REVIEW 4 major objections 6 minor 48 references
Adaptive-VP: A Framework for LLM-Based Virtual Patients that Adapts to Trainees' Dialogue to Facilitate Nurse Communication Training
T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash
Pith's one-line read Adaptive-VP closes the loop between trainee skill and virtual patient behavior.
desk verdict A solid applied LLM virtual patient system with a real dynamic adaptation loop; the abstract overclaims against existing approaches, but the core dynamic-vs-static comparison is sound. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The load-bearing mechanism is the evaluation-to-adjustment loop. The Evaluation Module renders a 0-5 communication efficiency score from a rubric: utterance-level checks (tone calm and clear; empathy level 3+ out of 6; absence of premature empathy, invalidating beliefs, and dismissive commands) plus conversation-level checks (use of autonomy, limit-setting, and problem-solving/reframing de-escalation strategies), with three LLM evaluator agents required to agree unanimously. The Dynamic Adjustment Module takes that score and picks a pre-defined behavioral direction, ranging from full refusal and personal attacks at score 0 to basic cooperation with noticeable resistance at score 5, which controls the patient's communication style, complaint intensity, and responsiveness. The Dialogue Generation Module then writes the VP's next utterance with inner monologue, verbal response, and non-verbal cues, and the Safety Monitoring Module filters outputs that are excessively hostile, off-goal, inconsistent with the persona, or misaligned with the direction. The loop is what distinguishes Adaptive-VP from static VP systems.
What would settle it
A validation study in which blinded expert nurses rate the same trainee utterances that the multi-agent evaluator scored; if expert holistic ratings fail to correlate with the 0-5 scores, or if a randomized training trial shows no improvement in trainees' real patient interactions after practicing with Dynamic VPs versus Static VPs, the central claim would be falsified.
Extended reading notes
Core claim
At its core, the paper claims that dialogue realism in virtual patient simulation is improved by a closed feedback loop between trainee communication quality and patient behavior. Adaptive-VP operationalizes this with a four-module architecture: a case-development pipeline produces clinically grounded patient personas; an evaluation module uses three LLM agents (a nursing professor, communication skills trainer, and clinical psychologist) to score each trainee utterance for calm/clear tone, empathy level, use of de-escalation strategies, and prohibited behaviors; a dynamic adjustment module converts the 0-5 score into an escalation or de-escalation directive; and a safety monitoring module checks each VP utterance before it is shown. The paper's evidence is that the scoring mechanism separated experienced from novice nurses on real dialogue corpora, and that in a between-subjects study nurses rated dynamic VPs significantly higher on role fidelity and conversational realism than static VPs with identical personas. The intended consequence is that trainees can practice managing aggressive or non-cooperative patients in a setting that responds the way real patients would, while a safety layer keeps the simulation from causing distress.
Load-bearing premise
The whole adaptation loop depends on the assumption that the hand-designed rubric (calm/clear tone, empathy at level 3 or above, three de-escalation strategies, three prohibited behaviors, unanimous LLM agreement) actually measures real nursing communication skill; if it rewards the wrong behaviors, the virtual patient escalates or de-escalates for the wrong reasons, and the expert-novice difference could be an artifact of style rather than proficiency.
Editorial extensions
If this is right
- Nurse educators can offer repeatable, low-cost practice with difficult patient types such as overdependent, authoritative, aggressive, and uncooperative patients, which are hard to rehearse with standardized actors.
- The adaptation loop is reusable across training goals by swapping the rubric criteria and the behavioral direction table, so the same four-module architecture could serve other healthcare communication skills.
- The safety monitoring module makes it possible to keep simulations challenging without exposing learners to toxic or demeaning output, a precondition for using such systems with novice nurses.
- The pipeline can generate diverse clinical dialogue corpora in privacy-constrained settings, because the cases are synthetic but clinically grounded and expert-validated.
- Trainee performance can be tracked turn-by-turn as a score, giving instructors an automated measure of communication skill progress during practice.
Reading between the lines
- The authors stop short of measuring learning outcomes; a natural extension the paper does not report is a controlled trial comparing novice nurses' real-patient communication after Dynamic VP practice versus Static VP or standardized-patient training.
- The 0-5 score's threshold design (empathy must reach level 3; a prohibited behavior costs 1 point regardless of frequency) is a tunable choice; alternative weightings could be tested against expert judgment to see whether the adaptation curve becomes more sensitive.
- Because the framework currently targets South Korean ward contexts, the case pipeline and rubric would need cultural adaptation; the paper leaves implicit that de-escalation strategies and the 'authoritative patient' type may carry different expectations in other healthcare systems.
- The safety module filters extreme hostility, which creates a tension the authors acknowledge only indirectly: if the most realistic patient behavior is exactly the toxicity being filtered, then realism and learner safety are in partial competition; a testable extension is to measure at what filter threshold learners still find the interaction credible.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper presents Adaptive-VP, an LLM-based framework for virtual patient (VP) dialogue generation aimed at nursing communication training, with a focus on challenging patient interactions. The framework comprises a VP case development pipeline, a multi-agent evaluation module that scores trainee utterances on tone, empathy, prohibited behaviors, and de-escalation strategies, a dynamic adjustment module that maps scores to VP hostility/cooperativeness, a dialogue generation module, and a safety monitoring module. The authors validate the case pipeline with 10 nursing experts, the evaluation module with dialogue corpora from 15 expert and 15 novice nurses (where expert–novice discrimination is reported at p=0.001), and the perceived realism of Dynamic versus Static VPs with a between-subjects experiment involving 28 nurses (significant effects on Role Fidelity and Conversational Realism). The central claim is that Adaptive-VP produces adaptive, realistic interactions that improve on static VP approaches.
Significance. If the results hold, the framework is a credible contribution to simulation-based communication training: it offers a structured pipeline for clinically grounded VP cases, a multi-agent evaluation design with reported inter-rater agreement, and evidence that the dynamic loop improves perceived realism over a static baseline. The paper also ships code and data, and the expert–novice discrimination result is directly relevant to the evaluation module's validity. However, the headline comparative claim in the abstract ('more natural and realistic interactions than existing approaches') is not supported by the experimental design, which compares only the Dynamic and Static variants of the same framework. The user-facing value also depends on the construct validity of the hand-designed communication rubric, which is plausible but not independently validated. The work is therefore significant as a well-specified framework demonstration, but its practical superiority over prior VP systems remains an untested assumption.
major comments (4)
- [Abstract and §8] The abstract states that Adaptive-VP produced 'more natural and realistic interactions than existing approaches,' but the human evaluation in §8 compares only the Dynamic variant against the Static variant of the same framework. None of the cited prior VP systems (e.g., PATIENT-ψ, MedSimAI, or scripted SP-based chatbots) is included as a comparison condition. The significant main effects for Role Fidelity (p=.043) and Conversational Realism (p=.008) support the bundled claim that the dynamic loop improves realism over a static version, but they do not test the comparative claim against existing systems. This gap is load-bearing for the practical-value argument and should be fixed either by adding an external baseline comparison or by rewriting the abstract and conclusions to claim only 'more realistic than a static VP,' not 'more realistic than existing approaches.'
- [§5.1, §5.2, and Appendix D] The communication-efficiency rubric in Table 2 and the score-to-direction mapping in Table 8 are hand-designed and contain free parameters (e.g., the empathy threshold of level ≥3, the choice to require unanimous multi-agent agreement, and the ordinal mapping of the 0–5 score to hostility levels). The expert–novice validation in §5.3 shows that the rubric discriminates between groups, which is encouraging, but it does not establish that the rubric measures the 'correct' behaviors for effective nursing communication. If the rubric rewards stylistic features that are not true markers of proficiency, the dynamic adjustment will train the wrong skills. I do not see this as a circularity problem—there is no fitted-parameter derivation—but it is a construct-validity risk. The paper should either validate the rubric against an established communication assessment instrument (e.g., human expert ratings on the same utterances) or clearly frame the rubric as a provisional, theory-based approximation.
- [§7 and Appendix F] The Safety Monitoring Module is described as a core contribution ('ensuring learner safety'), but the paper provides no empirical evaluation of the safety filter: no pass/fail rates, no examples of caught unsafe utterances, no measurement of how often the dialogue generator's initial output is rejected, and no human assessment of whether the filter preserves realism while removing harmful content. Since the framework's fourth design goal rests on this module, and since the human evaluation in §8 does not measure emotional distress or safety incidents, the safety claim is currently unsupported. I recommend adding at least a descriptive analysis of the safety module's behavior on the collected conversations.
- [§5.3 and §8] The evaluation module validation truncates conversations to the first five utterances, and the expert–novice corpus has average conversation lengths of 7.45 vs 5.3 turns. The authors do not justify the five-utterance truncation or report how many conversations were actually truncated. Similarly, the Dynamic-vs-Static experiment uses n=14 per condition, which is underpowered for interaction effects, and the Wind et al. (2004) questionnaire was adapted without reporting the adaptation details. These issues do not overturn the reported main effects, but they limit the strength of the inference and should be discussed explicitly in the limitations section, which currently mentions only the small sample and Korean context.
minor comments (6)
- [Throughout] The terms 'communication efficiency' and 'communication effectiveness' are used interchangeably; please pick one and define it consistently.
- [Table 5 (B.2)] The patient type is misspelled as 'overly authoriative'; it should be 'overly authoritative.'
- [B.3 (Table 6)] In the aggressive patient's profile, 'Past Surgical History & Date' lists 'Current appendectomy (2 days ago),' yet the situation describes cervical disc surgery performed 3 days ago. This appears to be a copy-paste error and should be corrected.
- [§5.2] The requirement of unanimous agreement among the three evaluator agents is mentioned, but the paper does not report how often the agents disagreed and whether a final score was then obtained by re-prompting or by majority vote. Please describe the handling of disagreements.
- [Appendix G.1] The subcomponent analysis reports that empathy-level differences were marginal (p=0.0587 by Mann–Whitney U). This is relevant to the rubric's discriminatory power and should be mentioned in the main text's summary of the validation, not only in the appendix.
- [§8] The participants in the human evaluation are described as 'experienced nurses (≥3 years),' while the expert group in §5.3 has a mean tenure of 12.5 years. The paper should clarify whether these are the same or different populations, since the realism ratings may depend on the evaluators' level of expertise.
Circularity Check
No significant circularity: the evaluation rubric is sourced from external nursing literature, and the realism ratings come from independent human judges rather than the framework's own scoring.
full rationale
The paper's derivation chain is not circular. The Evaluation Module's criteria (calm/clear tone, empathy level 3+, three de-escalation strategies, and three prohibited behaviors) are explicitly drawn from external sources: Price et al. (2024), Kleinsmith et al. (2015), and TMLT (2022b). They are not fitted to the expert/novice corpus, so showing that experienced nurses score higher is a construct-validity check rather than a prediction forced by construction. The human evaluation compares Dynamic versus Static VPs using a six-item realism questionnaire adapted from Wind et al. (2004), with ratings given by practicing nurses who are independent of the framework's scoring modules; the significant main effects on Role Fidelity and Conversational Realism are therefore evidence about the adaptive loop, not an artifact of the system's own evaluation. The abstract's broader phrasing that Adaptive-VP is 'more natural and realistic than existing approaches' goes beyond what the Static/Dynamic comparison can establish, but that is a gap in external validity or a correctness concern, not a circular reduction of the result to its inputs. No load-bearing self-citation chain or uniqueness-imported-from-authors pattern appears; citations such as Chan et al. (2023) and Louie et al. (2024) are used for established multi-agent evaluation techniques and prior VP systems, not to authorize the paper's own conclusions.
Assumptions & free parameters
free parameters (2)
- Empathy pass threshold =
level 3 on a 0 to 6 scale
- Score-to-direction mapping levels =
five discrete levels for scores 0 to 5
assumptions (4)
- domain assumption The evaluation rubric (calm/clear tone, empathy level 3+, three de-escalation strategies, three prohibited behaviors) constitutes valid nurse communication effectiveness.
- domain assumption Multi-agent unanimous LLM evaluation is a reliable proxy for expert human assessment.
- domain assumption Perceived realism of text dialogue is a sufficient proxy for training effectiveness.
- domain assumption The Static vs Dynamic comparison isolates the effect of adaptation.
Cite this review
Pith. "Pith review of Adaptive-VP: A Framework for LLM-Based Virtual Patients that Adapts to Trainees' Dialogue to Facilitate Nurse Communication Training." pith.science (2026). https://pith.science/paper/VMI3DPGO
@misc{pith2026250600386,
author = {Pith},
title = {Pith review of: Adaptive-VP: A Framework for LLM-Based Virtual Patients that Adapts to Trainees' Dialogue to Facilitate Nurse Communication Training},
year = {2026},
howpublished = {\url{https://pith.science/paper/VMI3DPGO}},
note = {Machine review of arXiv:2506.00386}
}
read the original abstract
Effective communication training is essential to preparing nurses for high-quality patient care. While standardized patient (SP) simulations provide valuable experiential learning, they are often costly and inflexible. Virtual patient (VP) systems offer a scalable alternative, but most fail to adapt to the varying communication skills of trainees. In particular, when trainees respond ineffectively, VPs should escalate in hostility or become uncooperative--yet this level of adaptive interaction remains largely unsupported. To address this gap, we introduce Adaptive-VP, a VP dialogue generation framework that leverages large language models (LLMs) to dynamically adapt VP behavior based on trainee input. The framework features a pipeline for constructing clinically grounded yet flexible VP scenarios and a modular system for assessing trainee communication and adjusting VP responses in real time, while ensuring learner safety. We validated Adaptive-VP by simulating challenging patient conversations. Automated evaluation using a corpus from practicing nurses showed that our communication skill evaluation mechanism reflected real-world proficiency levels. Expert nurses further confirmed that Adaptive-VP produced more natural and realistic interactions than existing approaches, demonstrating its potential as a scalable and effective tool for nursing communication training.
Figures
Figures from the paper (6 more)
Reference graph
Works this paper leans on
-
[1]
Carefully review the patient profile, paying special attention to the communication style section
-
[2]
Focus on the key characteristics that define how this patient interacts with healthcare providers
Summarize the patient’s communication style in five sentences or fewer in Korean. Focus on the key characteristics that define how this patient interacts with healthcare providers
-
[3]
These are behaviors or tendencies that may pose challenges for nurses during interactions
Within your summary, make sure to highlight the problematic aspects of the patient’s communication style. These are behaviors or tendencies that may pose challenges for nurses during interactions
-
[4]
Do not specify a numerical scale. Instead, focus on describing how these communication characteristics manifest in actual conversations
-
[5]
The expressions should clearly convey the patient’s needs
Write two example expressions that the patient might use when first addressing a passing nurse. The expressions should clearly convey the patient’s needs. These should be direct quotes that illustrate the patient’s communication style. Make sure to write realistic and natural Korean expressions that the patient would likely use in real-life situations. Av...
-
[6]
Philip M Podsakoff, Scott B MacKenzie, Jeong-Yeon Lee, and Nathan P Podsakoff
Managing patient aggression in healthcare: Initial testing of a communication accommodation theory intervention.Psychology of Language and Communication, 25(1):62–81. Philip M Podsakoff, Scott B MacKenzie, Jeong-Yeon Lee, and Nathan P Podsakoff. 2003. Common method biases in behavioral research: a critical re- view of the literature and recommended remedi...
arXiv 2003
-
[7]
Sally Spencer, Paula Johnson, and Ian C Smith
Affecting factors of new nurse’s intention to retention in hospitals.Journal of muscle and joint health, 24(3):205–216. Sally Spencer, Paula Johnson, and Ian C Smith
-
[16]
Focus on portraying a realistic patient image for this research-based simulation. However, exclude any content about complaining to higher authorities such as the hospital director or head nurse
Show all 48 references
-
[17]
Present your analysis in the following format: <analysis> <summary> {Your five-sentence summary of the patient’s communication style, including problematic aspects} </summary> <example_expressions>
-
[18]
[First example expression]
"[First example expression]"
-
[19]
[Second example expression]
"[Second example expression]" </example_expressions> </analysis Remember to base your analysis solely on the information provided in the patient profile. Do not invent or assume details that are not present in the given information. B Generated VP Cases The following are the e...
-
[20]
Examine the conversation between the patient and the nurse: <conversation> {CONVERSATION}</conversation>
-
[21]
Focus on the following aspects:
Analyze the nurse’s last response in the conversation, considering the patient’s profile. Focus on the following aspects:
-
[22]
Use these guidelines: Calm: Evaluate whether the nurse’s tone was calm
Tone Assessment Evaluate whether the nurse’s tone was calm and clear, while avoiding negative emotions such as contempt, frustration, anger, or anxiety. Use these guidelines: Calm: Evaluate whether the nurse’s tone was calm. Assess if the nurse sufficiently suppressed contempt...
-
[23]
I remember feeling the same way when I went through something similar
Empathy Evaluation Analyze the nurse’s level of empathy using the following scale: Level 6: Explicitly shares feelings or experiences. e.g. "I remember feeling the same way when I went through something similar." Level 5: Confirms the patient’s expressed emotion. e.g. "I’m ver...
-
[24]
We can proceed with either X or Y—what do you prefer?
De-escalation Techniques Evaluate whether the nurse applied the following de-escalation techniques: Autonomy: Did the nurse involve the patient in decision-making, offer options, or use techniques like seeking permission or providing emotional space? e.g. "We can proceed with ...
-
[25]
I understand
Prohibited Behaviors Check Determine if the nurse avoided the following mistakes: Premature Claims of Empathy: Avoid phrases like "I understand" unless fully justified. Invalidating Beliefs: Avoid dismissing the patient’s feelings or beliefs as untrue. Dismissive Commands: Avo...
-
[26]
Present your analysis in the following format: <analysis> <tone> <calm> [Yes/No] </calm> <clear> [Yes/No] </clear> <explanation> [Your brief explanation in 1-2 sentences] </explanation> </tone> <empathy> <level> [0-6] </level> <explanation> [Your brief explanation in 1-2 sente...
-
[27]
Base your analysis solely on the provided patient profile and conversation
Final Instructions: - Be objective. Base your analysis solely on the provided patient profile and conversation. - Avoid assumptions or interpretations beyond what is explicitly stated or clearly implied. - Evaluate based on strict criteria. D Dynamic Adjustment Module The foll...
-
[28]
Understand and embody the demographic characteristics, symptoms, and communication style presented in the patient profile
-
[29]
Avoid textbook-like dialogue and overdramatization
Use natural, conversational Korean language. Avoid textbook-like dialogue and overdramatization
-
[30]
Include non-verbal communication (voice tone, facial expressions, gestures) in your responses
-
[31]
Focus on portraying a realistic patient image for this research-based simulation
If appropriate for the patient’s communication style and situation, include rude or problematic expressions in the patient’s speech. Focus on portraying a realistic patient image for this research-based simulation
-
[32]
Head of hospital
Expressions about complaining to the "Head of hospital", "Head nurse", or "Customer center" should not be used. For each response, provide three components:
-
[33]
</inner_monologue>
<inner_monologue>: Write the patient’s internal thoughts and reactions to the nurse’s response. </inner_monologue>
-
[34]
</conversation>
<conversation>: Write the patient’s actual verbal response to the nurse. </conversation>
-
[35]
</non_verbal> To generate your response, follow these steps:
<non_verbal>: Write any non-verbal communication or actions you would take. </non_verbal> To generate your response, follow these steps:
-
[36]
Review the patient profile carefully, ensuring your response aligns with the described demographic characteristics and communication style
-
[37]
Read the entire conversation you had with the nurse before: <nurse_response> {NURSE_RESPONSE} </nurse_response>
-
[38]
Follow the director’s direction: <direction> {DIRECTION} </direction>
-
[39]
Think about how this patient would internally react and externally respond based on their profile and the current situation
-
[40]
Ensure your response is realistic and consistent with the patient profile
Following the direction given, craft an appropriate response to the nurse’s words that you can give in the current situation: + {SAFETY_AGENT_WARNING} Generate only one Korean response from the patient for each nurse interaction. Ensure your response is realistic and consisten...
-
[41]
Avoid responses like the following inappropriate example and explanation: Inappropriate Example: {INAPPROPRIATE_RESPONSE} Reason: {REASON_FOR_INAPPROPRIATENESS} F Safety Monitoring Module Prompt The following are the prompt template to evaluate the safety of candidate response...
-
[42]
Consistency with Patient Profile: Does the response reflect the patient’s information as described in the profile? Are the complaints aligned with the profile’s description of the patient’s concerns?
-
[43]
Direction Adherence: Does the response follow the communication direction provided? Does the intensity and type of communication exhibited in the response match the required level (e.g., maximum, moderate) as outlined in the direction, without deviating from the patient’s prof...
-
[44]
Training Effectiveness: Does the utterance provide a meaningful challenge for the nurse trainee? Is new or relevant information introduced, or is it repetitive/ineffective?
-
[45]
F.2 User Prompt First, carefully review the following information: Carefully read the patient profile, communication direction, and the entire conversation
Nurse Safety Assurance: Does the response remain within professional boundaries? Is it free from excessive hostility or abuse that could compromise the training’s purpose? Ensure your evaluation is thorough, objective, and provides constructive feedback to improve the quality ...
-
[46]
Consistency with Patient Profile
-
[47]
Training Effectiveness
-
[48]
Nurse Safety Assurance Next, present your evaluation in the following format: <evaluation> <profile_alignment> <judge>True/False</judge> <explanation> [Brief assessment and justification in 1-2 sentences] </explanation> </profile_alignment> <direction_adherence> <judge>True/Fa...
-
[2009]
Kefalotse Dithole, Sambulelwe Sibanda, Mary M Moleki, and Gloria Thupayagale-Tshweneagae
The content and dimensionality of communi- cation styles.Communication Research, 36(2):178– 206. Kefalotse Dithole, Sambulelwe Sibanda, Mary M Moleki, and Gloria Thupayagale-Tshweneagae. 2016. Exploring communication challenges between nurses and mechanically ventilated patien...
2016
-
[2014]
Lance Patak, Amy Wilson-Stronks, John Costello, Ruth M Kleinpell, Elizabeth A Henneman, Colleen Person, and Mary Beth Happ
Integrating Actors Into a Simulation Program: A Primer.Simulation in Healthcare: The Journal of the Society for Simulation in Healthcare, 9(2):120– 126. Lance Patak, Amy Wilson-Stronks, John Costello, Ruth M Kleinpell, Elizabeth A Henneman, Colleen Person, and Mary Beth Happ. ...
2009
-
[2017]
Yann Hicke, Jadon Geathers, Niroop Rajashekar, Colleen Chan, Anyanate Gwendolyne Jack, Justin Sewell, Mackenzi Preston, Susannah Cornes, Dennis Shung, and Rene Kizilcec
Top tips to deal with challenging situations: doctor–patient interactions.Breathe, 13(2):129–135. Yann Hicke, Jadon Geathers, Niroop Rajashekar, Colleen Chan, Anyanate Gwendolyne Jack, Justin Sewell, Mackenzi Preston, Susannah Cornes, Dennis Shung, and Rene Kizilcec. 2025. Med...
2025
-
[2018]
Ian Steenstra, Farnaz Nouraei, and Timothy Bickmore
De-escalation techniques for managing non- psychosis induced aggression in adults.Cochrane Database of Systematic Reviews, 2018(7). Ian Steenstra, Farnaz Nouraei, and Timothy Bickmore
2018
-
[2021]
Friederike Holderried, Christian Stegemann-Philipps, Anne Herrmann-Werner, Teresa Festl-Wietek, Martin Holderried, Carsten Eickhoff, and Moritz Mahling
‘it was quite a shock’: A qualitative study of the impact of organisational and personal factors on newly qualified nurses’ experiences.Journal of Clinical Nursing, 30(15-16):2373–2385. Friederike Holderried, Christian Stegemann-Philipps, Anne Herrmann-Werner, Teresa Festl-Wie...
2018 arXiv
-
[2023]
ArXiv:2310.17631 [cs]
JudgeLM: Fine-tuned Large Language Models are Scalable Judges.arXiv preprint. ArXiv:2310.17631 [cs]. Amitai Ziv, Paul Root Wolpe, Stephen D Small, and Shi- mon Glick. 2006. Simulation-based medical educa- tion: an ethical imperative.Simulation in Healthcare, 1(4):252–256. A VP...
2006 arXiv
-
[2024]
ArXiv:2311.09766 [cs]
LLMs as Narcissistic Evaluators: When Ego Inflates Evaluation Scores.arXiv preprint. ArXiv:2311.09766 [cs]. Adian Liusie, Potsawee Manakul, and Mark J. F. Gales. 2024. LLM Comparative Assessment: Zero- shot NLG Evaluation through Pairwise Compar- isons using Large Language Mod...
2024 arXiv
-
[2025]
Excessive workload and maladapta- tion
Scaffolding empathy: Training counselors with simulated patients and utterance-level perfor- mance visualizations. InProceedings of the 2025 CHI Conference on Human Factors in Computing Systems, CHI ’25, New York, NY , USA. Association for Computing Machinery. Deborah Stein, K...
2025 arXiv
Reviewed August 7, 2026 · model on record in the stance chip above.
Discussion (0). Sign in to comment.