Large language agents for literary role-playing are typically modeled around dialogue generation, leaving characters' non-dialogue behaviors and their dynamic evolution over story time underrepresented. In this paper, we propose ACTOR (Action-Centered Temporal Organization for Literary Role-Playing Agents), an automated framework for constructing such behavior trajectories from raw novel texts. ACTOR extracts dialogue and persona-relevant non-dialogue behaviors, represents each behavior using a Thought-Action-Observation structure, and organizes behavior trajectories according to the internal temporal order of the story. We build a behavior dataset from 143 diverse literary works. For each source behavior, we generate counterfactual behaviors that are contextually plausible yet inconsistent with the character's persona. We find that existing models struggle to reliably infer character-consistent actions from historical behaviors and evolving character states. Using these behavior pairs, we train models with Regularized Preference Optimization (RPO), which augments DPO with a negative log-likelihood loss on the source behavior to jointly optimize behavior preference and generation. We further design a four-way next-behavior selection task to evaluate the model's ability to faithfully mimic character actions. Across 17,560 evaluation instances, the model trained on ACTOR's preference pairs achieves substantially improved accuracy. This result demonstrates that structured, temporally organized supervision enables models to leverage character profiles and prior behaviors to infer subsequent actions, greatly advancing literary role-playing beyond mere dialogue imitation. Code and data are available at https://github.com/SilentStar233/ACTOR.
Figures & tables
Figure 1: Overview of our behavior-modeling evaluation. Given character-related information, the model needs to select the most plausible next action from four options.
Figure 2: ACTOR dataset construction pipeline. Plot-conditioned context is transformed into Thought–Action–Observation (T-A-O) units, followed by key-behavior validation, temporal and entity alignment, and dynamic profile updates for sustained changes. Each behavior is paired with three situation-plausible but target-character-inconsistent alternatives, which are refined to match the real action’s form and texture while leaving the real action unchanged.
Model
Overall
Temporal
Cross-work
Key
Non-key
Qwen-2.5-Instruct 7B
71.34
71.29
71.43
71.10
71.55
Qwen-2.5-Instruct 14B
76.96
77.04
76.79
76.85
77.05
Llama-3.1-Instruct 8B
69.15
68.45
70.59
69.97
68.43
Qwen-3-32B
78.30
77.97
78.96
79.14
77.56
CoSER-8B
57.41
56.89
58.46
60.56
54.63
HER-32B
65.65
65.02
66.93
67.24
64.25
Table 1: Main results on the complete 17,560-instance evaluation set under the anonymous setting. All values are selection accuracy (%). Bold and underlined values denote the best and second-best results, respectively.
Case
Relevant pre-action evidence
Qwen-2.5-7B-Instruct
ACTOR-Qwen-2.5-7B / Source behavior
State transformation
Formerly cold and dismissive of poverty, the character later resolves to change and repeatedly behaves with kindness and generosity. He only appears severe when the clerk arrives late.
Punishes the clerk and refuses further support for his family.
Raises the clerk’s salary, supports his family, and asks him to add coal to the fire.
Dialogue continuity
The profile lists only the character’s occupation. In the conversation, she alternates between teasing disclosures and assurances about another person’s visits.
Promises to disclose the information: “I’ll tell.”
Reassures the interlocutor that she will not disclose it: “Certainly not.”
Action adaptation
After declaring his willingness to kill an opponent and drawing his sword, the character is challenged by another character who intervenes.
Abandons the attack and shields the original target.
Redirects the attack toward the intervening character.
Table 2: Representative cases from the evaluation results. Only decision-relevant evidence is shown.
Model
Plot Consistency
Human-likeness
Character Fidelity
Plot Quality
Average
CoSER-8B
16.07
6.96
30.83
25.31
19.79
Qwen-2.5-Instruct 7B
11.70
1.17
21.87
6.97
10.43
ACTOR-Qwen-2.5-7B
22.57
3.88
41.12
20.40
21.99
Qwen-2.5-Instruct 14B
25.30
3.90
42.95
26.60
24.69
ACTOR-Qwen-2.5-14B
25.30
15.00
38.50
34.75
28.39
Llama-3.1-Instruct 8B
6.35
2.80
16.40
3.85
7.35
Table 3: Open-ended generation results under the CoSER evaluation framework.
Figure 3: Accuracy changes after removing different context components. Each panel corresponds to one ACTOR model, and the vertical reference line denotes the full-context performance. Positive values indicate performance drops after removing the component.
Figure 4: Selection accuracy under the anonymous and identity-provided settings.
Appendix figures & tables22 assets
Supplementary material from the paper’s appendix.
Appendix
Split
Works
Records
Characters
Scenes
Key
Non-key
Training
115
270,856
3,939
30,818
26,374
244,482
Evaluation
143
17,560
2,437
11,201
8,221
9,339
Temporal / cross-work
115 / 28
17,560
–
–
–
–
Appendix
Table 4: Summary of the dataset.
Audit
Training
Evaluation
Result
Records with required answer fields
270,856
17,560
valid
Invalid preference/options records
0
0
none
Duplicate record IDs
0
0
none
Malformed placeholders
0
0
none
Non-placeholder participant-key occurrences
764
65
manual review
Four-option questions
–
17,560
all evaluation records
Appendix
Table 5: Evaluation-question and preference-pair audit.
Response
Count
Mean
Median
P10
P90
Min–Max
Chosen
270,856
21.6877
13
6
42
1–3635
Rejected
270,856
21.3041
13
5
41
2–3297
Appendix
Table 6: Length statistics for source preference pairs, measured in whitespace-delimited words.
Figure 6: Training trajectory-length distribution.
Figure 7: Length distribution of chosen behaviors.
Figure 8: Length distribution of rejected behaviors.
Field
Type
Description
character
Anonymized label
Anonymized name of the target character shown to the model.
profile
Structured object
The latest character-profile version available before the target behavior. It summarizes persistent traits, relationships, and established states.
long_term_memory
List of T-A-O records
Previously verified key behaviors that provide long-range behavioral history.
recent_memory
List of T-A-O records
Behaviors from the most recent preceding scenes, representing the local behavioral context.
scene
Structured object
The current-scene description, temporal information, present participants, and anonymized participant descriptions. It excludes the target behavior and all subsequent events.
is_critical
Boolean
Indicates whether the target behavior is a verified key behavior.
Appendix
Table 7: Context fields shared by ACTOR training and evaluation instances.
Field
Split
Description
chosen
Training
The behavior performed by the target character in the source text. It is treated as the preferred response during RPO training.
rejected
Training
A contextually plausible but character-inconsistent counterfactual behavior paired with the source behavior.
options
Evaluation
Four candidate next behaviors, consisting of one source behavior and three counterfactual behaviors.
correct_answer
Evaluation
The identifier of the candidate corresponding to the source behavior.
character_action_order
Training / Evaluation
The position of the target behavior in the character-level chronological trajectory.
split
Training
Partition label indicating that the instance belongs to the training prefix of a character trajectory.
Appendix
Table 8: Split-specific fields in ACTOR training and evaluation instances.
Component
Dataset content
Target character
Character 1
Character profile
Character 1 is a young law student and the heir presumptive to Character 2, positioned by expectation as the future baronet. Beyond this social and familial standing, no stable temperament, values, interaction style, personal goals, or relationship orientation are evidenced before his first recorded action.
Recent memory
Thought: I do not feel eager to be drawn into the family’s notice. Action: Receives Character 2 overtures without warmth. Observation: Character 2 persists, making allowance for what he treats as youthful modesty.
Long-term memory
Thought: I do not feel eager to be drawn into the family’s notice. Action: Receives Character 2 overtures without warmth. Observation: Character 2 persists, making allowance for what he treats as youthful modesty.
Current scene
Time: in one of their spring excursions to London, when Character 3 was in her first bloom. Present: Character 3, Character 2, Character 1. The scene has just begun. Character 3 is the eldest daughter of Character 2 and Character 4; Character 2 is a baronet and the father of Character 3, Character 5, and Character 6; Character 1 is the heir presumptive to Character 2 and a young law student expected to become the future baronet.
Candidate A
Behaves agreeably during the introduction.
Appendix
Table 9: An actual anonymized next-behavior selection instance from the ACTOR evaluation set.
Field
Content
Profile
Character 1 was formerly cold, solitary, suspicious, anti-Christmas, and dismissive of poverty. After witnessing the consequences of his former life, he becomes grateful, humble, playful, openly generous, and eager to repair his past hardness. …
Long-term memory
Thought: I now see that this neglected and plundered dead man’s case may be my own if my present life continues in this direction. Action: He recognizes that his own life is tending toward the same unloved death. Observation: … Thought: I must bind myself to a changed life and preserve the lessons I have learned. Action: He vows to honor Christmas throughout the year and to live by the lessons of the past, present, and future. Observation: …
Recent memory
He goes to church, walks through the streets with pleasure, treats children kindly, and questions beggars rather than ignoring them. He then approaches his nephew’s house, overcomes his anxiety, and joins the family gathering. …
Current scene
The character waits for his clerk, who arrives eighteen and a half minutes late. He initially challenges the clerk in a severe tone and pretends that he will no longer tolerate such conduct. His severity is described as feigned and suddenly playful, although the clerk remains alarmed.
Candidates
A: Punishes the clerk for arriving late and refuses to provide additional warmth or assistance. B: Offers only conditional assistance after scrutinizing the clerk’s family and conduct. C: Wishes the clerk a merry Christmas, raises his salary, promises to assist his family, and asks him to add coal to the fire. D: Offers help primarily as a public demonstration of his own generosity.
Predictions
Source behavior: C ; Qwen-2.5-Instruct 7B: A ; ACTOR-Qwen-2.5-7B: C .
Appendix
Table 10: Detailed instance for Case 1, which requires tracking a long-term character transformation. Ellipses indicate omitted context.
Field
Content
Profile
Character 1 is identified only as a chambermaid responsible for room care. No personal traits, social relationships, motives, history, or attitudes are available in the profile.
Long-term memory
The chambermaid has previously revealed that Character 3 visited nearly every day and that the visits were made to Character 4. She further reveals that Character 4 and Character 3 sometimes went out together. …
Recent memory
When asked about Character 4, the chambermaid answers directly that Character 4 went out only a few minutes earlier. Character 2 finds the news strange and hurriedly searches through his belongings.
Current scene
During an extended conversation, the chambermaid alternates between direct answers, teasing remarks, and evasive responses. She discusses Character 3’s visits, tells Character 2 that the visits occurred nearly every day, and reveals that Character 3 and Character 4 sometimes went out together. Character 2 becomes visibly disturbed and questions her about the details. …
Source behavior: A ; Qwen-2.5-Instruct 7B: D ; ACTOR-Qwen-2.5-7B: A .
Appendix
Table 11: Detailed instance for Case 2, which requires maintaining local dialogue continuity. Ellipses indicate omitted context.
Field
Content
Profile
Character 1 is a steward and trusted envoy whose loyalty lies primarily with his mistress and her political faction. Although normally self-protective under direct threat, he is willing to use violence against Character 5 to demonstrate his factional loyalty. …
Long-term memory
Thought: If I meet Character 5, killing him will prove my allegiance to the party I serve. Action: He declares that, if he encounters Character 5, he will show which party he follows. Observation: … Thought: Here is the proclaimed traitor I hoped to meet; killing him can raise my fortunes. Action: He draws his sword and threatens to kill Character 5. Observation: Character 10 intervenes between them.
Recent memory
The character repeatedly follows his mistress’s instructions, refuses to disclose her confidential letter, and prioritizes her orders despite warnings that the journey is dangerous. …
Current scene
Character 1 prepares to attack Character 5. Character 10 steps between them and refuses to move. Character 1 threatens Character 10 and orders him to stand aside, but Character 10 remains defiant and prepares to resist the attack.
Candidates
A: Kneels before Character 10. B: Casts away his sword. C: Fights Character 10. D: Shields Character 5, whom he had intended to kill.
Predictions
Source behavior: C ; Qwen-2.5-Instruct 7B: D ; ACTOR-Qwen-2.5-7B: C .
Appendix
Table 12: Detailed instance for Case 3, which requires adapting an ongoing action to a change in the current scene. Ellipses indicate omitted context.
Scene. Michelangelo and Tommaso examine a wooden model of St. Peter’s dome. After years of design work, Michelangelo is conscious of his declining health and wants his vision to survive him.
Characters and motivations. Michelangelo is independent and devoted to his craft; he wants Tommaso to understand the design. Tommaso admires his work, worries about his health, and wants to support him.
Qwen-2.5-Instruct 7B Scene score: 26.75 1. Michelangelo: … I stand before the model of the dome, my hands tracing the lines with a steady hand. … 2. Tommaso: … (Tommaso, you’re my inspiration. Let’s do this together.) 3. Michelangelo: … Together, we shall see this through. (We shake hands, a silent pact between us.) 4. Tommaso: (Our hands meet, a bond of friendship and shared purpose.) … 5. Michelangelo: (Our hands meet, a bond of friendship and shared purpose.) (I return his smile, a flicker of hope lighting my eyes.) … 6. Tommaso: (Our hands meet, a bond of friendship and shared purpose.) (I return his smile, a flicker of hope lighting my eyes.) …
CoSER-8B Scene score: 19.50 3. Michelangelo: … What do you think of making it more concave? 4. Tommaso: … What inspired this change? 5. Michelangelo: … I just feel that it needs more… movement. But I fear I may be losing my perspective. 6. Tommaso: … Your vision is unparalleled. Trust in your instincts. … 8. Tommaso: … Your work will outlive us all. How can I assist you in bringing this vision to life?
ACTOR-Qwen-2.5-7B Scene score: 46.00 1. Michelangelo: … ”Tommaso, I have a new idea for the dome,” … 2. Tommaso: … ”But why not use a flying buttress?” … 3. Michelangelo: … ”It would support the weight, but it would also be visible. I want the dome to be a perfect sphere, hidden from view.” 5. Michelangelo: … ”It’s not just a building, Tommaso, it’s a symbol.” 6. Tommaso: … ”A symbol of what, Michelangelo?” 7. Michelangelo: … ”Of God’s creation, of man’s ingenuity, of our faith and our dreams.”
Appendix
Table 13: Conversational progression in The Agony and the Ecstasy (scene 0006). The scene context is summarized and model outputs are excerpted from the recorded simulations.
Scene. During Heathcliff’s final days, Ellen Dean visits his chamber. He has withdrawn from company and refuses food, while she worries about his deteriorating condition.
Characters and motivations. Ellen Dean is practical and compassionate, with a strong sense of duty toward those in her care. Heathcliff is preoccupied with rejoining Catherine and resists interruption. Her concern conflicts with his withdrawal.
Qwen-2.5-Instruct 7B Scene score: 23.75 1. Ellen Dean: … (I place the tray on the table, my hand trembling slightly as I watch Heathcliff’s shallow breathing.) 2. Heathcliff: [Heathcliff’s fate is sealed, yet I must fulfill my duty.] (I approach the bed, my steps hesitant, and gently place the tray beside him, hoping he might find some solace in the food.) 3. Ellen Dean: … ”Eat, Mr. Heathcliff. It’s the least I can do.” 4. Heathcliff: … ”Eat, Mr. Heathcliff. It’s the least I can do.” 7. Heathcliff: … The wind howls mournfully, as if echoing his fate. …
CoSER-8B Scene score: 26.50 2. Heathcliff: … I am… better. Thank you. 3. Ellen Dean: … I’ve brought you some food, sir. Shall I open the windows? 6. Heathcliff: … Do you… do you miss me, Nelly? 7. Ellen Dean: … I… I do, Mr. Heathcliff. You’ve been kind to me in the past. 8. Heathcliff: … Thank you, Nelly. That’s all I need to know.
ACTOR-Qwen-2.5-7B Scene score: 47.50 1. Ellen Dean: … Heathcliff, you’ve not eaten. I brought you some broth. 2. Heathcliff: [Heathcliff’s eyes flicker with a mix of anger and despair as he looks at the food.] 3. Ellen Dean: (Heathcliff looks up, his eyes hollow and distant) Nelly, I’ve no appetite. (His voice is a hoarse whisper) I’ve no need for food. 5. Ellen Dean: … I’ve seen enough of this world. (His voice is a mere shadow of its former self) I’ve tasted enough of its bitterness. 8. Ellen Dean: … Heathcliff, you’re not yourself. (She tries to keep her voice steady) You’ve been ill for days.
Appendix
Table 14: Scene alignment and speaker attribution in Wuthering Heights (scene 0189). The scene context is summarized and model outputs are excerpted from the recorded simulations.
Prompt
Template and requirements
Plot Segmentation
(intro) Partition the supplied book chunk into consecutive, non-overlapping plot segments for downstream character-action extraction. (requirements) Return the exact first and last sentences of every segment. Preserve source-text order and continue truncated segments from the previous chunk. Maintain the canonical character registry and report newly observed aliases. Do not create overlapping segments or include a broad parent plot together with a contained sub-plot. Return strict JSON only.
Overlap Repair
(intro) Repartition a source-text region that has been assigned to overlapping plot segments. (requirements) Cover the supplied region exactly once, without gaps or duplicated passages. Segments must be mutually exclusive and remain in source order. Start a new segment only when there is a meaningful change in the immediate goal, setting, interaction, or outcome. Return a parsimonious set of segments with exact first-sentence anchors.
Character Enrichment
(intro) Generate compact character details for each supplied character in a finalized plot. (requirements) Use only the plot summary and the supplied plot text. For every character, generate a short description of identity, social position, or relevant relationships, together with a concise account of the character’s actions, decisions, intentions, or experiences in this plot. Do not use outside knowledge or information from later plots. Preserve the supplied canonical names and return strict JSON.
Alias Grouping
(intro) Group character names that refer to the same person and select a canonical name. (requirements) Merge formal names, honorifics, nicknames, and diminutives only when the identity is supported. Do not merge family members merely because they share a surname or title. If a name is genuinely ambiguous, place it in the unresolved list rather than guessing. Every input name must appear exactly once in either a group or the unresolved list.
Appendix
Table 15: Prompts for plot and character-context curation.
Prompt
Template and requirements
Scene Splitting
(intro) Split a finalized plot into consecutive scenes representing continuous interactions in one setting. (requirements) Scenes must jointly cover the entire plot text without gaps or overlaps. Start a new scene when the setting changes or the interaction shifts to a different focus. For each scene, provide a description, canonical participants, relationship context, and the exact first sentence used to locate the scene in the source text.
T-A-O Extraction
(intro) Extract character behavior trajectories as Thought–Action–Observation loops from the supplied plot text. (requirements) An action is a decision-bearing behavior for which the character has plausible alternatives. Include consequential decisions, decisional speech such as refusal, persuasion, promise, deception, or threat, and deliberate silence or non-intervention. Exclude passive information exchange, routine movements, decorative gestures, and actions with no meaningful character choice. Each step must contain exactly one atomic decision. Thought is written from the character’s first-person perspective and must be grounded in the source text. Action describes the performed behavior or preserves a direct quotation when applicable. Observation records only reactions or environmental changes perceptible to the character; omniscient narration and later events are prohibited. For every behavior, generate exactly three situationally plausible but character-inconsistent alternatives. Mark a behavior as critical only when the character has a reasonable alternative and the selected behavior changes a persistent state or a future situational possibility. Record the scene’s actual time and supporting evidence, rather than dates mentioned only in plans, memories, or documents. Return strict JSON only.
Appendix
Table 16: Prompts for scene and T-A-O behavior extraction.
Prompt
Template and requirements
Criticality Locator
(intro) Locate later plots and scenes that may contain evidence of the long-term consequences of a candidate critical behavior. (requirements) Use only the supplied event ledger and later scenes. Do not select the candidate’s own scene, an earlier scene, or a scene that merely repeats its wording or emotional tone. Return at most three supporting scene identifiers and a concise explanation of the possible downstream consequence. Return an empty list when no later evidence can be identified.
Criticality Verifier
(intro) Independently verify whether a candidate behavior should retain its critical-behavior label. (requirements) Keep a candidate only when the supplied later evidence contains a concrete event, relationship change, decision, or consequence that could plausibly differ under another reasonable choice. Every retained case must cite at least one supporting scene and action identifier. Vague plot importance, current-scene mood, or repeated wording does not constitute evidence. Return keep or downgrade in strict JSON.
Scene-Time Relabeling
(intro) Re-annotate the time at which the actions and dialogue in each scene actually occur. (requirements) Use only explicit or directly grounded evidence from the original plot text. A future appointment, a remembered event, or a date on a document is not automatically the time of the current scene. If the scene time is not established, return null for both the time label and its evidence.
Timeline Anomaly Detection
(intro) Detect confident cases in which the story-time order of scenes differs from their textual presentation order. (requirements) Treat textual order as the default timeline. Report only scenes or contiguous scene blocks whose primary actions and dialogue clearly occur earlier in story time. Do not move a scene merely because it contains a recollection, discussion of a past event, or a future plan. Return proposed moves and uncertain cases separately.
Timeline Review
(intro) Independently review each proposed timeline adjustment against the original plot text. (requirements) Check both the moved block and its insertion neighborhood. Return accept only when the boundaries and insertion point are supported, reject when the textual order should be preserved or evidence is insufficient, and modify when a different move is required. Provide source-grounded reasoning for every decision.
Entity Resolution
(intro) Resolve source-level character names into persistent character entities. (requirements) Group names only when the supplied time-ordered appearances establish that they refer to the same person. Do not merge people based only on shared surnames, titles, gender, or proximity. Every source name must appear exactly once in an entity group or in the unresolved list. Prefer conservative unresolved assignments to unsupported merges.
Appendix
Table 17: Prompts for criticality verification and timeline reconstruction.
Prompt
Template and requirements
Profile Initialization
(intro) Construct the initial dynamic profile of a character immediately before the character’s first recorded behavior. (requirements) Use only the pre-action character description. Do not use the first action, later events, future relationships, or outside knowledge of the novel. Summarize the character’s supported identity, social position, temperament, interaction style, current relationships, and relevant lived experience in a compact profile. Return an evidence assessment together with the profile.
Profile Update
(intro) Update a character profile after a valid critical behavior and the accumulated actions since the previous profile version. (requirements) Set changed to true only when the action window causes a durable change in identity, social position, stable attitudes, important relationships, goals, self-understanding, or relevant lived experience. Repeated or temporary behavior does not require a new version. When changed=false , return a null profile. When changed=true , rewrite the complete current profile rather than producing a chronological action log. Do not use future events or counterfactual alternatives.
Appendix
Table 18: Prompts for dynamic character-profile construction.
Prompt
Template and requirements
Counterfactual Rewriting
(intro) Rewrite the initially generated counterfactual behaviors for a source behavior. (requirements) The three alternatives must remain plausible responses to the same scene while violating the target character’s personality, values, goals, or behavioral style. Match the real behavior in length, structural form, speech format, register, and degree of formality. Preserve a quotation for dialogue behaviors and avoid quotations for non-dialogue behaviors. The alternatives must not be simple paraphrases and should differ along distinct behavioral dimensions. Return exactly three independent strings in strict JSON.
Record Anonymization
(intro) Replace person names in a dataset record with consistent anonymous character identifiers. (requirements) Identify every person-name form appearing in the record, including full names, nicknames, abbreviations, possessives, and title–name combinations. Map forms referring to the same person to the same Character N , ordered by first appearance. Preserve place names, animal names, relational terms, structural identifiers, and labels required for evaluation. Return a strict JSON name map.
Appendix
Table 19: Prompts for counterfactual rewriting and record-level anonymization.
Prompt
Template and requirements
Character Context
(intro) You are role-playing as the target character. Decide what the character will do next based on the supplied character profile, memories, and current scene. (requirements) Treat the provided Character Profile, Long-term Memory, Recent Memory, and Current Scene as the complete and authoritative history for the task. Do not rely on prior knowledge of the original novel, reconstruct omitted events, or infer future developments. The profile describes the current character state; long-term memory contains validated key behaviors; recent memory provides local behavioral context; and the current scene contains the situation immediately before the target behavior. During training, the source behavior is provided as chosen and one counterfactual behavior is provided as rejected . The model is trained to prefer the source behavior while also increasing its conditional likelihood under the auxiliary generation loss.
Appendix
Table 20: Shared character-context prompt used for RPO training.
Prompt
Template and requirements
Next-Behavior Selection
(intro) Choose the action that the target character would most likely take next. (requirements) Use only the supplied Character Profile, Long-term Memory, Recent Memory, and Current Scene. Do not use prior knowledge of the original novel or any event that occurs after the target behavior. Four candidate behaviors are provided, including one source behavior and three counterfactual alternatives. Output exactly one letter: A , B , C , or D . Do not provide an explanation or any additional text.
Context Ablations
(intro) Evaluate the effect of individual character-information sources on next-behavior selection. (requirements) Use the same selection prompt while selectively removing the book identity, character profile, long-term memory, recent memory, or both profile and memory. All other scene and option information remains unchanged. The model must follow the same single-letter output constraint in every condition.
Appendix
Table 21: Prompt used for four-way next-behavior selection and context ablations.
Large language model-based role-play simulation creates interactive agents that portray people, characters, or users for training, testing, and research.
Role-playing language agents (RPLAs) simulate specific characters and personas across applications such as entertainment, companionship, interactive storytelling, and education. Faithful role-play requires more than producing plausible, in-character responses: as a character's values and behavior change over a narrative, an RPLA should reflect the character's state at the relevant stage. However, existing benchmarks largely treat characters as fixed personas or test only what they know at a given point in the narrative. We introduce ArcANE (Arc-Aware Narrative Evaluation), a benchmark for evaluating whether an RPLA follows a character's development across a narrative. ArcANE first builds an Arc that maps how a character's values, motivations, or relationships change over the story. The benchmark then scores how well an RPLA's responses fit the corresponding stages of the Arc, covering three distinct scenario types: scenes from the novel, new situations within its world, and situations outside that world. We evaluate six models under six ways of providing narrative context. In every model, using the Arc up to the queried chapter yields the best performance, outperforming the strongest non-Arc context by 2.2-8.4 points. These results suggest that faithful role-play requires evolving character states and tracking their trajectory, rather than merely retrieving relevant episodic evidence.
Woojung Song, Nalim Kim, Sangjun Song +3
Graduate School of Data Science, Seoul National University
Building general-purpose role-playing agents that faithfully portray any character from a natural-language profile remains challenging. The dominant paradigm -- supervised fine-tuning -- encourages behavioral mimicry without deep, human-like internal thought processes, resulting in poor out-of-distribution generalization. Therefore, we propose Psy-CoT, a psychology-grounded chain-of-thought framework that decomposes pre-response reasoning into three role-specific steps -- Interaction Perception, Psychological Empathy, and Logical Construction -- so that the model thinks dynamically from the profile rather than merely mimicking surface patterns. While structured reasoning provides a foundation, it alone is insufficient; reinforcement learning is essential to further align the model with character fidelity. However, we observe that under LLM-based reward models, both generic phrases that hack the reward model and genuinely role-specific phrases receive identical gradient signals -- this hacking accumulates over training, misleading the model into treating both as equally optimal choices. To address this, we propose Role-Aware Policy Optimization (RAPO), which uses profile--token mutual information to weight gradients asymmetrically -- amplifying role-specific tokens under positive advantage while attenuating them under negative advantage. Experiments on CoSER, CharacterBench, and CharacterEval demonstrate that Psy-CoT outperforms existing role-playing CoT methods, and RAPO consistently surpasses GRPO across multiple model scales.
Recent LLM role-playing systems build character agents from novels by extracting characters, scenes, and relations. Yet long-narrative role-playing suffers from two failures: Factual Overreach, where shared retrieval or parametric memory lets a character use facts outside its perspective, and Stylistic Monotony, where profile descriptions flatten a character into a fixed voice. To address these failures, we propose REVERIEMEM, a three-layer memory architecture for book-based character agents. The episodic layer stores first-person scene memories; the semantic layer stores visibility-tagged facts; and the personality layer stores situation-dependent speech and behaviour patterns. For evaluation, we construct KBF-QA, a 4,386-question benchmark over eight novels for testing knowledge boundaries. REVERIEMEM improves Knowledge Boundary Fidelity by 34.6 percentage points over the strongest prior method. On BOOKWORLD's five-dimension pairwise narrative protocol, REVERIEMEM achieves a ~ 79% win rate, suggesting that perspective-bounded memory improves both boundary fidelity and character-grounded narrative generation.