Interactive Virtual Personas

• Interactive Virtual Personas are computational agents that enable real-time, multimodal conversational interactions, blending static profile data with dynamic dialogue.
• They are constructed using diverse pipelines such as LLM-driven narrative generation, speech-to-text transcription, and multimodal synthesis for realistic and context-aware simulations.
• IVPs are applied in human-centered design, accessibility support, and healthcare simulation, providing actionable insights and innovative interactive solutions.

Interactive Virtual Personas (IVPs) are computational personas that can be interacted with rather than only read about. In the recent literature, the term has been used most explicitly for “multimodal, LLM-driven, conversational user simulations” that designers can interview, brainstorm with, and ask for feedback in real time, but adjacent work also includes embodied conversational personas, real-time talking avatars, voice-enabled virtual patients, AI guides in social VR, and persistent multi-persona conversational configurations (Deep et al., 26 Aug 2025, Bowden et al., 2017, Cheng et al., 2024, Botero et al., 1 Nov 2025, Collins et al., 2024, Taheri et al., 2024). A central distinction separates AI-generated personas, whose primary artifact is a static profile, from AI-simulated personas, whose primary artifact is interactive dialogue; several systems occupy intermediate positions, including VR workflows where personas act as shared discussion proxies or semi-interactive design artifacts rather than autonomous embodied agents (Deep et al., 26 Aug 2025, Wang et al., 4 Feb 2026, Wang et al., 5 Mar 2026).

1. Conceptual scope and lineage

IVPs inherit from older work on virtual humans and embodied conversational agents. Review-oriented work on virtual humans frames realism as a multidimensional problem involving appearance, movement, emotional expression, social response, and contextual appropriateness rather than appearance alone, while the “Virtual Human Journalist” treated the core challenge as building a humanoid conversational agent capable of eliciting information through sustained, multimodal conversation with speech, gestures, facial expressions, and emotional responsiveness (Montanha et al., 2023, Bowden et al., 2017). In that lineage, an IVP is less a single technology than a configuration in which identity, interaction, and some degree of persistence are coupled.

Recent IVP work has sharpened the conceptual boundary between static persona artifacts and interactive persona agents. “AI-generated personas” are defined as systems for “Creating or enriching persona descriptions,” typically yielding static profiles with goals, needs, pain points, quotes, demographics and images, whereas “AI-simulated personas” are defined as systems for “Enabling real-time, conversational interaction with a persona,” yielding an “Interactive AI agent producing dialogic behaviour involving responses and role-play” (Deep et al., 26 Aug 2025). This distinction is useful because several systems contribute to IVPs only partially. In VR-supported requirements engineering, for example, personas generated from speech transcripts act as “discussion proxies and role identities more than as live agents,” and are characterized as “static or semi-static generated personas in immersive space, not fully interactive persona agents” (Wang et al., 4 Feb 2026). Accessibility personas generated from VR app-store reviews likewise remain “semi-interactive design artifacts”: interactive in how they are requested, explored, and compared, but not autonomous virtual beings (Wang et al., 5 Mar 2026). This suggests that IVPs are best understood as a spectrum running from static role artifacts embedded in interactive workflows to fully conversational, embodied, and stateful agents.

2. Persona construction and representation

The literature uses several distinct pipelines for constructing persona representations. One line of work derives personas from live or recorded human traces. In VR-supported requirements discussion, participant audio is transcribed through Azure speech-to-text and analyzed by a GPT-4-based pipeline to produce personas containing demographic attributes, inferred biographies, accessibility requirements, pain points, challenges, expertise, and user needs; the same paper explicitly notes that the system may “infer or fabricate” missing details, creating a risk of hallucination or synthetic overreach (Wang et al., 4 Feb 2026). In accessibility-focused VR education, reviews from the Meta Quest Store and Steam VR store are filtered by disability-related keywords and fuzzy matching, embedded with a sentence-transformer model, stored in Chroma, and passed through a GPT-4o RAG pipeline to generate a brief biography, pain points, representative quotes, explicit requirements, and a DALL·E 3 profile image (Wang et al., 5 Mar 2026).

A second line uses richer, semi-structured psychological representations. SPIRIT’s Painter infers personas from public social media posts and outputs a JSON plus a third-person narrative. The structured portion includes personality_big5, primal_world_beliefs, values_and_identities, life_experiences, opinions_and_beliefs, interaction_style, and meta, each with uncertainty-aware fields such as confidence and rationale (Li et al., 28 Mar 2026). Another approach, developed for political opinion simulation, conditions personas on long synthetic backstories generated as multi-turn interview transcripts. It then matches virtual personas to human respondents using maximum-weight bipartite matching, with edge weights defined as $$w(e_{ij}) = w(h_i, v_j) = \prod_{l=1}^k P(d_{jl}=t_{il})$$ (Kang et al., 16 Apr 2025). That work argues that short demographic prompts produce “shallow binding,” whereas long, coherent narratives support “deep binding” in which a persona better preserves ingroup, outgroup, and meta-perception structure (Kang et al., 16 Apr 2025).

A third line emphasizes scale and coverage. Persona Generators defines a generator function $G_{\phi,\theta}$ that produces synthetic populations from a context, diversity axes, and a target population size, and optimizes generator code for support coverage rather than density matching (Paglieri et al., 3 Feb 2026). Persona Hub pushes this logic further by automatically curating 1,015,863,523 personas from web text using Text-to-Persona and Persona-to-Persona expansion, then deduplicating them with MinHash and embedding similarity thresholds (Ge et al., 2024). These systems do not themselves yield interactive agents, but they provide infrastructure for population-scale IVP initialization.

3. Architectures, embodiment, and interaction mechanics

IVP architectures range from prompt-configured conversational systems to explicit embodied controllers. In design support, the customized GPT-4 IVP “Alice” was implemented using the OpenAI GPTs editor as a “voice-enabled, multimodal interface” with voice as the primary mode, optional text input/output, image input for wireframe critique, preserved conversation history, and prompt engineering intended to produce short, first-person, informal, role-consistent replies across user research, ideation, and prototype evaluation (Deep et al., 26 Aug 2025). In product design, Personagram instead deemphasizes chat and operationalizes personas through a structured multimodal pipeline: persona attributes drive GPT-4o-mini product inference, Google Image Search retrieves references, a multimodal model extracts aesthetic, behavioral, and contextual features, and Flux generates concept images from recombined features (Kim et al., 5 Feb 2026).

Embodied IVPs add avatar synthesis, animation, and situated action. RITA turns a single user-uploaded image into a responsive talking avatar by separating expensive offline generation from lightweight online inference: an avatar-specific frame library is precomputed, LLMs generate dialogue, TTS generates response audio, speech hyperparameters are matched against stored hyperparameters, and RIFE performs video interpolation. The paper reports average runtimes of 0.09 seconds for hyperparameter embedding and frame matching and 3.97 seconds for interpolation (Cheng et al., 2024). MASK offers a different embodiment strategy: persona is compiled into a deterministic finite-state controller over a discrete nonverbal action space. The observation space has 72 combinations derived from raised hands, distance, gaze, hand velocity, and approach/static/leave; the state is defined as $s=(f,m)$ over 12 facial expressions and 13 motions; and persona-specific transition rules are precomputed as $T(s' \mid s,o)$ (Park et al., 2024). For social VR accessibility, an AI guide in Unity on Meta Quest 2 uses user speech, first-person and bird’s-eye screenshots, and object metadata to classify queries into holistic description, specific visual question, navigation request, beacon request, or other conversation, then responds through speech, movement, audio cues, and haptics across six distinct personas including Human, Guide Dog, White Cane, Robot, Bird, and Invisible (Collins et al., 2024).

Other systems privilege voice and profile control rather than visual embodiment. The voice-enabled virtual patient system uses Amazon Transcribe, Anthropic’s Claude Sonnet 3.7, and TTS back ends including Amazon Polly, ElevenLabs, and Cartesia to simulate patients whose personas are defined by item-level MADRS symptom targets, demographics, biography, and communication style such as “cooperative” or “guarded” (Botero et al., 1 Nov 2025). Virtual Buddy uses a web-based interface with GPT-4 and MongoDB to maintain multiple persistent persona profiles defined by Role, Personality, and Needs, allowing one user to select among several reusable “virtual buddies” instead of interacting with a single companion (Taheri et al., 2024). These architectures show that IVP interactivity can be realized through different combinations of dialogue, embodiment, persistence, and world-grounded action.

4. Application domains

Human-centred design is one of the most explicit IVP application areas. In the Alice study, professional UX designers used an IVP as stakeholder, co-creator, and tester across user research, ideation, and prototype evaluation, especially valuing rapid interviewing and wireframe feedback (Deep et al., 26 Aug 2025). Personagram addresses a related problem from the opposite direction: instead of treating the persona as a speaking character, it turns the persona into a structured design instrument that links census-based user profiles to product references, feature extraction, and image-based ideation (Kim et al., 5 Feb 2026). This suggests two distinct design uses for IVPs: conversational surrogate users and multimodal persona workbenches.

Accessibility and immersive systems form a second major domain. Virtual Buddy reframes persona multiplicity as an accessibility mechanism for people with hand motor disabilities by reducing repeated setup prompts through persistent topic- and role-specialized buddies (Taheri et al., 2024). The AI Guide paper reframes accessibility support in social VR as a socially embodied, persona-driven companion rather than background assistive feedback (Collins et al., 2024). In requirements engineering, automatically generated personas support accessibility-focused discussion in VR meeting rooms, while review-grounded persona generation in VR education is used to surface latent accessibility requirements and support empathy development (Wang et al., 4 Feb 2026, Wang et al., 5 Mar 2026).

Healthcare and research simulation form a third domain. The voice-enabled virtual patient system treats IVPs as scalable, realistic patient simulations for standardized clinical assessment training (Botero et al., 1 Nov 2025). VLM personas are used as proxy participants in embodied HCI, where each persona interprets video from a first-person street-crossing perspective, makes sequential crossing decisions, and later reports subjective ratings (Gui et al., 18 Feb 2026). At population scale, SPIRIT persona banks are used as weighted virtual respondent panels for stable attitudes and emerging public-opinion questions (Li et al., 28 Mar 2026). Persona Hub and Persona Generators extend the idea further by supplying synthetic populations, user simulations, game NPCs, and tool/function specifications (Ge et al., 2024, Paglieri et al., 3 Feb 2026).

5. Empirical evidence and evaluation patterns

Evidence for IVPs is mixed: some systems have direct user studies, some have expert validation, and others remain largely conceptual. In design support, the Alice study involved eight professional UX, UI, and product designers and found that usefulness, relevance, flexibility, and authenticity clustered around a median of 6 on a 7-point scale, with prototype evaluation judged strongest; “real-person believability” differed by experience level, with juniors at Md = 6 and seniors at Md = 4 (Deep et al., 26 Aug 2025). In product design, Personagram was evaluated with 12 professional physical product designers and showed 187 button interactions versus 92 text queries in the chat baseline with $W = 55, p = 0.002$, alongside higher perceived transparency (5.33 vs 3.92, $W = 10.5, p = 0.026$) and better NASA-TLX performance ratings (5.67 vs 5.00, $W = 0.0, p = 0.027$) (Kim et al., 5 Feb 2026).

In VR and embodiment studies, the evidence is more heterogeneous. The VR-supported requirements discussion system reported 18 participants, social presence Mean = 5.21, SD = 0.81, system usability Mean = 5.19, SD = 0.78, and VR workload Mean = 3.57, SD = 0.85, with a significant workload difference favoring VR over manually created personas in a physical meeting room at $p < .001$ (Wang et al., 4 Feb 2026). MASK, which evaluates persona recognition rather than task performance, reported 108 participants and an overall character-recognition accuracy of 76.7%, above the authors’ success threshold of 67%, with Cowardly Lion, Minion, and Scrooge easier to identify than Spock (Park et al., 2024). RITA reports system-oriented runtime evidence rather than human-subject outcome measures, with 0.09 seconds for hyperparameter embedding and frame matching and 3.97 seconds for interpolation (Cheng et al., 2024).

Clinical and simulation-oriented IVPs have some of the strongest fidelity-style evaluations. In the virtual patient system, 5 experienced clinical raters conducted 20 simulated MADRS interviews across 4 personas, producing a mean item difference of 0.52 with SD = 0.75 between clinician-assigned and configured scores and total-score inter-rater reliability of ICC(2,1) = 0.90, 95% CI = 0.68-0.99 (Botero et al., 1 Nov 2025). In embodied HCI simulation, VLM personas produced average crossing times of 5.25 s versus 5.07 s for human participants with $p = 0.8465$, but also showed lower variability and “hyper-confidence,” with confidence ratings of 4.53 versus 3.50 for humans at $p < 0.001$ (Gui et al., 18 Feb 2026). In repeated social interaction, persona-conditioned agents in Split or Steal showed that mutual Split dominated at roughly 74 percent of rounds, exploitation occurred in fewer than 11 percent of rounds, and Prosocial and Principled personas were more cooperative while Analytical personas were more likely to exploit the VH (Leon et al., 3 May 2026). Across these studies, the dominant evaluation pattern is that IVPs can often be made useful, coherent, or behaviorally distinctive, but evidence for long-horizon realism and general-purpose validity remains limited.

6. Controversies, limits, and research directions

The strongest recurring controversy is whether IVPs should be treated as substitutes for real people. The Alice study states that IVPs should be viewed as “a complement to, not a replacement for, real user engagement,” and identifies over-optimism as a persistent problem: the IVP was “useful, but a bit too optimistic,” often affirming ideas without surfacing costs, constraints, or failure modes (Deep et al., 26 Aug 2025). Related work on political simulation argues that many persona methods remain shallow and stereotype-like unless they achieve “deep binding,” meaning that the persona preserves higher-order social cognition rather than merely echoing outgroup expectations (Kang et al., 16 Apr 2025). This suggests that IVP realism must be evaluated at the level of perspective structure, not just surface fluency.

A second recurring issue is that higher realism is not always better. The Virtual Human Journalist paper argues that “naturalness should not always be seen as a desirable goal” and that suppressing naturalness or altering personality cues can sometimes yield more desirable results because highly human-like presentation raises expectations for empathy, prosody, and contextual understanding that current systems cannot meet (Bowden et al., 2017). Review work on virtual humans similarly highlights the uncanny valley, perceptual comfort, bias transfer, and the need for inclusivity and cultural sensitivity (Montanha et al., 2023). In social media contexts, audience discourse around virtual influencers is structurally different from discourse around human influencers, with an appearance-discourse cluster absent from human-influencer comments and substantially more negative sentiment in psychologically sensitive domains such as mental health, body image, and artificial identity (Chaudhry et al., 25 Mar 2026). Together, these results indicate that virtuality itself changes how personas are interpreted.

A third concern is grounding and trustworthiness. Several systems openly rely on inferential generation that may fabricate or overreach. The VR requirements discussion system can “reasonably infer or fabricate” missing demographic details and provides no validation mechanism for generated persona attributes (Wang et al., 4 Feb 2026). Review-grounded VR personas have limited fidelity evaluation and weakly specified demographic inference (Wang et al., 5 Mar 2026). The virtual patient paper notes that robust guardrails and continuous validation will be essential because LLMs remain susceptible to factual inconsistency and clinically implausible generations (Botero et al., 1 Nov 2025). Virtual Buddy raises a related ethical issue from another direction: over-attachment to highly human-like chatbots, with the authors suggesting that one possible response is to make such systems “look less human” (Taheri et al., 2024).

Future work across the literature converges on several directions: richer embodiment, better grounding, stronger memory, and more careful evaluation. Proposed extensions include embodied persona agents in VR, turn-taking and dialogue capabilities, persistent memory across discussion, real-time updates as conversation unfolds, scenario-based simulations, multimodal representations, prompt libraries designed to elicit critique rather than sycophancy, portfolios of diverse personas rather than reliance on a single archetype, and broader support coverage so that rare or long-tail user types are not excluded (Wang et al., 4 Feb 2026, Wang et al., 5 Mar 2026, Deep et al., 26 Aug 2025, Paglieri et al., 3 Feb 2026). A plausible implication is that mature IVPs will combine three properties that are often separated in current work: psychologically grounded persona representations, embodied and world-aware interaction loops, and evaluation protocols that test not only usefulness or plausibility but also diversity, calibration, and failure under non-average conditions.