LabaLab

Research Projects

Ongoing research lines in language, affect, and behaviour in AI: social robots, conversational agents, and the safety of language models in sensitive settings.

Social robots

Robots in homes, clinics, workplaces and welfare services, and what people do with them.

Robot as a Window: Social Robots Mediating Parent–Child Understanding in Autism Play

Make-believe play gives parents a window into how their child thinks and feels. For many autistic children, imagination is present but hard to show in everyday play, so that window can stay closed. This US–Israel project with the University of Tennessee, Knoxville develops social robots as mediators: a predictable, engaging play partner for the child that also helps parents notice the themes, emotions and intentions expressed in play. The work moves from participatory design with autism specialists and parents, to laboratory studies of how robot-mediated play unfolds and how parents perceive it, to home deployments with families in the United States and Israel.

Supported by the NSF–BSF program (US–Israel Binational Science Foundation).

Topics: Social Robotics, Autism, Play, Families, Participatory Design, Human-Robot Interaction

People: Dr. Amit Rogel (Postdoc); Itay Kohn (Masters Student); Shir Ben Amram (Masters Student); Noa Lehmann (Research Assistant); Dr. Samuel Olatunji (Collaborator, University of Tennessee, Knoxville); Dr. Rachael Bevill Burns (Collaborator, University of Tennessee, Knoxville)

Partner institutions: University of Tennessee, Knoxville

Funded by: U.S. National Science Foundation, United States–Israel Binational Science Foundation

Robotic Social Cognition Training for Autistic Adults

This project addresses a major barrier behind the employment gap for autistic adults: social communication challenges in job interviews and in everyday workplace interaction. It develops a socially assistive coaching intervention using the Furhat robot, designed to provide consistent, low pressure practice for skills like emotion recognition and perspective taking, alongside strategies for handling interview stress and workplace misunderstandings.

At the core of the system is an expressive cues engine that can control and vary the robot’s facial expressions, gaze, speech, and prosody. This makes it possible to structure training in a gradual, adaptive way. Sessions can begin with clear and exaggerated social signals, then progressively shift toward more natural and subtle cues as users gain confidence and accuracy. The training content is organized into two complementary modules. The perspective taking module uses short scenarios and structured prompts to support “perspective checks,” helping users reason about what another person might think or feel and how to respond appropriately. The emotion recognition module trains users to identify emotional states from facial and vocal signals, while also practising appropriate expression and conversational responses.

The project is developed and evaluated in collaboration with ALUT through two applied studies. One focuses on interview readiness, using baseline, post intervention, and follow up mock interviews to measure change over time. The second focuses on workplace integration, using a workplace adapted SSPA together with staff ratings of collaboration and communication quality. The overall goal is to deliver an intervention that is both scientifically grounded and practically deployable, with a pathway to scaling through open source release and adaptation to additional employment support settings.

Topics: Social Robotics, Autism, Emotion Recognition, Social Cognition, Affective Computing, Facial Expression, Human-Robot Interaction

People: Yuval Zohar (Masters Student); Yuval Zipora (Masters Student)

Partner institutions: ALUT

Funded by: ABC Robotics Initiative, Ben-Gurion University of the Negev

Robots in Welfare Services

Together with Israel's Ministry of Welfare and Social Affairs, we placed the same robot in seven very different services: day centres for older adults, homes for adults with intellectual disability, youth-protection facilities and social services. LabaLab studies what the robot became in each place, and why the same machine is indispensable in one service and a burden in another.

The programme began in 2023 with a steering committee, a shared vendor and five robots, then seven, across services that had almost nothing else in common. Each service named its own robot and decided what it would do: morning rounds, calling people to lunch, music and games, and later therapeutic work with clinicians. The pilot was judged a success, and in 2026 the programme is expanding to 80 locations with 100 robots, selected through a public call.

In autumn 2026 we travelled to Japan with the Ministry's leadership to see how Japan integrates social robots into welfare services, and which implementation models can work in Israel.

Topics: Social Robotics, Welfare Services, Field Deployment, Older Adults, Policy

Partner institutions: Ministry of Welfare and Social Affairs

Antagonistic AI

Most conversational AI is built to agree. This project studies the opposite on purpose: antagonistic behaviour such as dismissiveness, sarcasm, condescension and confrontation, which matters for training, rehearsal and stress-testing, and for understanding how people react when an agent disagrees with them.

On the language-model side, we built a reproducible way to generate and evaluate antagonism across many models, and examine how safety training quietly softens antagonistic replies and whether antagonism persists once it is trained in.

On the robot side, we built an operator console that puts a human between the language model and the robot. The operator sets the behaviour live, every reply waits in a review gate before the robot speaks it, and safety checks flag replies that are too harsh or too soft. The console runs on several social robots, so antagonistic robot behaviour can be studied with participants in a controlled, low-risk way.

Topics: Antagonistic AI, Large Language Models, AI Safety, Social Robotics, Human-Robot Interaction

Expressive Robots with Minimal Bodies

Not every social robot has an expressive face. This project studies how emotion comes across through minimal motion, using the Reachy Mini robot, and how people perceive it, alongside work on facial expressions on the Furhat robot.

Topics: Social Robotics, Emotion Expression, Human Perception

People: Dr. Amit Rogel (Postdoc)

Related publications:

  • Emotional Expression in Low-Degrees-of-Freedom Robots: Assessing Perception with Reachy Mini (2026)

Language models and conversational AI

How language models behave when people turn to them for care, advice and support, and how to keep that safe.

Conversational AI for Real-Time Clinical Case Formulation

This project builds a conversational AI platform for clinical measurement and monitoring that can be used alongside therapy. The motivation is that most validated mental health measures rely on retrospective self-report, are typically collected infrequently, and can be too taxing for vulnerable patients to complete regularly. As a result, clinicians often have limited visibility into day-to-day mechanisms of change, and research frequently captures only broad symptom change between intake and endpoint.

The platform focuses on quantifying clinical case formulation, meaning an individualized picture of how symptoms, beliefs, emotions, and behaviours interact to maintain distress and where intervention should focus. Instead of collecting symptom checklists through chat, it uses brief, targeted conversational interactions to identify clinically meaningful components and map their relationships over time. The design is guided by established cognitive models of PTSD, anxiety, and depression, with an initial focus on PTSD. The system adapts based on patient responses and clinician input, and it is designed to be highly explainable so each coded component and formulation can be traced back to specific conversational moments.

Evaluation is staged. First, the conversationally derived outputs are validated against traditional momentary assessments and reviewed by experienced clinicians, with iterative refinement until the tool reliably captures clinically meaningful information. Next, clinical usefulness is tested through patient feedback on user experience and clinician agreement with the system’s conclusions. Finally, randomized trials assess whether using the tool improves formulation quality, treatment planning, and outcomes compared to usual care without the tool. Ethics and safety are integrated throughout via involvement of people with lived experience in design and piloting, close clinical oversight during early testing, and a deployment approach intended to support therapy without replacing clinician judgement.

Topics: Generative AI, Clinical Measurement, Case Formulation, Conversational AI, PTSD, Human-AI Interaction, Human-Agent Interaction

People: Amit Oren (Masters Student); Dr. Dean Ariel (Collaborator, Clalit Health Services); Dr. Nimrod Hertz (Collaborator, University of Cambridge)

Partner institutions: MRC-CBU University of Cambridge

Related publications:

  • Generating Clinical Vignettes that Preserve Cognitive Formulations (2026)
  • Towards a CONCORD Between Patient Perspective and Ecological Case Formulation: Conversational Large Language Models for Personalised Mental Health (2026)

Psychosocial Safety for LLM Agents

A unified research and tooling direction that combines stable persona-driven dialogue simulation with multi-agent psychosocial safety evaluation, enabling realistic stress-testing and safer deployment of LLMs in sensitive conversational settings.

Topics: LLM Safety, Multi-Agent Evaluation, Dialogue Simulation, Human-AI Interaction, Human-Agent Interaction, Generative AI, Conversational AI

People: Han Luo (Alumni)

Related publications:

  • SPASM: Stable Persona-driven Agent Simulation for Multi-turn Dialogue Generation (2026)
  • DialogGuard: Multi-Agent Psychosocial Safety Evaluation Interface of Sensitive LLM Responses (2026)

Adaptation and Alignment in Communication between Autistic Adults and AI Agents

Autistic adults increasingly turn to AI agents for communication, everyday tasks and emotional support. Many find them easier to talk to than people: patient, available and free of social judgement. Yet these agents are trained and aligned on data that reflect majority, neurotypical ways of talking, so their default replies can cost effort and cause misunderstanding.

Framed through the double empathy problem, which treats autistic–neurotypical communication difficulty as a mismatch between two communication styles rather than a deficit on one side, this project studies both sides of that mismatch. It follows autistic and neurotypical adults as they use AI agents in daily life over several weeks, rating their interactions in the moment, and builds an open corpus annotated with autistic raters. From it we derive communication profiles, design requirements and an autistic-informed benchmark for AI agents, then test experimentally whether agents aligned with autistic preferences lift the burden, for whom, and whether it matters who does the adapting.

The work is carried out with autistic researchers and raters, and community partners.

Topics: Autism, Large Language Models, Human-AI Interaction, Double Empathy, Accessibility

People: Keren Orzech (Masters Student)

Evaluating Language Models for Health and Life Advice

People increasingly ask chatbots for medical and personal advice. This line of work measures where that goes wrong: models that recommend unnecessary care or hold back necessary care, safety that drops outside English, empathic replies that stand in for actually helping, and structured clinical tasks that simpler decision models handle more reliably. Related outputs include Annals of Internal Medicine (2026), Communications Health (2026), NeurIPS 2026, and two medRxiv preprints.

Topics: Large Language Models, Clinical AI, AI Safety, Multilingual Evaluation

People: Dr. Dean Ariel (Collaborator, Clalit Health Services)

Related publications:

  • Overuse and False Restraint in Large Language Model Recommendations for Choosing Wisely Scenarios (2026)
  • When Empathy Misses the Goal: A Benchmark for Goal Displacement in LLM Advice (2026)
  • Linguistic inequity in clinical AI evaluation: a scoping review of 806 studies across 62 languages (2026)
  • Decision models can offload structured clinical tasks from large language models (2026)
  • Consumer chatbots gave similarly empathic answers whether safe or unsafe: a physician-rated evaluation in six languages (2026)

When Words and Voice Disagree: Emotion in Audio Language Models

Audio-capable language models hear both the words and the voice. When the two conflict, such as a cheerful sentence said sadly, current evaluations break down. This project introduces an incongruence-aware, threshold-free way to evaluate how these models handle that conflict.

Topics: Affective Computing, Speech, Multimodal LLMs

Related publications:

  • Incongruence-Aware Evaluation of Multimodal Speech Emotion Recognition in Audio Large Language Models (2026)

Lexi Takes the Lead: Affective Interventions by a Proactive Conversational Agent

Most conversational AI waits to be asked. An agent that starts conversations on its own, and remembers what people told it, could build closeness, or could feel intrusive and push people away. We call this tension the reactance trade-off.

Lexi is a language-model chat companion delivered as an Android app. In a longitudinal field study, participants talk with Lexi in one of three conditions that differ only in how the agent makes contact: it never reaches out, it sends neutral invitations to chat, or it sends warm, personal check-ins that follow up on things the participant shared earlier. The agent's persona, its memory and the timing of its messages are the same in every condition. The study combines logged behaviour (replies to the agent's messages, how often and how long people talk, what they disclose) with in-app questionnaires on self-disclosure, privacy and surveillance concerns, reactance, mind perception and loneliness.

Co-supervised by Dr. Tal Golan.

Topics: Conversational AI, Large Language Models, Proactive Agents, Self-Disclosure, Privacy, Field Study

People: Itay Kohn (Masters Student)

AUT-Guard: Expert-Grounded AI for Autism Inclusion and Misinformation Response

AUT-Guard is a cohesive research project led by LabaLab (Ben-Gurion University of the Negev) in partnership with ALUT that integrates two complementary aims: strengthening inclusion for autistic youth in everyday community settings and countering autism- and disability-related misinformation where it spreads online. The project develops a modular, expert-grounded architecture that transforms curated clinical expertise and lived experience into adaptive, context-aware responses, using retrieval-augmented generation to produce practical guidance that can be tuned in tone and detail for different audiences and situations. In its “offline” deployment, AUT-Guard serves as a mobile conversational partner for youth movement guides, providing just-in-time support for challenges such as sensory sensitivities, social misunderstandings, routine disruptions, and questions from children, parents, or staff; this deployment is evaluated through participatory co-design followed by a longitudinal field study that tracks changes in guides’ self-efficacy, autism knowledge, coping and responsiveness, and inclusion climate alongside qualitative interviews and real-world usage patterns. In its “online” deployment, the same architecture powers an ALUT-facing workspace for responding to misinformation, enabling staff (and partner organisations) to paste or flag questionable content, receive grounded draft replies, and refine them for publication as comments, posts, or embedded website chatbots. By studying both deployments together, AUT-Guard advances a unified model of responsible AI support that links micro-level, moment-to-moment inclusion work with macro-level public communication, while generating empirical evidence about when people seek AI guidance, what support they find useful, and how expert oversight can be maintained in sensitive domains.

Topics: Autism, Neurodiversity, Inclusion, RAG, Support, Youth, Community, Misinformation, Disability

Partner institutions: ALUT, Shamoon College of Engineering