PaperlayerView exact revision on arXiv

Source-linked AI summary

Beyond Profile: From Surface-Level Facts to Deep Persona Simulation in LLMs

Zixiao Wang, Duzhen Zhang, Ishita Agrawal, Shen Gao, Le Song, Xiuying Chen

arXiv:2502.12988v3cs.CL

TL;DR

Existing persona-simulation methods often capture biographical facts or limited dialogue rather than deeper worldviews and ideological patterns. CharacterBot uses Lu Xun’s writings, four training tasks, and CharLoRA to model linguistic and thought patterns; experiments report consistent improvements over baselines, while the authors caution that the simulated depth reflects textual pattern learning rather than genuine cognition.

  • Problem

    Existing persona-simulation approaches often emphasize surface profiles or narrow conversational patterns, leaving deeper worldviews, beliefs, and ideological dimensions insufficiently represented.

    Source: §1 · p. 2

  • Method

    CharacterBot trains on Lu Xun’s corpus using pre-training plus multiple-choice QA, generative QA, and style transfer, with CharLoRA integrating shared and task-specific persona knowledge.

    Source: §1 · p. 2, §4.1 · p. 4

  • Results

    CharacterBot outperforms baselines across linguistic accuracy, style alignment, and opinion comprehension, including 0.880 multiple-choice accuracy versus 0.788 for Tongyi Xingchen and 0.787 for Qwen2.5-7B.

    Source: §1 · p. 2, §5.5 · p. 7, §7 · p. 9

  • Takeaways & Limitations

    The work supports modeling deep character personas from textual works through multitask learning and shared task-specific adaptation.

    Source: §1 · p. 2, §7 · p. 9

    The simulated persona reflects statistical patterns in Lu Xun’s texts rather than genuine comprehension, consciousness, or human-like abstract thought.

    Source: Limitations · p. 9

Abstract

from arXiv · show

Previous approaches to persona simulation large language models (LLMs) have typically relied on learning basic biographical information, or using limited role-play dialogue datasets to capture a character's responses. However, a holistic representation of an individual goes beyond surface-level facts or conversations to deeper thoughts and thinking. In this work, we introduce CharacterBot, a model designed to replicate both the linguistic patterns and distinctive thought patterns as manifested in the textual works of a character. Using Lu Xun, a renowned Chinese writer as a case study, we propose four training tasks derived from his 17 essay collections. These include a pre-training task focused on mastering external linguistic structures and knowledge, as well as three fine-tuning tasks: multiple-choice question answering, generative question answering, and style transfer, each aligning the LLM with Lu Xun's internal ideation and writing style. To optimize learning across these tasks, we introduce a CharLoRA parameter updating mechanism, where a general linguistic style expert collaborates with other task-specific experts to better study both the language style and the understanding of deeper thoughts. We evaluate CharacterBot on three tasks for linguistic accuracy and opinion comprehension, demonstrating that it significantly outperforms the baselines on our adapted metrics. We hope this work inspires future research on deep character persona simulation LLMs while considering the importance of ethical standards.

1 Introduction

Character persona simulation should move beyond biographical facts and limited dialogue to represent deeper worldviews, beliefs, and stylistic patterns. CharacterBot addresses this goal through Lu Xun’s writings, multitask training, and CharLoRA, with reported gains over baselines and explicit ethical safeguards.

  • Current persona-simulation methods often rely on profile descriptors, dialogue data, or prompt injection, whereas robust simulation should capture worldviews, ethical frameworks, context-dependent viewpoints, and foundational beliefs.Source: §1 · pp. 1–2
  • CharacterBot uses 17 Lu Xun essay collections containing 638 works to learn linguistic style, cognitive frameworks, ideological viewpoints, and authorial style.Source: §1 · p. 2
  • CharLoRA combines shared linguistic and cognitive knowledge with task-specific persona patterns, enabling cross-task synthesis and more consistent deep persona representation.Source: §1 · p. 2
  • CharacterBot significantly improves style and ideological depth over role-playing baselines, with human evaluations confirming more contextually accurate and stylistically aligned outputs.Source: §1 · p. 2
  • High-fidelity persona simulation requires respectful representation, avoidance of harmful misrepresentation, and clear distinction from authentic works, especially for historical figures.Source: §1 · p. 2
  • Experiments report gains in linguistic accuracy, style preservation, and opinion comprehension while simulating complex character personas.Source: §1 · p. 2

2 Related Work

Related work spans prompt-based simulation, profile- and dialogue-based fine-tuning, style transfer, and parameter-efficient adaptation. These approaches improve persona or stylistic control but generally leave deeper ideological traits insufficiently modeled.

Character Persona Simulation

  • Prompt-based methods inject character settings into prompts, but they often fail to capture intricate traits and dynamic behaviors.Source: §2 · pp. 2–3
  • Fine-tuning methods use profiles, knowledge bases, or novel summaries, yet commonly reproduce surface style while missing deeper character traits.Source: §2 · p. 3

Style Transfer

  • Text style-transfer methods alter stylistic attributes while preserving meaning, but predominantly address linguistic style rather than deeper ideological dimensions.Source: §2 · p. 3

Parameter-Efficient Fine-Tuning

  • Parameter-efficient methods include Prefix Tuning, Prompt Tuning, LoRA, and QLoRA, with LoRA using low-rank matrices for efficient adaptation.Source: §2 · p. 3
  • Recent LoRA variants retrieve, share, or merge modules for mixed-task, personalized, and multitask adaptation, motivating the paper’s deep-persona extension of LoRA.Source: §2 · p. 3

3 Problem Formulation

The paper defines deep persona simulation as generating character-specific responses and stylistic outputs aligned with inferred cognitive and linguistic patterns from a textual corpus. CharacterBot evaluates this capability through multiple-choice QA, generative QA, and style transfer.

  • Deep persona simulation targets consistent style, tone, recurring themes, and inferred cognitive patterns rather than an ontological replication of an inner self.Source: §3 · p. 3
  • A single model performs multiple tasks that jointly assess understanding and generation of character-specific responses.Source: §3 · p. 3
  • Multiple-choice QA provides a question and candidate answers, and the model selects the answer corresponding to the character’s perspective.Source: §3 · p. 3
  • Generative QA asks the model to produce an answer reflecting the character’s knowledge, opinions, or personality.Source: §3 · p. 3
  • Style transfer rewrites an input text into an output aligned with a specified persona style.Source: §3 · p. 3

4 Method

CharacterBot’s method pre-trains on author-perspective-reframed essays, then fine-tunes three complementary tasks and uses CharLoRA to preserve persona knowledge while adapting across tasks.

4.1 Pre-training

  • CharacterBot pre-trains on Lu Xun’s corpus with Authorial Perspective Reframing, which makes authorial attribution explicit and strengthens perspective alignment.APR converts first-person narratives into third-person descriptions with cues such as “the author argues” or “the author critiques.”Source: §4.1 · p. 4

4.2 Fine-tuning

  • Fine-tuning uses multiple-choice answering, generative question answering, and style transfer to align the model with the character’s ideas, ideology, and writing style.The tasks respectively test essay-grounded choice selection, ideological responses faithful to essay arguments, and stylistic rewriting that preserves meaning and intent.Source: §4.2 · pp. 4–5

4.3 CharLoRA

  • CharLoRA adapts a frozen base model through trainable low-rank matrices, integrating persona-specific traits while reducing trainable parameters relative to full fine-tuning.The low-rank update captures writing style, tone, and ideological perspectives while preserving the original model’s general knowledge.Source: §4.3 · p. 5
  • During downstream training, CharLoRA shares Apt for global persona attributes, gives each task its own Bi matrix, and freezes inactive task matrices to prevent interference.This hybrid design preserves a consistent voice while allowing task-specific adaptation for question answering and stylistically consistent generation.Source: §4.3 · pp. 5–6

5 Experiments

Experiments use Lu Xun’s 17 essay collections to train and evaluate CharacterBot across three persona-simulation tasks. CharacterBot achieves the strongest reported results in ideological question answering, generative fidelity, and style transfer.

5.1 Datasets

  • CharacterBot is trained from 17 essay collections and 638 essays, with three derived fine-tuning datasets split into training, validation, and test sets.Source: §5.1 · p. 6

5.2 Baselines

  • Baselines include general-purpose, role-playing, and Lu Xun style-transfer models, while implementation uses Qwen2.5-7B-Instruct with CharLoRA rank 64.Source: §5.2 · p. 6, §5.3 · p. 6

5.4 Evaluation Metrics

  • The evaluation covers accuracy, content and style scores, and BLEU, ROUGE-1, and style matching, with scoring supported by DeepSeek-V3 and two literary experts.Source: §5.4 · p. 7

5.5 Main Results

  • 0.880 accuracy on multiple-choice questions exceeds Tongyi Xingchen (0.788) and Qwen2.5-7B (0.787), with the task testing complex ideological nuances.Source: §5.5 · p. 7
  • 3.214 content and 2.885 style scores in generative question answering surpass GPT-4o’s 2.542 and Tongyi Xingchen’s 2.823 style score.Source: §5.5 · p. 7
  • 0.293 BLEU, 0.410 ROUGE-1, and 0.937 style matching in style transfer outperform baselines, including LuXun-GPT overall.Source: §5.5 · p. 7

6 Analysis and Discussion

Ablation and human-evaluation analyses attribute CharacterBot’s gains to the full CharLoRA and APR design and report stronger content and stylistic fidelity than leading baselines.

6.1 Ablation Study

  • The full architecture achieves the best performance across all three tasks, confirming contributions from both CharLoRA and APR.Source: §6.1 · p. 8
  • VanillaLoRA drops across content, human-evaluated style, automatic style matching, BLEU, and ROUGE-1, while Multi-LoRA improves slightly but remains below CharLoRA.Source: §6.1 · p. 8
  • Without APR, style score falls to 2.536 and style matching to 0.895, linking APR to third-person reframing that decouples narrative perspective from content.Source: §6.1 · p. 8

6.2 Human Evaluation

  • Human evaluation by two literary experts yields Cohen’s kappa of 0.72 and finds CharacterBot superior to GPT-4o and Tongyi Xingchen in content and stylistic fidelity.Source: §6.2 · p. 8
  • Case studies show CharacterBot avoids distractor bias, faulty attribution, causal misinterpretation, and metaphor-induced meaning distortion better than baselines.Source: §6.2 · pp. 8–9, Table 4 · p. 9

7 Conclusion

The paper presents CharacterBot as a multitask Lu Xun persona simulator combining linguistic modeling with deeper thought-process modeling. Its experiments report better linguistic accuracy and opinion comprehension than baselines, with future work targeting broader personas and genres.

  • CharacterBot combines linguistic-structure pre-training, question-answering and style-transfer fine-tuning, and CharLoRA collaboration between general and task-specific experts.Source: §7 · p. 9
  • Experiments show higher linguistic accuracy and opinion comprehension than baselines, while future work extends toward diverse personas and multilayered simulations.Source: §7 · p. 9

Limitations

CharacterBot’s simulated depth depends on statistical pattern reproduction from source texts rather than genuine human-like cognition, and its current evidence base is limited in scope.

  • CharacterBot reproduces recurring arguments, themes, and stylistic choices without possessing genuine comprehension, consciousness, or human-like abstract thought.Its simulated depth depends on source-text richness and pattern-matching capability rather than intrinsic cognitive faculty.Source: Limitations · p. 9
  • The current approach primarily uses personal essays and reflective writings, leaving novels and fictional works as an open challenge for more holistic simulation.Extending coverage would require understanding thematic elements, narrative structures, and symbolic representations.Source: Limitations · p. 9
  • Broader tasks and evaluations are needed to assess and improve persona depth, fidelity, and plausibility while addressing bias and nuanced abstract concepts.Source: Limitations · pp. 9–10

Ethical Considerations

The paper frames deep persona simulation as requiring ethical safeguards around authenticity, intellectual property, privacy, and misuse, while grounding the Lu Xun dataset in public-domain essays.

Ethical Considerations

  • Deep persona simulation raises risks of misrepresenting an author’s intent or historical context, enabling convincing misinformation or forgery, and violating privacy or intellectual-property rights.Source: Ethical Considerations · p. 10
  • The authors require clear AI labeling, restricted educational or controlled creative use, ethical data sourcing and consent, technical safeguards, and continuing ethical review.Source: Ethical Considerations · p. 10
  • The ethical stance emphasizes respect for the original author, transparency of AI generation, and use within appropriate boundaries.Source: Ethical Considerations · p. 10

A Essay Lists

  • The dataset contains 638 Lu Xun articles from 17 essay collections sourced from Wikisource, with all texts in the public domain for scholarly use.The compilation supports analysis of linguistic patterns, thematic evolution, and rhetorical strategies across his essays.Source: A Essay Lists · p. 12

B Data Examples

The data examples operationalize persona understanding through multiple-choice judgment, open-ended viewpoint generation, and style transfer that preserves meaning while adapting language.

B Data Examples

  • The multiple-choice task tests whether the model selects answers from candidate options according to the character’s semantic perspective.Source: B Data Examples · p. 12, Table 6 · p. 13
  • The generative question-answering task asks the model to articulate viewpoints in response to open-ended questions, testing reasoning and language expression.Source: B Data Examples · p. 12, Table 7 · p. 13
  • The style-transfer task requires reconstructing Lu Xun’s language style while preserving the original meaning.Source: B Data Examples · p. 13, Table 8 · p. 13
  • Together, the tasks evaluate closed choice, open generation, and style migration as a multidimensional benchmark for language understanding and generation.Source: B Data Examples · p. 13

C Case Study

The case studies identify recurring baseline errors and show CharacterBot preserving contextual interpretation, semantic integrity, and rhetorical intent; the accompanying prompts specify data construction and evaluation procedures.

C Case Study

  • Baseline errors include distractor bias, faulty attribution, context neglect, misaligned metaphor, and concept drift.Source: C Case Study · p. 13
  • The case studies compare CharacterBot with GPT-4o, Llama3.1-8B, and Tongyi Xingchen in Chinese, with red marking errors and blue marking correct responses.Source: C Case Study · p. 13, Table 9 · p. 15, Table 10 · p. 16
  • In generative question answering, CharacterBot correctly attributes Lu Sao’s punishment to the influence of his ideas, unlike a baseline that misidentifies the context and invokes Rousseau.Source: C Case Study · p. 13
  • In style transfer, CharacterBot preserves semantic integrity and rhetorical intent where GPT-4o shifts “eliminating disasters” toward “changing our fate.”Source: C Case Study · pp. 13–14

D Prompts

Related papers

Loading 2502.12988v3…