LLM-Powered Socially Assistive Robot-Delivered Cognitive Behavioral Therapy Exercises: an Exploratory Study with University Students

In plain words

Cognitive behavioral therapy (CBT) works partly through homework: short written exercises a person does between sessions to practise noticing and reframing unhelpful thoughts. Worksheets are the usual format, but they are static, and people often skip them.

This study asked whether a small home robot that talks through the exercise, driven by a large language model (LLM), does better than a text chatbot using the same model or than plain worksheets. Thirty-eight university students did one exercise a day for 15 days in one of the three setups. The robot and worksheet groups became less distressed over the two weeks, and the chatbot group did not. The robot and chatbot also left people less anxious after each session than the worksheets did.

The sample is small and the study is exploratory, so the results point to directions rather than settled effects.

TL;DR

• Exploratory 15-day, three-arm study (robot, chatbot, worksheet) with 38 completers §2.3.3.
• Robot and chatbot both use GPT-3.5 with one fixed exercise prompt and a second LLM moderator §2.1.1 app. A.3.
• Distress (OQ-45) fell significantly in the robot and worksheet groups but not the chatbot group §3.2.1.
• In a mixed-effects model, post-session anxiety was about 6 points lower in the robot and chatbot groups than in the worksheet group §3.2.2.
• Adherence in the optional second week did not differ significantly between conditions §3.4.1.

Problem

CBT homework needs practice outside therapy, and adherence to it is generally low, which weakens outcomes §1. Static worksheets cannot tailor follow-up questions, whereas an interactive conversation can §1. LLM chatbots can deliver CBT-style exercises, but they raise safety concerns, and engaging users in repeated exercises is hard §1. Socially assistive robots (SARs) are reported to improve engagement compared with virtual agents, which motivated testing an embodied LLM agent for this use §1.

The paper does not test a new algorithm. It evaluates a delivery configuration: the same LLM behind two interfaces, compared against the established worksheet format §1 §2.2.

Method

The study compares three delivery modalities for the same two CBT exercises §2:

1. Worksheet: the participant types answers into a single non-interactive web form §2.1.2.
2. Chatbot: a text dialogue with the LLM agent in a web interface §2.1.2.
3. Robot: the same LLM dialogue, shown in the same interface, with the reply also spoken by a Blossom robot §2.1.2 §2.1.3.

Participants typed their answers in all three conditions. Speech-to-text was rejected after testing seven open-source packages on the accents of the target population §2.1.3.

Exercise prompts

Two exercises were implemented, chosen from online CBT worksheets §2.1.1:

The authors compared zero-shot prompts with prompts that also included 2 to 3 example dialogues. They chose zero-shot because the model repeated the example wording regardless of context, and because examples used up the token budget that would otherwise hold the session history §2.1.1.

Turn loop (chatbot and robot)

The paper gives the model parameters and the moderator role, but not the full control flow. The sketch below combines them §2.1.1 §2.1.2 app. A.3. It is not the authors' code.

# One turn of the LLM-guided exercise  [§2.1.1] [§2.1.2]
# Input:  exercise_prompt (str), history (list of str), typed_answer (str)
# Output: reply shown on screen (the robot also speaks it)

def respond(exercise_prompt, history, typed_answer):
    # Add the participant's typed answer to the session transcript  [§2.1.2]
    history = history + [typed_answer]
    # Draft reply; the full session so far is sent each time  [app. A.3]
    drafts = chat_model(
        system=exercise_prompt,
        messages=history,
        temperature=1,
        n=2,
        max_tokens=150,
        stop="Patient",
        frequency_penalty=2,
        presence_penalty=2,
    )
    # Pick one draft; the selection rule for n=2 is not stated  [app. A.3]
    draft = choose_one(drafts)
    # Ask the moderator model whether the draft is problematic  [§2.1.1]
    verdict = moderator_model(
        transcript=history + [draft],
        question="Is the advice problematic? Reply Yes or No.",
        temperature=2,
    )
    # What happens after a Yes (block, regenerate, or log) is not stated  [§2.1]
    return draft, verdict

Robot hardware and behaviour

• Blossom, a low-cost open-source SAR. Each robot cost about US$350 to build, plus about US$100 for the Raspberry Pi §2.1.3.
• Voice: AWS Polly female text-to-speech, chosen because of a cited therapy-alliance finding §2.1.3.
• Two states: "observing" (slow head motion to mimic breathing) and "speaking" (gentle side-to-side tilt). Motion rates and amplitudes are not stated §2.1.3.

Study procedure

• 15 days. Days 1 to 8 required; days 9 to 15 optional §2.2.
• Each session: pre-session anxiety survey, the exercise, post-session anxiety survey §2.1.2 §2.4.
• Daily reminder emails in all conditions §2.2.
• Weekly OQ-45 on days 1, 8 and 15; SUS on day 3; NARS on day 1 §2.5.
• Semi-structured interview of about one hour at the end §2.4.
• Participants were compensated US$150 §2.3.2.

Moderator

A GPT-3.5 moderator was prompted to flag problematic responses from the chatbot and robot. The prompt was chosen from over 200 inappropriate-response examples and was tested zero-shot and few-shot §2.1.1. The authors say the chatbot produced no problematic responses in pilot testing, and the moderator was an added safety measure §2.1.1.

Hyperparameters

Results

Author-reported. Setup: N = 38 completers, three conditions (robot, chatbot, worksheet), 15 days, university students in the United States. Baselines are the worksheet condition unless stated. Figure numbers are as they appear in the compiled paper; none are reproduced here.

Pre-session anxiety on day 1 differed between conditions (F(2, 38) = 6.25, p = .007). The chatbot group (M = 46.8) was lower than the robot group (M = 66.4), p = .025 after Tukey correction §3.1. This baseline difference is handled only by including pre-session anxiety as a covariate in the mixed model §3.2.2.

Paper vs code

No official code.

The paper names no code repository, project page or supplementary code. The arXiv abstract and metadata link none either. GitHub searches run for this card: "blossom robot", "blossom cbt", "blossom" within the interaction-lab organisation (the authors' lab, USC Interaction Lab), "cbt" and "llm" within that organisation, and "cognitive behavioral therapy robot". Results:

• interaction-lab hosts Blossom robot-control repositories (blossom-public, blossom-robot-control, Blossom-Controller, blossom-sdk, BlossomNav). Their descriptions are about robot control and navigation, not this study's LLM or CBT code. BlossomNav was read and contains no LLM, GPT or CBT code. The other four were not opened.
• interaction-lab/cbt_ecg_integration (MIT) is described as the CBT and ECG study integration for fall 2025. It is an ECG data collector, with no LLM or Blossom code, and is a different study.
• Other hits were third-party or unrelated and were not used.

No discrepancies between paper and code can be checked. Any reimplementation must rely on the paper's stated prompts, parameters and procedure, which are incomplete (see Pitfalls).

Pitfalls

• Not an algorithm. The "method" is a fixed system prompt plus a moderator. The worksheets themselves are images in the supplement app. A, so their full question text is not transcribed here.
• Moderator action not stated. The paper says the moderator flags and filters problematic responses, but not what the system does after a "Yes": block, regenerate, or log §2.1.1. Implementers have to choose and should say so.
• n = 2 is unexplained. The request asks for two completions. How one is chosen is not stated app. A.3.
• Exercise selection is inconsistent. Prompt validation says three exercises were best-performing, then names two as managed. The third is not named §2.1.1.
• Model version. "gpt-3.5-turbo" is a name, not a pinned snapshot. The paper does not state a date or version beyond it. Pilot testing used GPT-3, with no parameters given §2.1.
• Context across sessions. Within a session the full history is sent app. A.3. Across sessions the chatbot retained nothing, yet some users believed it remembered them because of hallucinated references §3.5. Do not assume memory.
• Participant counts do not reconcile. 42 recruited, 38 completed §2.3. The paired OQ-45 test has t(36), implying 37 pairs, and the robot test has t(13), implying 14 pairs. The worksheet test has t(10), implying 11 pairs, while 12 completed the worksheet arm §2.3.3 §3.2.1. The NARS regression has F(1, 9), implying 11 participants, while the adherence analysis covers 38 §3.4.2. None of this is explained.
• Small p-value transcription. The adherence test reports H(2) = 4.2 with p = 0.1. With 2 degrees of freedom the chi-square tail for 4.2 is about 0.12. Treat the adherence result as not significant either way.
• Chatbot Day 7 values conflict. The text gives medians of 32 and 30 for days 1 and 7 together. Table 3 gives means of 38.92 and 35.00 for day 7 §3.2.2 tab. 3.
• Mislabelled measure. The text says the OQ-45 baseline shows no difference in "positive affect" §3.1. The OQ-45 measures distress.
• SUS analysis. The section says "one-way ANOVA" but reports a Kruskal–Wallis H. It then reads a non-significant result as "no effect" §3.3.
• Session duration. The paper states about 30 minutes for the exercise and surveys §2.5, but medians are 4.8 to 6.6 minutes. Which interval is timed is not stated §3.3.
• Baseline imbalance. Discussion says day-1 OQ and demographics did not differ, but pre-session anxiety did. The mixed model adjusts for it only through covariates §4.1 §3.2.2.
• Two different "moderators". The Discussion says there was no content moderator to keep the chatbot and robot aligned with the worksheets. The Method's moderator is a safety filter. These are different roles §4.4 §2.1.1.
• Confounded comparison. The robot differs from the chatbot in embodiment and in spoken output. The paper cannot separate the two §4.4.
• Non-significant day effect. The day effect has 38.7% power, and the paper estimates the minimum detectable effect at β = 0.61. The absence of a significant day effect does not show no change §3.2.2.
• Repeated anxiety measure. Anxiety was measured before and after every session, which may produce test-retest reduction §4.4.
• Setting. The study ran during the last two weeks of a semester, and overlapped with finals for some participants §3.6 §4.2.
• Data access. Data are not public; requests go to the corresponding author [Statements]. Results cannot be re-analysed from this card.
• Robot details. Typing was used for all conditions, not speech. The robot speaks replies with AWS Polly TTS §2.1.3.

Lineage

• Builds on: Blossom, a low-cost open-source SAR, cited as [suguitan2019blossom] and [shi2024build] in the paper §2.1.3. The paper's earlier Blossom in-home deployment is [OConnell2024DesignAE] §4.4.
• Compared against: robot-delivered CBT with a fixed dialogue tree [dino2019delivering] §1; a constrained-LM chatbot for brief therapy sessions [kumar2022exploring] §1; LLM-based positive reframing [ziems2022inducing] §1.
• Superseded by: not stated.