Character is Destiny: Can Role-Playing Language Agents Make Persona-Driven Decisions?

Personas, identity, and agents2025ACL AnthologyApproved editorial review

Authors: Rui Xu, Xintao Wang, Jiangjie Chen, Siyu Yuan, Xinfeng Yuan, Jiaqing Liang, Zulong Chen, Xiaoqing Dong, Yanghua Xiao

Keywords: Large Language Models, Personality, Persona, Model Evaluation, LLM Evaluation

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

9
Authors
13
Findings
20
Limitations
10
Evidence

Editorial summary

English

The Findings of EMNLP 2025 paper introduces LIFECHOICE as a narrative decision-reconstruction benchmark: given a character, context preceding a decision, a scenario, and four options, a role-playing agent must select the action the character took in the novel. The paper reports 2,512 decision points from 470 books. It uses Supersummary character descriptions, chapter summaries, and book analyses as expert-written material. GPT-4 selects decisions, motivations, and relevant chapters and, while seeing the gold decision and motivation, constructs the scenario, question, correct option, and three distractors. Ten English-speaking university students filter items, and GPT-4o assigns one motivation category per example. Profile methods include GPT-3.5 recursive and progressive summaries, expert descriptions, BM25 or `text-embedding-ada-002` memories, direct concatenation, and CHARMAP. CHARMAP locates question-relevant episodes in a description and uses them to retrieve memories. In a single run per condition, the best reported accuracy is 65.28% for o1+CHARMAP versus 61.12% for o1+direct concatenation, a descriptive 4.16-point difference. The paper also reports 92.01% human accuracy with raw text, but this comes from three students and six unfamiliar novels under a small protocol with no exact question count, assignment design, agreement, or uncertainty; it is not a matched benchmark-wide ceiling. Audit of the official repository materially changes the interpretation. It releases only 1,576 records from 383 books, 62.74% of reported items and 81.49% of books, and omits original books, CHARMAP, profile construction, retrieval, outputs, analysis, and the exact environment. Its only script tests older or generic models unlike those in the paper, puts only the name under Description, treats `input_text` as Memory, duplicates option letters, and saves no predictions. The public subset has ten invalid rows and extreme position imbalance: B is correct 837 times (53.11%), C 618 (39.21%), A 48, and D 63. Always choosing B scores 53.11%, above or near many description-only and memory-only results and far from a uniform 25% chance assumption. Whether the full set has the same imbalance cannot be checked because it is unreleased. Direct semantic leakage also appears: the first scenario says “Upon refusal, Leduc threatens to expel her” and then asks what Amelia decides; the gold answer is to reject Leduc. Because GPT-4 saw the decision and motivation while writing questions and the paper lacks question-only, random-profile, shuffled-character, option-order, or answer-blind rewrite controls, accuracy can measure wording and generator cues in addition to persona understanding. Claims that CHARMAP “significantly advances” performance and model gaps are “insignificant” have no statistical tests, intervals, repeated runs, or variance. The contamination analysis uses Douban popularity and 30 post-cutoff books as proxies; it cannot establish absence of plots or Supersummary text from training. The defensible contribution is an interesting narrative role-playing benchmark and a promising retrieval method; the released evidence cannot reproduce its tables or establish personality understanding, human fidelity, real-world choice prediction, or statistically reliable superiority.

Español

La versión publicada en Findings of EMNLP 2025 presenta LIFECHOICE como un benchmark de reconstrucción de decisiones narrativas: dado un personaje, el contexto anterior a una decisión, un escenario y cuatro opciones, un agente de role-playing debe elegir la acción que el personaje tomó en la novela. El paper declara 2.512 puntos de decisión de 470 libros. Usa contenido de Supersummary, descripciones de personajes, resúmenes de capítulos y análisis de libros, como material escrito por expertos literarios. GPT-4 selecciona decisiones, motivaciones y capítulos y, viendo la decisión correcta y su motivación, construye el escenario, la pregunta, una opción correcta y tres distractores. Diez estudiantes universitarios angloparlantes filtran los ítems; GPT-4o asigna una categoría motivacional por ejemplo. Para construir perfiles compara resúmenes recursivos y progresivos con GPT-3.5, descripciones expertas, memoria BM25 o embedding `text-embedding-ada-002`, concatenación directa y CHARMAP. CHARMAP localiza en la descripción episodios relevantes para la pregunta y los usa como consulta para recuperar memorias. En una sola ejecución por condición, la mejor precisión publicada es 65,28 % con o1+CHARMAP, frente a 61,12 % con o1+concatenación directa; la diferencia descriptiva es 4,16 puntos. El paper también informa 92,01 % humano leyendo texto original, pero ese resultado procede de tres estudiantes, seis novelas que no conocían y un protocolo pequeño cuya cantidad exacta de preguntas, asignación, acuerdo e intervalos no se publican; no es un techo emparejado sobre los 2.512 ítems. La auditoría del repositorio oficial cambia sustancialmente la lectura. Solo libera 1.576 registros de 383 libros, 62,74 % de los ítems y 81,49 % de los libros declarados, y omite libros, CHARMAP, construcción de perfiles, retrieval, salidas, análisis y entorno exacto. El único script prueba modelos antiguos o genéricos distintos de los del paper, coloca solo el nombre bajo Description, usa `input_text` como Memory, duplica las letras de opción y no guarda predicciones. El subconjunto público tiene diez filas inválidas y un sesgo de posición extremo: B es correcta 837 veces (53,11 %), C 618 (39,21 %), A 48 y D 63. Elegir siempre B logra 53,11 %, por encima o cerca de numerosos resultados de descripción o memoria y muy lejos del supuesto azar uniforme del 25 %. No puede saberse si el conjunto completo mantiene este sesgo porque no está publicado. Hay también fuga semántica directa: el primer escenario dice “Upon refusal, Leduc threatens to expel her” y después pregunta qué decide Amelia; la respuesta es rechazar a Leduc. Como GPT-4 recibió decisión y motivación al redactar las preguntas y no se publican controles question-only, de perfil aleatorio, personaje permutado, orden de opciones o reescritura ciega a la respuesta, la precisión puede medir claves del enunciado y del generador además de comprensión de persona. Las expresiones “significantly advances” y diferencias “insignificant” entre modelos carecen de pruebas, intervalos, repeticiones o varianza. El análisis de contaminación usa popularidad en Douban y 30 libros posteriores al cutoff como proxies; no demuestra ausencia de trama o de Supersummary en entrenamiento. La contribución defendible es un benchmark interesante de role-playing narrativo y un método de recuperación prometedor; la evidencia liberada no permite reproducir sus tablas ni establecer comprensión de personalidad, fidelidad humana, predicción de decisiones reales o superioridad estadísticamente fiable.

Research question

To what extent can LLMs and role-playing agents reconstruct the original decision of a fictional character from their preceding history, and does a specific scenario profile combining description and retrieved memory improve accuracy?

Method

LIFECHOICE transforms decisions from novels into four-option questions. GPT-4 selects decision points and motivations from Supersummary, locates chapters, and generates scenario, question, correct option, and distractors. Ten students filter questions with a 0-2 rubric across four dimensions and a mean threshold greater than six; GPT-4o classifies motivations. Summarized or expert descriptions, BM25/embedding memory, concatenation, and CHARMAP are compared with seven model families; one run and accuracy are reported, plus BLEU, ROUGE-L, and NLI in an appendix. The independent audit profiles the 1,576 public records, calculates position baselines, inspects the single script, and contrasts the released scope with the tables.

Sample: The paper declares 2,512 decision points from 470 books. The public artifact contains 1,576 unique book-character pairs from 383 books in 392 groups; the input texts have a median of 24,993 characters, p95 of 53,174, and maximum of 71,086. Ten rows are not valid A-D questions. The construction control uses ten students; the human comparison uses three students and six novels, with no exact n of questions published.

Findings

  • The authoritative source is the PDF of Findings of EMNLP 2025, DOI 10.18653/v1/2025.findings-emnlp.813, pages 15038-15059; its 22 pages were rendered and visually inspected.
  • The PDF is titled Character is Destiny: Can Role-Playing Language Agents Make Persona-Driven Decisions?, while the ACL record uses Can Persona-assigned Language Models Make Personal Choices?.
  • The best reported accuracy is o1+CHARMAP 65.28%, compared to o1+concatenation 61.12%; no inferential test is published for the 4.16 points.
  • Expert descriptions reach 53.04-56.90%, better than automatic summaries 38.23-48.92%; memory alone achieves 27.98-35.80%.
  • CHARMAP descriptively outperforms direct concatenation in the reported models and obtains 61.43-65.28%.
  • The small human evaluation publishes 92.01% with original text, 66.83% with concatenation, and 72.17% with CHARMAP; GPT-4o gives 60.07% and 64.22% in the latter two conditions.
  • The official repository only releases 1,576 of 2,512 records and 383 of 470 books, with no CHARMAP code or reproducible results.
  • In the public subset, B is gold in 53.11% and C in 39.21%; always-B achieves 53.11%, while A and D sum only 7.05%.
  • Ten public records are N/A or Not applicable, have a single option, and cannot be scored with the A-D evaluator.
  • The first public scenario filters the decision through the phrase Upon refusal before asking whether Amelia rejects the proposition.
  • The only public runner does not implement CHARMAP, does not use the main models from the paper, duplicates option labels, and does not preserve a per-item trace.
  • Table 15 scores Poor as 1 in Comprehensiveness, the same as Average, although the other dimensions and the declared design use 0-2.
  • The NLI of BM25 and embedding in Table 6 exactly match their MCQ accuracies, 27.98, 28.51, 35.19, and 35.80, without any artifact to clarify whether this is coincidence or error.

Limitations

  • The full set of 2,512 items and 470 books is not released; only the public subset is audited.
  • The release does not include books, CHARMAP, profile construction, retrieval, outputs, splits, analysis, or exact environment.
  • The always-B of 53.11% in the subset invalidates using 25% as a sufficient baseline; the full distribution is unknown.
  • Construction with GPT-4 knows the answer and motivation; there is at least one direct leak in the scenario.
  • There is no question-only baseline, without profile, random profile, permuted character, permuted options, or length/position control.
  • There is no blind rewriting to the answer or systematic audit of semantic or stylistic cues.
  • GPT-4 participates in construction and models from the same family participate in evaluation, with risk of style/provider circularity.
  • A single accuracy is reported per condition, with no seeds, repetitions, temperature, intervals, or variance.
  • Significantly advances and insignificant model gap are not supported by statistical tests.
  • The human evaluation uses only three students and six novels; n of questions, assignment, agreement, and intervals are missing.
  • The manual rubric does not publish per-item ratings, before/after counts, adjudication, or inter-rater reliability; it contains a scoring error.
  • The contamination analysis uses Douban reviews as a proxy and does not prove absence in pretraining or in Supersummary.
  • Entity replacement may preserve identifiable events and alter narrative signals; its effectiveness is not quantified.
  • Gender and motivations do not report cell sizes, intervals, or multiple correction; causal explanations are speculative.
  • The temporal analysis includes only 40 characters and five fractions without uncertainty.
  • BLEU, ROUGE-L, and XNLI entailment are not calibrated against human behavioral equivalence.
  • The exact matches between NLI and MCQ in four rows of Table 6 cannot be audited without outputs/code.
  • A fictional character's decision is a plot choice designed by an author, not an observation of real human behavior.
  • The title differs between PDF and ACL record; ACL concatenates Xiaoqing Dong as Xiaoqingdong.
  • The Apache-2.0 license of the repo does not by itself clarify the rights of content derived from third parties; no legal conclusion is formulated.

What the study does not establish

  • It does not demonstrate that LLMs possess personality, desires, values, or a human mind.
  • It does not demonstrate psychometric fidelity or trait stability.
  • It does not demonstrate that agents predict real people's decisions.
  • It does not separate persona understanding from cues in scenario, response position, generator style, or prior knowledge of the plot.
  • It does not establish that CHARMAP statistically outperforms concatenation or other models.
  • It does not establish a general human-model gap across all of LIFECHOICE.
  • It does not demonstrate absence of contamination from novels, summaries, or Supersummary.
  • It does not reproduce the main tables from the official repository.
  • It does not validate a 25% chance baseline given the highly imbalanced public distribution.
  • It does not prove that higher accuracy corresponds to better safety, autonomy, or utility in real personal agents.

Traceability

Scope: Full text

Version: Findings of EMNLP 2025, Anthology ID 2025.findings-emnlp.813, DOI 10.18653/v1/2025.findings-emnlp.813, pages 15038-15059, 22 pages; PDF title is Character is Destiny: Can Role-Playing Language Agents Make Persona-Driven Decisions?

Consulted source: https://aclanthology.org/2025.findings-emnlp.813/

Review: Codex complete bilingual full-text fidelity pass using the published Findings version, all-page visual inspection, official repository commit audit, exact public-data profiling, option-position baseline calculation, malformed-row audit, scenario leakage inspection, released-runner review, title and author reconciliation, statistical-claim review, contamination-control review, human-evaluation review, and construct-boundary assessment; summaries written from the paper and artifact evidence rather than abstract keywords, 2026-07-15

Approval: Codex fidelity pass, 2026-07-15

English translation: approved, 2026-07-18

Models evaluated

  • Llama-4-Maverick
  • qwen3-235b-a22b
  • Claude-3.5-Sonnet
  • DeepSeek-R1
  • gpt-4o
  • o1
  • gemini-2.5-pro
  • GPT-4 for decision selection, gold motivation extraction, chapter location, and multiple-choice construction
  • GPT-4o for motivation-category assignment and selected analyses
  • GPT-3.5 for recursive and progressive summary construction
  • Legacy public runner: gpt-3.5-turbo, generic gpt-4, claude-3-sonnet-20240229, and gemini-pro

Instruments and metrics

  • Four-option multiple-choice narrative decision reconstruction
  • CHARMAP scenario-specific description localization and memory retrieval
  • BM25 retrieval
  • OpenAI text-embedding-ada-002 retrieval
  • Manual four-dimension question rubric: comprehensiveness, logical consistency, challenge level, and alignment with character motivation
  • GPT-4o Aristophanes-inspired motivation taxonomy
  • Accuracy
  • BLEU and ROUGE-L for free behavior generation
  • mDeBERTa-v3-base-xnli entailment probability as NLI score
  • Independent public-subset option-distribution and always-B baseline audit

Data used

  • Reported LIFECHOICE: 2,512 fictional-character decision points from 470 books
  • Supersummary character descriptions, chapter summaries, and book analyses used under research authorization
  • Official GitHub public subset at commit 37266c38648cd56574b0cb2140857171a752cf38: 1,576 records, 383 books, 10 malformed records
  • Thirty popularity-stratified books and thirty post-August-6-2024 control books for reported leakage analysis
  • Small human-evaluation subset of six unfamiliar novels

Evidence and location

  • PDF title, authors, abstract, DOI, and scope: Published page 15038 and ACL Anthology record 2025.findings-emnlp.813 checked 15 July 2026
  • LIFECHOICE construction and manual control: Published pages 15040-15041 and 15052, Sections 3.1 and Appendix D
  • Methods, models, and main results: Published pages 15042-15045, Sections 4-5 and Tables 3-5
  • Human evaluation and declared training leak: Published pages 15043-15044, Table 4 and Analysis and Mitigation of Data Leakage
  • Generative evaluation and repeated values: Published pages 15049-15050, Appendix A.1 and Table 6
  • Construction, role-playing, and CHARMAP prompts: Published pages 15056-15058, Tables 10-14
  • Manual rubric error: Published page 15059, Table 15
  • Public coverage, option distribution, invalid rows, scenario leak, and runner limits: Official LifeChoice repository commit 37266c38648cd56574b0cb2140857171a752cf38 independently profiled 15 July 2026
  • Complete artifact and validity audit: reports/verification/article-191-lifechoice-leakage-and-artifact-audit.json
  • Complete visual inspection: All 22 published PDF pages rendered and visually inspected on 15 July 2026