An evolutionary model of personality traits related to cooperative behavior using a large language model

Society, culture, and collective behavior2024Scientific ReportsApproved editorial review

Authors: Reiji Suzuki, Takaya Arita

Keywords: evolutionary computation, personality traits, cooperative behavior, large language models, game theory, evolutionary dynamics, behavioral traits

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

2
Authors
14
Findings
26
Limitations
14
Evidence

Editorial summary

English

This paper proposes an evolutionary model in which an English sentence of roughly ten words acts as a personality “gene” related to cooperation or selfishness. Llama-2-13B-Chat-GPTQ does not play every round as a conversational agent: it is used once per unique sentence to compile that sentence into a deterministic phenotype, a 16-action C/D lookup table covering all possible two-round histories of the iterated Prisoner's Dilemma. If a response contains no identifiable choice, generation is retried up to ten times and then a random action is assigned. The table is cached for every later occurrence of that exact sentence. The LLM re-enters at mutation: with probability .05 it rewrites the gene toward a randomly selected cooperative or selfish tone. The simulation uses 30 individuals, 1,000 generations, 20-round round-robin games, .05 action noise, and payoffs R=4, T=5, S=0, P=1. Seven initial genes generated by GPT-4 span a selfish-to-cooperative gradient. To examine initial mapping, each gene is compiled 50 times at temperature .9; extremes average roughly .1 and .7 cooperation, while intermediate descriptions vary more. That variability is then omitted: each newly encountered gene retains one deterministic sample. Across 15 evolutionary trials, mean cooperation in the language model is .31, with a mode near .18 and a tail toward higher values; mean cosine distance between population-average genes 20 generations apart is .05, with a strong peak near .02. The authors interpret this as long defection-dominated stagnation interrupted by cooperative invasions. One illustrative run stays near .05 cooperation until generation 300, rises to .55, peaks near .75 around generation 850, and drops to .15 near generation 900. A control directly encodes the 16 actions and mutates by flipping one or two; it reaches .50 mean cooperation and .07 cosine distance, with smoother and more similar trajectories across trials. This comparison shows that linguistic rewriting creates irregular mutation sizes and a different evolutionary topology; it does not isolate “personality,” because representation, mutation operator, and genotype-phenotype mapping all change together. The lexical analysis associates frequent words with outcomes: “gently,” “fosters,” “establishes,” and “harmony” occur in high-cooperation genes; “trampling,” “trumps,” “disregard,” “blatant,” and “skepticism” occur with defection. These are associations among words co-occurring in selected sentences, not causal effects of individual tokens; there is no ablation, regression, or context control. The actual contribution is showing that a quantized LLM can act as a semantic operator inside an artificial-life model and produce nontrivial dynamics. It does not demonstrate the evolution of human personality or psychological traits in agents. “Personality” is operationalized as phrases about cooperation and self-interest, without a Big Five instrument, human validation, or correspondence with real behavior; the authors explicitly acknowledge that this correspondence remains unknown. Reproducibility is limited: code and data are not released, only available on request, and the paper supplies no seeds, environment, exact GPTQ model revision or branch, library versions, GPT-4 prompt used to create the initial genes, parsing-failure rate, or fallback rate. The model repository remains accessible and offers multiple quantizations, but the article does not pin the one used. The findings are therefore exploratory evidence about the evolution of linguistic representations compiled by an LLM in one game, not psychological or social evidence about people.

Español

Este trabajo propone un modelo evolutivo en el que una frase inglesa de unas diez palabras funciona como “gen” de personalidad relacionado con cooperación o egoísmo. Llama-2-13B-Chat-GPTQ no juega cada ronda como agente conversacional: se usa una vez por cada frase única para compilarla en un fenotipo determinista, una tabla de 16 acciones C/D que cubre todas las historias posibles de dos rondas del dilema del prisionero. Si una respuesta no contiene una elección identificable, se regenera hasta diez veces y después se asigna una acción aleatoria. La tabla queda cacheada para todas las apariciones futuras de esa frase. El LLM vuelve a intervenir en la mutación: con probabilidad 0,05, reescribe el gen hacia un tono cooperativo o egoísta elegido al azar. La simulación usa 30 individuos, 1.000 generaciones, torneos round-robin de 20 rondas, ruido de acción 0,05 y payoffs R=4, T=5, S=0 y P=1. Siete genes iniciales generados por GPT-4 cubren una gradación de egoísmo a cooperación. Para comprobar el mapeo inicial, cada gen se compila 50 veces con temperatura 0,9; los extremos producen, en promedio, aproximadamente 0,1 y 0,7 de cooperación, mientras las descripciones intermedias son más variables. Sin embargo, esa variabilidad se omite después: cada gen nuevo conserva una única muestra determinista. En 15 ensayos evolutivos, la cooperación media del modelo lingüístico es 0,31, con una moda cercana a 0,18 y cola hacia valores altos; la distancia coseno media entre genes promedio separados 20 generaciones es 0,05, con fuerte pico alrededor de 0,02. Los autores interpretan esto como largos estancamientos dominados por defección interrumpidos por invasiones cooperativas. Un ensayo ilustrativo pasa de aproximadamente 0,05 de cooperación hasta la generación 300, sube a 0,55, alcanza cerca de 0,75 hacia la 850 y cae a 0,15 hacia la 900. El control codifica directamente las 16 acciones y muta volteando una o dos: logra cooperación media 0,50 y distancia coseno 0,07, con trayectorias más graduales y parecidas entre ensayos. La comparación muestra que introducir una reescritura lingüística crea mutaciones de tamaño irregular y una topología evolutiva distinta; no aísla “personalidad”, porque también cambia simultáneamente representación, operador de mutación y mapeo genotipo-fenotipo. El análisis léxico asocia palabras frecuentes con resultados: “gently”, “fosters”, “establishes” y “harmony” aparecen en genes con cooperación alta; “trampling”, “trumps”, “disregard”, “blatant” y “skepticism” con defección. Son asociaciones entre palabras que coexisten dentro de frases seleccionadas, no efectos causales de tokens individuales; no hay ablación, regresión ni control por contexto. La contribución real es demostrar que un LLM cuantizado puede servir como operador semántico dentro de un modelo de vida artificial y producir dinámicas no triviales. No demuestra evolución de personalidad humana ni que los agentes posean rasgos psicológicos. “Personalidad” se operacionaliza como frases sobre cooperación/beneficio propio, sin instrumento Big Five, validación humana o correspondencia con conducta real; los autores reconocen que esa correspondencia queda sin comprobar. La reproducibilidad es limitada: no se publican código ni datos, solo disponibles a petición, ni semillas, entorno, commit o rama exacta del modelo GPTQ, versión de librerías, prompt de GPT-4 que creó los genes iniciales, fallos de parsing o tasas de fallback. El repositorio de modelo sigue accesible y ofrece múltiples cuantizaciones, pero el artículo no fija cuál usó. Por tanto, los resultados son evidencia exploratoria sobre evolución de representaciones lingüísticas compiladas por un LLM en un juego muy concreto, no evidencia psicológica o social sobre personas.

Research question

Can an LLM map linguistic descriptions of cooperation and selfishness to prisoner's dilemma strategies and act as a mutation operator to generate evolutionary dynamics distinct from a direct encoding of actions?

Method

Each unique textual gene is compiled with Llama-2-13B-Chat-GPTQ into a deterministic table of 16 actions. A population of 30 tables evolves for 1,000 generations via roulette-wheel selection by payoff in tournaments of 20 rounds, textual mutation with LLM and action noise. 15 trials and a control with a binary genotype of 16 actions are run; MiniLM embeddings, UMAP, cooperation, cosine distance and lexical associations describe the trajectories.

Sample: Fifteen trials with populations of 30 individuals over 1,000 generations and fifteen control trials. Each fitness comes from a round-robin tournament of 20-round matches. The initial mapping is explored with 50 generations of phenotype for each of seven genes, but evolution caches a single table per unique gene.

Findings

  • The seven initial genes ordered from selfish to cooperative produce, on average, action tables consistent with that gradation.
  • Extreme genes produce more consistent strategies than intermediate ones in the 50 initial samples.
  • Each unique gene is then reduced to a deterministic table of 16 actions and reused.
  • Mean cooperation across the 15 linguistic trials is 0.31.
  • The linguistic distribution has its main peak near 0.18 and a tail toward high cooperation.
  • The mean cosine distance between intervals of 20 generations is 0.05.
  • The trajectories show long stagnations and abrupt shifts between defection and cooperation.
  • The binary control reaches mean cooperation 0.50 and cosine distance 0.07.
  • The control evolves more gradually and with lower spatial variation across trials.
  • Words of harmony and collaboration are descriptively associated with high cooperation.
  • Words of contempt or self-interest are descriptively associated with defection.
  • The paper acknowledges that it is not known whether the LLM mapping coincides with human relationships between personality and behavior.
  • The model's Hugging Face repository remains accessible, but offers multiple quantizations and the paper does not fix a revision.
  • There is no official repository of code or data for the experiment.

Limitations

  • Personality is reduced to phrases of cooperation and selfishness, without psychological measurement.
  • Big Five dimensions are not used or validated despite being cited as motivation.
  • There are no human participants or annotators who validate phrase, strategy or interpretation.
  • The LLM does not act during each round; it compiles a table once.
  • A stochastic sample at temperature 0.9 is frozen as a deterministic phenotype.
  • The regeneration and random assignment rates after ten failures are not published.
  • Caching by exact string can treat nearly identical paraphrases as distinct genes.
  • The control changes representation, mapping and mutation operator at the same time.
  • Linguistic mutations do not have a controlled distance comparable to flipping one or two bits.
  • The seven initial genes depend on GPT-4, whose version and prompt are not specified.
  • Only one quantized local LLM and a single game type are used.
  • There is no robustness to other models, prompts, temperatures, payoffs or population sizes.
  • Fifteen trials do not include intervals, inferential tests or power analysis.
  • The choice of trial 13 for the narrative does not follow a predeclared criterion.
  • UMAP is a non-unique projection and does not necessarily preserve global structure.
  • Words are analyzed by marginal association within highly correlated phrases.
  • There is no ablation showing that an isolated word causes the attributed behavior.
  • The threshold of 500 appearances is arbitrary and not statistically justified.
  • Table 1 contains the typographical error 00.404.
  • No code, datasets, seeds or execution environment are published.
  • No commit, branch or exact GPTQ configuration of the model is fixed.
  • Versions of Transformers, AutoGPTQ, Sentence Transformers or UMAP are not documented.
  • Data are only available on reasonable request.
  • Inference from an artificial simulation to human behavioral evolution is speculative.
  • Generated phrases inherit semantic biases from GPT-4 and Llama 2.
  • Effects of safety, culture, language or demographic variation are not evaluated.

What the study does not establish

  • It does not demonstrate that agents have personality.
  • It does not demonstrate evolution of human psychological traits.
  • It does not validate a Big Five measure.
  • It does not demonstrate that phrases predict real human behavior.
  • It does not demonstrate that an LLM simulates human decisions.
  • It does not demonstrate that individual words cause cooperation or defection.
  • It does not demonstrate that linguistic mutations are biologically plausible.
  • It does not demonstrate spontaneous emergence independent of the cooperative/selfish bias imposed on the mutator.
  • It does not demonstrate superiority of the linguistic model; the control obtains higher mean cooperation.
  • It does not generalize to other games, models, languages or societies.
  • It does not establish stability under another sample of the same gene.
  • It does not allow exact reproduction of the figures from public artifacts.
  • It does not demonstrate consciousness, intention, preferences or understanding of the LLM.
  • It does not demonstrate that higher cooperation is always higher fitness or better individual outcome.

Traceability

Scope: Full text

Version: Scientific Reports 14:5989, version of record published 19 Mar 2024, DOI 10.1038/s41598-024-55903-y

Consulted source: https://www.nature.com/articles/s41598-024-55903-y.pdf

Review: Codex full-text, bilingual-fidelity, visual, evolutionary-game, LLM-genotype-phenotype, prompt, control, lexical-association, personality-construct, model-artifact, statistics, reproducibility and data-availability audit, 2026-07-15

Approval: Codex fidelity pass, 2026-07-15

English translation: approved, 2026-07-18

Models evaluated

  • TheBloke/Llama-2-13B-chat-GPTQ
  • GPT-4 for generating seven initial personality descriptions
  • sentence-transformers/paraphrase-MiniLM-L6-v2 for embeddings
  • UMAP for two-dimensional visualization

Instruments and metrics

  • Iterated Prisoner's Dilemma
  • 16-action deterministic strategy table with memory of two rounds
  • Roulette-wheel evolutionary selection
  • LLM semantic mutation operator
  • Proportion of cooperation
  • Cosine distance between population-average gene embeddings
  • Word-outcome association table
  • Direct behavioral-genotype control

Data used

  • Seven GPT-4-generated initial personality descriptions
  • 15 evolutionary trials of 1,000 generations
  • 15 direct-genotype control trials
  • Generated natural-language mutation lineages
  • No public code or experiment dataset; available on reasonable request

Evidence and location

  • Official publication: Scientific Reports 14:5989, published 19 Mar 2024, DOI 10.1038/s41598-024-55903-y
  • Complete source: .cache/editorial-sources/article-108/source.pdf; 9 pages; sha256 5b8a9df601e311ee81dbee60a8ebc8e6b9e27565f88e5bf35d6eed59705dfa84
  • Genotype-phenotype mapping: Full text p. 3, Methods and Figure 1
  • Prompts and fallback: Full text pp. 3-4, Methods and Figure 2
  • Simulation parameters and models: Full text pp. 3-4, Experiments and analyses
  • Seven genes and 50 generations: Full text pp. 4-5, Figure 3
  • Illustrative evolutionary trial: Full text pp. 4-6, Figure 4
  • Aggregate results: Full text p. 6, Figure 5
  • Direct control: Full text pp. 6-7, Figure 6
  • Lexical associations: Full text pp. 7-8, Table 1
  • Recognized limits: Full text p. 8, Conclusion
  • Data availability: Full text p. 8, Data availability: on reasonable request
  • Current GPTQ model: https://huggingface.co/TheBloke/Llama-2-13B-chat-GPTQ; multiple quantization branches; checked 15 Jul 2026
  • Visual inspection: All 9 PDF pages rendered and visually inspected, including six figures, prompts and Table 1; checked 15 Jul 2026