usa precios por token de Nebius y amortiza inputs de creación de vectores, pero excluye GPU, tiempo de extracción, carga del modelo y overhead de intervenir activaciones. No sostiene que sea el método absolutamente más barato.

Los análisis de composición muestran interacciones, no control aditivo limpio. La curva de uno a siete vectores cambia simultáneamente identidad y cantidad: todas las condiciones desde dos incluyen Creative Professional y el default duplica Futurist. No permite estimar un efecto causal del número de personas. En doce prompts neutrales, Gemini 2.5 Pro, Gemini 2 Flash y GPT-4o-mini reparten exactamente 100 puntos entre cinco roles ad hoc. El vector Creative eleva Creative a 43,2 y Environmentalist eleva Environmental a 30,8. Pero la fusión Creative+Environmentalist obtiene 43,9 en Creative y solo 12,3 en Environmental, por debajo tanto de Environmentalist solo como del 14,9 de SA. Los autores admiten que el vocabulario creativo domina al juez. Esto contradice un control simultáneo fiable en el texto, aunque las activaciones se proyecten positivamente sobre ambas direcciones. Esa proyección tampoco es validación independiente: se proyecta el cambio causado por añadir una media de vectores sobre esos mismos vectores. Confirma la dirección inyectada, no su significado psicológico, la causalidad de la calidad creativa ni la expresión observada.

El repositorio oficial es valioso: publica 92 scripts Python sintácticamente válidos, cuatro datasets, 50 vectores serializados, ratings humanos, prompts, tablas de proyección y resultados suficientes para reproducir porcentajes. Pero no es una reproducción extremo a extremo limpia. El README expande mal BILLY, afirma licencia MIT sin que exista LICENSE y dirige el quick start a seis rutas ausentes. Ciento cuarenta JSON quedan como punteros en un clon sin Git LFS y no se documenta git-lfs pull. No hay tests, CI, contenedor, entorno bloqueado o seeds. Hay rutas absolutas de máquinas de autores. El generador publicado usa umbrales adaptativos basados en medianas, exige coherence>=50, conserva al menos tres casos y trunca a los primeros 20, distinto del >50/<50 descrito. La API genera la respuesta antes de añadir el supuesto prompt multirrol, /reset llama a un método comentado, los modos de persona fallan con cuatro vectores y tres métodos de fusión anunciados no implementan una alternativa estable. La conclusión fiel es que BILLY aporta evidencia prometedora de que una intervención de activación concreta mejora puntuaciones de originalidad en estos benchmarks y reduce radicalmente coste frente a una discusión de cuatro agentes. No demuestra personalidad coherente, creatividad verdadera, control interpretable independiente, consenso humano fiable ni generalización; sus artefactos permiten auditoría parcial, no todavía reproducción documentada de producción.

Research question

Can the mean of several activation vectors associated with professional roles transfer part of the creativity improvement attributed to a multiagent discussion to a single inference, with lower latency and cost?

Method

Extract by contrast role vectors from five positive prompts, 20 questions, neutral answers and GPT-4o-mini scores; average vectors of several roles and add the resulting direction at layer 20 with coefficient 2. Evaluate three models on four tasks of 100 questions against four baselines, with GPT-4o-mini judge, paired t-tests, a human sample of eleven volunteers and analysis of composition, projection, traits and cost.

Sample: Three open models by four tasks of up to 100 questions; five methods in the LLM table; 88 paired t-tests with n=98 or 100; human evaluation of three methods, three outputs per task, four tasks, two metrics and eleven volunteers; twelve neutral prompts for role expression; ten neutral questions and ten random samples for projection.

Findings

  • BILLY leads originality in 10 of 12 model-task crosses, not in the two Gemma Scientific and Similarities crosses; elaboration is competitive but not uniformly superior.
  • The 55 official JSON files reproduce 65/88 tests with p<0.05 and 58/88 with p<0.001; 57 survive a global Bonferroni, although the paper does not correct or report effect sizes.
  • BILLY has the highest human mean in eight rows, but all published Krippendorff alphas are between 0.1253 and 0.2279, insufficient reliability according to the script itself.
  • Against LLM Discussion, BILLY reports 19 s versus 513 s and 0.30 versus 25.50 dollars per 10,000 queries; against SA it is faster but more expensive per tokens.
  • The Creative+Environmentalist fusion does not preserve Environmentalist expression in the output rubric: 12.3, compared to 30.8 with ENV alone and 14.9 in SA.
  • The repository allows substantive audit of results, but its quick start/API and its exact correspondence with the written method are not resolved.

Limitations

  • Single principal judge GPT-4o-mini, without independent family replication for Table 2.
  • Rubrics 1-5 close to the ceiling, especially in Gemma, sensitive to style, length and judge preferences.
  • Eighty-eight tests without multiple correction, intervals, effect sizes or preregistered family.
  • AUT has 98 observations, but the text generalizes tasks of 100 without explaining the two absences.
  • Human-LLM correlations computed over only twelve means per metric and without intervals.
  • Inter-rater human reliability very low and omitted from the main body of the paper.
  • Selection of three human samples per task without seed or randomization detail.
  • Comparison of number of vectors confounded with identity and duplication of Futurist.
  • Projection onto the same injected vectors, without independent semantic validation.
  • Twelve ad hoc prompts and forced distribution of 100 points for role expression, not a validated psychological instrument.
  • Cost excludes extraction, GPU, loading and intervention; the >95% advantage only applies to the multiagent baseline.
  • Extraction code uses thresholds and filters different from the paper method.
  • No code/data LICENSE despite README claiming MIT; the CC BY license corresponds to the paper.
  • Quick start with six missing paths, Git LFS not documented, no tests, CI, container, seeds or locked environment.
  • No safety study, role stereotypes, persistence, long conversation or undesired effects.

What the study does not establish

  • That BILLY captures collective intelligence or all the benefits of multiagent systems.
  • A human personality, coherent, stable or psychometrically valid within the model.
  • That GPT-4o-mini distinguishes true creativity from diversity, style, length or judge preference.
  • Reliable human consensus or stable superiority among evaluators.
  • Significant and practically important advantage in every model, task, metric and baseline.
  • That increasing the number of people improves monotonically or causally.
  • Reliable simultaneous control of several people in the generated text.
  • That positive projection proves independent semantic meaning or causality of creativity.
  • That it is cheaper than a standard agent; Table 5 reports 0.30 versus 0.25 dollars.
  • Operation without training/extraction cost; requires generation, judgment and offline passes.
  • Generalization to other tasks, long conversations, closed models, other layers or sensitive contexts.
  • Licensed, tested and documented end-to-end reproduction from the current repository.

Traceability

Scope: Full text

Version: EACL 2026 Long Papers, ACL Anthology 2026.eacl-long.369; arXiv:2510.10157v2, submitted 2026-01-24; CC BY 4.0

Consulted source: https://arxiv.org/abs/2510.10157

Review: Codex 46-page visual, EACL/arXiv-v2 metadata, full-method, 55-Git-LFS-result, 88-test reproduction, human-IRR, vector-composition, projection, cost, repository-code, API and claim-boundary audit, 2026-07-16

Approval: Codex fidelity pass, 2026-07-16

English translation: approved, 2026-07-18

Models evaluated

  • Qwen2.5-7B-Instruct
  • Llama-3.1-8B-Instruct
  • Gemma-3-4B-it
  • GPT-4o-mini as creativity and extraction judge
  • Claude Sonnet 4 as positive-prompt and neutral-question generator
  • Gemini 2.5 Pro, Gemini 2 Flash and GPT-4o-mini as trait-distribution judges

Instruments and metrics

  • Wallach-Kogan Alternative Uses, Instances and Similarities tasks
  • Scientific Creativity task
  • GPT-4o-mini 1-5 originality and elaboration rubrics adapted from TTCT concepts
  • Two-sided paired t-tests over question-level LLM-judge scores
  • Human 1-5 originality and elaboration ratings
  • Krippendorff alpha and pairwise Kendall tau from released human ratings
  • Human-LLM Spearman and Pearson correlations over 12 aggregate points per metric
  • Residual-stream self-projection onto extracted persona vectors
  • Five-role forced 100-point trait-distribution rubric

Data used

  • GPT-4-augmented AUT: 100 prompts, 98 usable in released statistical files
  • GPT-4-augmented Instances: 100 prompts
  • GPT-4-augmented Similarities: 100 prompts
  • GPT-4-augmented Scientific Creativity: 100 prompts
  • Persona extraction data: five positive prompts and 20 trait-eliciting questions per role plus neutral responses
  • Human_Evaluation matrix: 11 raters by 72 scalar cells
  • Official Git LFS generation/evaluation JSONs and 50 serialized persona-vector files

Evidence and location

  • Publication metadata, authors, venue, DOI and pages: ACL Anthology 2026.eacl-long.369
  • Method, tables, appendices, license and limitations: arXiv:2510.10157v2 PDF, 46 pages, sha256 33daf21b58e161e2ca571774af7012d5847b81e27832116a1254a50b863ac9fa
  • Code, vectors, data, human ratings and statistical results: Bai1026/LLM_Persona commit c8f81acb562897b13d478766f65e946b3f3eceb8
  • Reproduction of 88 t-tests and percentages 74%/66%: 55 official Git LFS JSON result files under LLM_Discussion/Experiments/Rebuttal
  • Human reliability and structure of 11 evaluators: Human_Evaluation/data/human_scores.csv, krippendorff_alpha_results.csv and kendall_tau_results.csv
  • Audit of claims, statistics, code, data, cost and reproducibility: reports/verification/article-253-eacl-billy-persona-vector-creativity-statistics-code-data-and-claim-audit.json