Prompt-Based Role Playing

• Prompt-Based Role Playing is the practice of conditioning large language models with natural language role specifications to emulate specific personas, professions, or characters.
• PBRP employs diverse techniques—from simple role cards to memory injection and multi-agent pipelines—to steer style, tool use, and decision-making in tailored domains.
• Empirical evaluations reveal domain-dependent gains and limitations, highlighting challenges in achieving deep cognitive role shifts versus surface-level behavior imitation.

Prompt-Based Role Playing (PBRP) is the practice of conditioning a LLM with natural-language role specifications so that it responds as a particular persona, profession, or character. In the literature, this conditioning can be as simple as a role card in the system prompt, or as elaborate as memory injection, retrieval-augmented demonstrations, multi-agent prompt pipelines, and explicit reasoning scaffolds. Across recent work, PBRP is used to induce persona fidelity, stylistic consistency, domain-specific behavior, tool-use policies, and in some settings graded expertise; however, the evidence also shows that prompting alone often stabilizes surface behavior more reliably than it changes deep internal computation (Wang et al., 15 Jan 2026, Liang et al., 28 Oct 2025).

1. Conceptual foundations and historical positioning

PBRP occupies a specific place in the evolution of role-playing language agents. A recent survey organizes that evolution into three broad stages: rule/template-based role simulation, style imitation relying on pre-trained LLMs and prompt engineering, and cognitive simulation centered on personality-driven and decision-consistent modeling with memory mechanisms. Within that taxonomy, PBRP lies primarily in the style-imitation stage, while memory-augmented prompting and motivation-situation control begin to push it toward cognitive simulation (Wang et al., 15 Jan 2026).

The basic premise is straightforward: a model is told to act as a role, such as a mathematician, a tutor, a medical resident, a quiz contestant, a detective, or a fictional character. Role descriptions may specify identity, values, tone, speech habits, responsibilities, and constraints. RoleLLM formalized this as role profiles containing a role description, catchphrases, and structured dialogues, and then used these profiles for zero-shot prompting, few-shot dialogue engineering, and role-conditioned instruction tuning (Wang et al., 2023). Self-prompt tuning pushed the same idea further by training models to autonomously emit role prompts before answering, so that role selection becomes part of the model’s own generation process rather than an external prompting step (Kong et al., 2024).

PBRP has also been framed as a means of shaping decision-making rather than merely style. In MBTI-based experiments, persona prompts were used to test adaptability, exploration–exploitation trade-offs, reasoning, and safety, with reported differences across roles in flexibility, stability, exploratory behavior, and SD-3 trait scores (Shen et al., 2024). At the same time, other work treats PBRP more narrowly as an inference-time control interface for dialogue style, tool use, or persona consistency. This divergence in framing is central to the contemporary literature: some papers study PBRP as a lightweight behavioral scaffold, while others treat it as a possible route toward role-specific cognition.

2. Prompt architectures and control regimes

Prompt architectures for PBRP range from plain persona headers to structured, multi-component prompting systems. At minimum, a role prompt names the role and provides behavioral guidance. In medical QA, for example, one consistent template was [Background/Behavioral Guidance + Role Name] + "Please answer the question", instantiated for Medical Student, Resident, Attending Physician, Emergency Doctor, Surgeon, and other roles (Liang et al., 28 Oct 2025). In dialogue systems, role cards often also include worldview, scene state, recent function knowledge, and dialogue history (Ruangtanusak et al., 30 Aug 2025).

More elaborate systems layer additional control channels on top of the role prompt. Memory-augmented prompting injects episodic and semantic facts, relationship structure, and current situational context into the prompt at each turn, as in CHARMAP and MAP-style designs discussed in the survey literature (Wang et al., 15 Jan 2026). Role-sensitive NPC prompting can further combine symbolic schemas with fuzzy, numerical parameters such as guidance_intensity, evasiveness, and disclosure_prob, allowing symbolic constraints to coexist with graded improvisational freedom (Figueiredo et al., 29 Oct 2025). Retrieval-augmented role play goes in a different direction: rather than supplying a static profile, it conditions the model on curated reference demonstrations, catchphrases, and persona-authentic refusals, recasting role play as a retrieval problem (Rupprecht et al., 15 Sep 2025).

Several representative prompting regimes can be distinguished.

Regime	Core prompt content	Representative source
Role-card prompting	Identity, traits, style, constraints	(Wang et al., 2023)
Rule-based role prompting	Character Card, Scene, tool rules, one-call budget	(Ruangtanusak et al., 30 Aug 2025)
Retrieval-conditioned role play	Persona demonstrations, catchphrase buckets, guardrails	(Rupprecht et al., 15 Sep 2025)
Multi-agent prompting	Separate Eye, Ear, Brain, Mouth prompts	(Chen et al., 13 Apr 2026)
Thinking-before-speaking	Role_name (thinking): then Role_name (speaking):	(Zhang et al., 2024)
Identity-compositional prompting	Multiple identities combined hierarchically	(Sun et al., 2024)

Tool-augmented dialogue is one domain where prompt design becomes especially explicit. In the CPDC 2025 API track, four prompting strategies were compared: basic role prompting, human-crafted role prompting, automatic prompt optimization, and rule-based role prompting. The best system, rule-based role prompting, separated Voice from Action via a Character-Card/Scene-Contract and enforced Hard-Enforced Function calling with rules such as Action-first, Single-shot, Defer text, Schema-correct, and Ambiguity handling (Ruangtanusak et al., 30 Aug 2025). This is still PBRP, but it is PBRP under procedural constraints rather than free-form persona conditioning.

Multimodal extensions generalize the same logic. ActorMind treated speech role-playing as a four-agent chain-of-thought style pipeline: Eye Agent reads role profile and scene, Ear Agent derives emotion captions from spoken dialogue via SECAP, Brain Agent infers a descriptive emotional state, and Mouth Agent performs retrieval-augmented speech synthesis conditioned on that state (Chen et al., 13 Apr 2026). The paper explicitly maps these components back to text-only PBRP: Eye corresponds to persona and scene priming, Ear to emotion perception, Brain to role reasoning, and Mouth to stylized output generation.

3. Evaluation frameworks and measurement traditions

Evaluation in PBRP is unusually heterogeneous because the target construct is not just correctness. Recent frameworks measure role fidelity, narrative management, knowledge adherence, style, tool use, reward-model preference, and internal consistency, often simultaneously.

A broad survey of role-playing agents catalogs evaluation along role knowledge, personality fidelity, value alignment, and interactive hallucination, citing benchmarks such as RoleEval, CharacterEval, InCharacter, SHARP, TimeChara, RMTBench, and RVBench (Wang et al., 15 Jan 2026). RoleLLM’s RoleBench provided an earlier fine-grained benchmark dataset for role-playing with 168,093 samples across 100 roles, separating RAW, CUS, and SPE views for accuracy, customized style imitation, and role-specific knowledge (Wang et al., 2023).

Newer evaluation frameworks have become more targeted. RoleRMBench was introduced specifically for reward modeling in profile-based role play, covering seven capabilities: Introduction, Progression, Stitching, Scene Transition, Role Consistency, Instruction Following, Safety, Multi-turn Coherence, and Attractiveness, grouped under a pairwise reward-model evaluation protocol (Ding et al., 11 Dec 2025). FURINA-Bench instead evaluates per-utterance dimensions—Context Reliance, Factual Recall, Reflective Reasoning, Conversational Ability, and Preference Alignment—using a multi-agent construction pipeline and pairwise LLM judging with a bidirectional scoring rule (Wu et al., 8 Oct 2025).

Several task-specific metrics have also appeared. ActorMind introduced RP-MOS, a 1–5 subjective measure emphasizing Exact Delivery and Emotion Expression in speech role-playing (Chen et al., 13 Apr 2026). RAGs-to-Riches introduced two token-level ROUGE-style metrics, Intersection over Output and Intersection over References, to quantify how much a response reuses retrieved persona demonstrations (Rupprecht et al., 15 Sep 2025). The medical RPNA framework combined accuracy deltas from neuron ablation with Jensen–Shannon Divergence, CKA, PCA, K-means clustering, and silhouette coefficients to test whether role prompts produce distinct internal trajectories (Liang et al., 28 Oct 2025).

Classical prompt-generation work also used evaluator models and overlap metrics. The GPT-based role-playing prompt framework for dataset generation and evaluation used GPT-3.5 ranking on conversational ability, acting ability, and task response, then cross-validated those rankings with ROUGE-L against GPT-4 reference outputs (Liu et al., 2024). By contrast, MBTI-based decision-making work evaluated adaptability through flexibility and stability, exploration–exploitation via Kalman-filtered bandit regression, reasoning via MMLU accuracy, and safety through SD-3-style trait measurements (Shen et al., 2024).

This variety suggests that PBRP is not evaluated against a single latent target. Depending on the application, the benchmark may privilege canon knowledge, stylistic mimicry, reward-model preference, action policy, emotional delivery, or mechanistic separation.

4. Empirical performance: gains, null results, and domain dependence

Empirical results on PBRP are sharply domain-dependent. In some settings, carefully designed role prompts yield clear gains. In others, they have little measurable effect or improve only the presentation layer.

On reasoning and QA tasks, ORPP reported consistent improvements by optimizing within a role-playing prompt subspace rather than the full prompt space. For Qwen2.5-14B, accuracy improved from 43.43 to 45.45 on GPQA, from 67.21 to 70.29 on AGIEval-Math, and from 63.63 to 65.67 on MedQA. For Qwen2.5-32B, GPQA improved from 42.93 to 49.49 and MATH from 65.33 to 67.45 (Duan et al., 3 Jun 2025). SRPS, which extracts role-relevant SAE features and injects a steering vector into the residual stream, likewise reported zero-shot CoT gains such as Llama3.1-8B on CSQA improving from 31.86% to 39.80%, and Gemma2-9B on SVAMP increasing from 37.50% to 45.10% (Wang et al., 9 Jun 2025).

Self-prompt tuning reported more modest but broad improvements from role-annotated supervised fine-tuning. Mistral-Role improved over Mistral-LIMA on all eight reported benchmarks in average score, 45.0 to 48.4, with gains on MMLU, CommonsenseQA, StrategyQA, TruthfulQA, OpenBookQA, HumanEval, GSM8K, and Date Understanding; Llama-Role likewise improved average score from 37.3 to 40.6 over Llama-LIMA (Kong et al., 2024). In role-playing fine-tuning for character dialogue, generated datasets also produced substantial post-LoRA improvements: for example, Gemma-7B’s win rate against its original version was 94.87% in GPT-based evaluation, and ROUGE-L improved from 14.3428 to 26.6724 (Liu et al., 2024).

Prompt engineering shows especially strong effects when the task is not pure QA but structured interaction. In the CPDC 2025 API track, rule-based role prompting improved the overall score from 0.519 for the baseline to 0.571, with Task 1 rising to 0.531 and Task 2 remaining competitive at 0.611 (Ruangtanusak et al., 30 Aug 2025). In hostile-user role-playing scenarios, RAGs-to-Riches reported that its prompting strategy incorporated an average of 35% more tokens from reference demonstrations than zero-shot and standard ICL baselines, and across 453 interactions its models were consistently judged more authentic and more often in-character (Rupprecht et al., 15 Sep 2025). For speech role-playing, ActorMind achieved an average RP-MOS of 3.56 ± 0.27, compared with 3.05 ± 0.56 for IndexTTS and 1.00 ± 0.00 for Qwen-Omni in the reported setup (Chen et al., 13 Apr 2026).

Yet PBRP does not uniformly improve reasoning. The clearest counterexample comes from medical QA. Across six models and three datasets, role prompts in RPNA generally changed accuracy by under ±2%, with Cochran’s Q tests showing no statistically significant differences in most cases. On MedQA, Qwen2.5-72B scored 0.7227 as Medical Student, 0.7054 as Resident, 0.7069 as Expert Doctor, and 0.7014 as Surgeon; GPT-4o scored 0.8020, 0.7824, 0.7824, and 0.7941 for the same roles; DeepSeek-R1 varied only between 0.8884 and 0.8982 across listed roles (Liang et al., 28 Oct 2025). This study argues that role prompts in that setting modulate tone, hedging, and register, but not the underlying medical decision.

The same heterogeneity appears in subjective role simulation. Thinking Before Speaking reported that prompt-only methods often imitate tone without respecting knowledge boundaries, whereas supervised fine-tuning with explicit thinking and speaking traces improved Hallucination, Stability, and Adaptability scores; TBS_Llama3 reached an average of 6.81 on CharacterLLM metrics and 6.81 Overall on the new TBS metrics (Zhang et al., 2024). Symbolically Scaffolded Play found that a high-constraint prompt and a low-constraint prompt produced no reliable experiential differences in a within-subjects usability study, even though a later synthetic evaluation revealed clear role-dependent scaffold effects (Figueiredo et al., 29 Oct 2025).

A plausible implication is that PBRP is most effective when the role prompt is aligned with a task that rewards procedural discipline, persona-specific policy, or retrieval-conditioned style, and least effective when the task demands latent expertise that the model must already possess or learn through mechanisms beyond surface conditioning.

5. Mechanistic interpretations and the move beyond prompt-only control

One of the most active research directions asks whether PBRP merely steers outputs or actually changes internal computation. The current literature does not give a single answer, but it does establish several distinct mechanisms.

RPNA directly tested the hypothesis of role-specific cognition in medical LLMs. It identified role-sensitive neurons by comparing role-prompted and baseline hidden-state activations, ablated those neurons, and examined representation trajectories. The results were strongly negative for deep role separation: cross-role ablation produced nearly identical drops to within-role ablation, CKA similarity across roles was typically 0.96–1.00, silhouette scores ranged only from 0.11 to 0.14, and layer-wise JSD showed early differences that attenuated in deeper layers (Liang et al., 28 Oct 2025). In that setting, PBRP looked like early semantic framing plus stylistic encoding, followed by convergence to a shared reasoning backbone.

SRPS proposes a different account. Instead of trusting prompt wording alone, it maps role effects into sparse autoencoder latents, selects top-k role-relevant features via activation and frequency differences, and constructs a steering vector

$\mathbf{s} = \sum_{i \in K} \alpha_i \mathbf{v}_i$

that is injected into the residual stream as

$r_l' = r_l + \gamma \mathbf{s}.$

This makes role-playing interpretable at the feature level and controllable through layer choice, top-k selection, and injection strength (Wang et al., 9 Jun 2025). A closely related line extends this to persona facets: contrastive SAE training and trait-activated routing learn vectors aligned to the Big Five 30-facet model, then dynamically inject only the relevant facet vectors at a mid residual layer (Tang et al., 22 Feb 2026).

Other work responds to the limitations of prompt-only control by externalizing parts of the role logic. Beyond Dialogue removes profile–dialogue mismatch by adjusting the profile prompt per scenario and by training on “beyond dialogue” alignment tasks that explicitly map Character, Style, Emotion, Relationship, and Personality to dialogue evidence (Yu et al., 2024). Thinking Before Speaking inserts an explicit Role_name (thinking): ... trace before Role_name (speaking): ..., so that role-specific logic and knowledge boundaries become part of the supervised target rather than an implicit byproduct of prompting (Zhang et al., 2024). Codified Profiles go further still by representing character logic as executable functions such as parse_by_scene(scene) and check_condition(scene, question), replacing free-form role prose with explicit control flow and stochastic gates (Peng et al., 12 May 2025).

These systems are not prompt-only in a strict sense, but they remain part of the same intellectual trajectory. They respond to a recurring empirical observation: prompt text is easy to apply, but difficult to debug, weakly persistent, and often too implicit for stable role-conditioned behavior. This suggests that the long-term development of PBRP may lie less in longer persona descriptions than in hybrids that connect prompts to memory modules, structured latent control, or explicit behavioral logic.

6. Applications, limitations, and future directions

PBRP now spans a wide application range. In medical AI it is used to simulate clinical roles and expertise gradations (Liang et al., 28 Oct 2025). In tool-augmented agents it helps align in-character dialogue with function calling and action policies (Ruangtanusak et al., 30 Aug 2025). In speech systems it enables persona-conditioned verbal delivery with emotional cues (Chen et al., 13 Apr 2026). In game dialogue and NPC design it supports detectives, suspects, quest-givers, and procedurally generated characters, while raising role-specific questions about how much scaffolding different narrative roles should receive (Figueiredo et al., 29 Oct 2025). In benchmark design it underlies fully customizable role-playing evaluation pipelines such as FURINA-Builder, which generates multi-party dialogues and evaluates them on Context Reliance, Factual Recall, Reflective Reasoning, Conversational Ability, and Preference Alignment (Wu et al., 8 Oct 2025).

Several limitations recur across these domains. Prompt signal dilution in long contexts leads to persona drift and inconsistency, especially when memory, retrieval passages, and dialogue history compete for attention (Tang et al., 22 Feb 2026). Prompt-only systems can over-speak, hallucinate function calls, or mis-handle argument schemas in tool environments (Ruangtanusak et al., 30 Aug 2025). Voice and speech role-playing are sensitive to diarization, ASR, and prompt-length degradation (Chen et al., 13 Apr 2026). Role-playing reward models trained for generic helpfulness can fail badly on narrative coherence and stylistic fidelity, which motivated RoleRM and Continuous Implicit Preferences (Ding et al., 11 Dec 2025). FURINA further reports a broader Pareto frontier between RP performance and reliability: reasoning can improve role-play quality while simultaneously increasing role-play hallucinations, and model scale does not monotonically reduce hallucinations (Wu et al., 8 Oct 2025).

The future directions proposed across the literature are convergent. The survey work emphasizes personality evolution modeling, multi-agent collaborative narrative, multimodal immersive interaction, and integration with cognitive neuroscience (Wang et al., 15 Jan 2026). Mechanistic papers call for role-conditioned training objectives, modular architectures, representation steering, tool- and memory-integrated cognition, and causal circuit editing (Liang et al., 28 Oct 2025). Persona-control work argues for facet-level routing and contrastive latent supervision rather than relying on diluted text descriptions (Tang et al., 22 Feb 2026). Reward-model work suggests that subjective role alignment requires specialized, consistent preference supervision rather than general-purpose evaluators (Ding et al., 11 Dec 2025).

Taken together, these results depict PBRP not as a single technique but as a spectrum of control strategies. At one end lies prompt-only style imitation; at the other lie structured, memory-rich, and representation-aware systems that preserve the convenience of prompting while reducing its brittleness. The central unresolved question is no longer whether role prompts can change outputs—they clearly can—but under what conditions they change reasoning, policy, and long-horizon behavioral consistency rather than merely the way a model sounds.