Scientific Papers
Published research and the scientific foundations of the method.
Publications
Five publications since 2017, including research with Università della Svizzera italiana and Warsaw University of Technology, published through IATED, Springer Nature and Wiley.
Research on practice and transfer
Does practice in a behavioral simulator improve performance in situations never practiced?
- 2017· Peer-reviewed conference paperSpringer
A LifeLike experience to train user requirements elicitation skills
S. De Ascaniis, L. Cantoni (Università della Svizzera italiana), E. Sutinen, R. Talling (University of Turku)
Design, User Experience, and Usability, HCI International 2017 — Springer, LNCS vol. 10290, pp. 219–237
Controlled experiment with 69 Bachelor students at Università della Svizzera italiana (USI), randomly split into a practice group and a control group, using a Lifelike behavioural simulator developed with the USI eLearning Lab.
Key finding: After three weeks of practice the practice group's median performance rose from 51% to 59.5% (+8.5 points), against +3 points for the control group. In a listening test on an unseen conversation, the practice group scored a median of 6.6/10 against 4.9/10.
Why it matters: Structured practice improved performance more than exposure alone (+8.5 vs +3 points), and the gain showed up in conversations never practiced. Quantitative evidence that you get better by practicing, not by watching.
Read publication - 2017· Peer-reviewed journal articleWiley / ASME
Virtual patients in problem-based learning
H. Hamdy, A. Al-Moslih, G. Tavarnesi, A. Laus
Medical Education (Wiley / ASME)
Describes AI-driven virtual patient simulations, with professional actors and real-time branching decisions, used to train medical students' reasoning and communication skills.
Key finding: Students and faculty reported higher focus, accountability and involvement in managing the patient's problem.
Why it matters: Students and faculty practicing with virtual patients reported more focus, accountability and involvement in the case. Independent medical-education research confirming how learners respond to realistic simulated conversations.
Read publication
Research on the method
How the Skillgym method is designed: from competency model to observable behaviors, motivation, and what learners perceive.
- 2022· Conference paper
Motivation Strategies to Keep Trainees Engaged
E. Ciani, A. Laus
ICERI2022 — 15th International Conference of Education, Research and Innovation (IATED), pp. 491–499
How the Skillgym method embeds motivation and engagement theory into practice: multichannel feedback, challenge balanced against skill, and realistic, relevant scenarios.
Key finding: Challenge calibrated to the learner's skill, multichannel feedback and realistic scenarios are the levers that sustain engagement over repeated practice sessions.
Why it matters: Without engagement there is no repetition, and without repetition there is no behaviour change — motivation design is not a feature of the training, it is the training.
Read publication - 2022· Peer-reviewed conference paperSpringer
An Actionable Training: From Competency Model to Observable Behaviors While Empowering Communication
E. Ciani, A. Laus
The Learning Ideas Conference 2022 — Springer, LNNS vol. 581, pp. 104–113
Describes the architecture that turns a competency model into measurable, observable behaviours trained through AI-driven role play.
Key finding: The paper defines the mapping from a competency model to observable, measurable behaviours that AI-driven role play can train and assess.
Why it matters: This is how Skillgym turns your competency framework into practice and debrief criteria.
Read publication
Research on AI analysis
How rigorous definitions make a language model's analysis more accurate. This is research on analysis quality, not a training outcome.
- 2026· Peer-reviewed conference paperSpringer Nature
A Harm-Centric Operationalisation of Hateful Speech for Moderation and Education
E. Gajewska, M. Kulik, K. Budzynska, Ł. Kobyliński, A. Laus, B. Skowron
Advances in Computational Collective Intelligence, ICCCI 2026 — Springer Nature, CCIS vol. 3043, pp. 229–242
Proposes defining hateful speech by the unfair harm it causes rather than by surface-level keywords, and tests the definition with human annotators and a large language model (Gemini 2.0 Flash). Joint work with Warsaw University of Technology.
Key finding: Prompted with the harm-based definition, the language model reached an F1 score of 0.87 overall and 0.88 on the hate class, outperforming every conventional definition compared. Human annotators were also more accurate and more consistent.
Why it matters: A precise definition of the behaviour makes AI language analysis measurably more accurate. The same discipline Skillgym applies to conversation: define what to observe, then measure it.
Read publication
Program results, measured in client programs, are reported in Case Studies, with their limits. See Case Studies
Patent
The method behind Skillgym's simulations is protected by an Italian industrial invention patent, granted to LIFELIKE S.A. on 14 June 2023. It is evidence of proprietary innovation — separate from the publications above, which are evidence of research.
What the patent covers
A method for simulating interpersonal interactions: the simulation is driven by structured textual prompts, the learner's speech is transcribed and compared semantically, and the resulting classification and feedback steer the progression of the conversation.
Patent record
- Italian industrial invention patent, granted by the Italian Patent and Trademark Office (UIBM)
- Application 102021000011675, filed 6 May 2021
- Granted 14 June 2023
- Official title: “Metodo per la simulazione di interazioni interpersonali” (Method for simulating interpersonal interactions)
- Owner: LIFELIKE S.A. (100%)
- Classification G09B — education and training apparatus
What the research investigated
Two different questions, two different studies.
Does practice transfer to unfamiliar situations?
Whether practice with a behavioural simulator improves performance, and shows up when listening to a new, never-practiced conversation. (USI Lugano study.)
Can a better definition make AI analysis more accurate?
Whether a language model guided by a rigorous, harm-based definition classifies human language more accurately than with conventional definitions. (Springer Nature ICCCI 2026, with Warsaw University of Technology.)