Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

Integrating Reinforcement Learning and AI Agents for Adaptive Robotic Interaction and Assistance in Dementia Care

T0 review · 4 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A Q-learning caregiver trained inside an LLM-driven simulator learns to select assistance actions for people with dementia that score roughly twice the reward of random action selection in a shopping task.

desk verdict A useful simulation-only proof of concept whose headline claim overreaches; the 140-vs-70 result is mostly internal consistency of hand-set model parameters, but the paper is honest about its limits and deserves a serious referee. read the letter →

arxiv 2501.17206 v1 pith:DEYQLGJC submitted 2025-01-28 cs.AI cs.RO

Fengpei Yuan , Nehal Hasnaeen , Ran Zhang , Bryce Bible , Joseph Riley Taylor , Hairong Qi , Fenghui Yao , Xiaopeng Zhao This is my paper · ORCID
classification cs.AIcs.RO
keywords dementiacaresociallyassistiveroboticsreinforcementlearninglargelanguagemodelssimulationactivitiesofdailylivingQ-learninghuman-robotinteraction
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that a fully simulated environment can stand in for the scarce real-world data on how people with dementia interact with caregivers, and that an AI caregiver trained in that environment learns useful assistance strategies. The simulator combines a probabilistic model of four cognitive and emotional states (forgetfulness, confusion, anger, disengagement) with a large language model that generates realistic verbal and nonverbal behavior. A Q-learning agent then chooses among four assistance levels, and in the shopping scenario the learned policies reach an average return near 140, compared with roughly 70 for random action selection. The authors argue this shows the feasibility of LLM-enhanced reinforcement learning for adaptive, personalized dementia care, and that the same decision module works for computer-based agents as well as robots. If correct, this would give researchers a low-cost way to develop and test assistive AI before deployment.

What carries the argument

The machinery is a closed loop of four components: a probabilistic Markov-chain model with hand-set transition probabilities that turns the four statuses forgetfulness, confusion, anger, and disengagement into binary states; an LLM-based behavior module that renders those states as realistic verbal and nonverbal behavior from a person with dementia; an LLM-based perception module that reads that behavior back into a perceived state vector; and a Q-learning action-selection module over a 16-state, four-action Markov decision process. The reward function in Equation 1, with weights in Table 4, encodes two caregiving principles: keep the person in a positive affective state, and provide the minimum assistance that completes the task, with heavier penalties for anger and for directive assistance. The action execution module then converts the selected abstract action into a robot utterance. The Q-learning agent carries the argument because its learned policies are the evidence that the simulator is trainable and that the assistance decisions are sensible.

What would settle it

Record real PLWD-caregiver interactions during a shopping task, extract the actual probabilities of transitioning among forgetfulness, confusion, anger, and disengagement under each assistance type, then re-run the Q-learning policy; if the optimal actions under measured probabilities differ substantially from the simulated policy, the central claim fails. Alternatively, have dementia-care clinicians rate the generated robot utterances and simulated PLWD behaviors against video of real interactions, since a systematic mismatch would falsify the claim that the simulation produces realistic, tailored caregiving strategies.

Watch

Extended reading notes

Core claim

The paper's central claim is that an RL-based caregiver agent can be trained entirely inside a simulator, meaning a Markov-chain model of a person with dementia's cognitive and affective states coupled to an LLM that generates the person's behavior, and that the resulting Q-learning policy chooses assistance actions that substantially outperform random action selection, with average returns of approximately 140 versus 70 in the shopping scenario. The learned policies also appear clinically plausible: anger triggers supportive encouragement, combined forgetfulness and confusion trigger directive instruction, and confusion alone triggers non-directive cues. The authors further claim the framework extends to computer-based agents and that the generated LLM utterances in illustrative interactions are contextually appropriate and empathetic, while noting that the evaluation is based on RL performance metrics and the authors' own observations rather than external validation.

Load-bearing premise

The load-bearing premise is that the hand-set transition probabilities and reward weights, chosen by the authors from qualitative clinical observation rather than measured data, approximate how real people with dementia behave and what actually helps them; if they do not, the learned policies and returns are artifacts of the authors' modeling choices.

Editorial extensions

If this is right

  • If the paper is right, a Q-learning policy trained solely in this simulator chooses assistance actions with roughly twice the reward of random action selection in the shopping scenario.
  • The learned policies recommend specific assistance by state: supportive verbal assistance when the person is angry, directive verbal instruction when forgetful and confused, and non-directive cues when only confused, which suggests the reward function yields clinically plausible behavior.
  • The framework transfers to computer-based agents, so the same RL decision module could drive smartphone or desktop assistance, not only humanoid robots.
  • The open-source simulator gives other RL researchers a source of synthetic behavioral data for dementia-care tasks, addressing the scarcity of real-world interaction datasets.
  • The authors conclude that the RL/LLM combination demonstrates feasibility for adaptive assistance, while explicitly cautioning that higher RL returns do not by themselves guarantee better real-world caregiving outcomes.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A testable extension the paper leaves implicit: the same simulator could generate training data for policies that then get fine-tuned on a small amount of real PLWD-caregiver interaction, treating the simulator as a pretraining environment rather than a final validation.
  • Because the perception module misread uncertainty as forgetfulness, adding a task-procedure state to the observation space could reduce such errors; the paper notes the problem but does not propose this specific fix.
  • The reward function's weights are the real policy objective; if clinicians were asked to tune those weights, the learned behaviors would likely shift, so an empirical elicitation of caregiver preferences would be a natural next step.
  • If the simulator is valid, the most direct societal payoff is lower-cost development of assistive agents for dementia care, but that payoff depends on the real-world validation the paper explicitly defers.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes a simulation framework for dementia care in which a person living with dementia (PLWD) is modeled by a hand-specified Markov chain over four binary cognitive/affective states, with an LLM generating verbal and nonverbal behavior descriptions, while a robot caregiver uses Q-learning over a 16-state, 4-action MDP to select assistance types and an LLM to execute verbal actions. The authors report that the learned policies achieve an average return of roughly 140 versus about 70 for random action selection in a shopping scenario, and they illustrate qualitative interactions between the two agents. The paper claims that the RL system, enhanced by LLMs, effectively interprets and responds to the complex needs of PLWDs, while acknowledging several limitations including the lack of real-world validation and the reliance on author observation.

Significance. If the framework is taken as a simulation platform for generating hypotheses and for developing RL-based assistive strategies in a data-scarce domain, it has potential value: the architecture is clearly described, the state/action/reward structure is explicit, and the authors have made the simulator open-source. The paper also honestly documents limitations and future validation plans. However, the central evaluative claim—that the system effectively interprets and responds to the needs of PLWDs—is not supported by the evidence presented. The reported performance gap is an internal-consistency result within a simulator whose parameters were authored by the same team, and the qualitative evaluation is based on the authors' own impressions. The significance of the paper therefore depends on reframing it as a simulation and framework contribution rather than as an empirical demonstration of effective caregiving.

major comments (4)
  1. [Section 4.1 and Tables 3–4] The headline result (average return of about 140 vs. about 70) is not evidence of effective caregiving, because the transition probabilities and rewards are authored by the same team and are aligned with the conclusion. Table 3 sets directive assistance to reduce Forgetfulness/Confusion persistence to 5% and supportive assistance to reduce Anger/Disengagement persistence to 5%, while Table 4 assigns penalties to negative states and large rewards to subtask/task completion. A Q-learning agent reading these tables through simulation will naturally select the assistance types the authors already assumed to be effective. The 140-vs-70 gap therefore mostly confirms internal consistency between the assumed dynamics and the reward function; it says little about whether a real PLWD would benefit. Without sensitivity analysis over the hand-set probabilities/rewards or external validation, this result cannot support the abstract's claim of effective interpretation and response.
  2. [Sections 3.2.5 and 4.1] The final-policy selection procedure introduces selection bias. The paper states that the top five most frequent policies from the last 100 episodes were run for 10,000 experiments and the maximum-return policy was chosen, and Section 4.1 further reports evaluating policies with 1,000 experiments and selecting the one with the highest average return (144.3 vs. 133.5 and 139.2). Selecting the maximum from many stochastic replications inflates the expected return of the chosen policy, so the reported differences among candidate policies are not statistically meaningful without correction for multiple comparisons or at least error bars. The comparison between RL and random action selection (140 vs. 70) may still be robust, but the finer policy comparisons used to justify the final policy are not.
  3. [Section 4.2] The qualitative evaluation of the interactions relies entirely on the authors' own impressions ('Observations by the authors ... indicate that the interactions ... appeared reasonable'), and the authors explicitly acknowledge 'the limitation of relying on internal observations without formal evaluation.' The example in Fig. 11 actually shows the robot misperceiving the PLWD's uncertainty as forgetfulness, which undermines the claim that the system 'effectively interprets' PLWD behavior. The paper does not report any formal evaluation of the perception module's accuracy, nor any inter-rater or expert assessment of the generated interactions. At best, Section 4.2 provides illustrative examples of the framework's behavior, not evidence of effective interpretation and response.
  4. [Section 4.3 vs. Abstract/Conclusion] The paper's own limitation statement in Section 4.3 says 'Improvements in RL returns do not necessarily translate to enhanced caregiving outcomes for PLWDs,' and the section lists the lack of validation from external stakeholders and end-users as a key limitation. This directly contradicts the abstract's claim that the results 'demonstrate that the RL system, enhanced by LLMs, effectively interprets and responds to the complex needs of PLWDs, providing tailored caregiving strategies.' The conclusions should be scaled back to claims about the feasibility of the simulation framework, or the paper should present actual external validation.
minor comments (5)
  1. [Section 1] There is a typo in the Introduction: 'enhancing the qualify of life' should be 'quality of life.'
  2. [Table 3] The layout of Table 3 is difficult to parse: the base probabilities and the matrix of mutual influences are presented in the same table without clear subheadings, and the '+' signs on some entries are ambiguous (e.g., '+5%' suggests an additive effect, but the base probability is a baseline rate). Clarify the exact generative rule for combining base probabilities and mutual influences.
  3. [Equation (1)] The subscripts in Equation (1) are inconsistently formatted (e.g., 'T imestep', 'F lagsubtaskComplet', 'wsubtaskSkip') and should be cleaned up for readability.
  4. [General] The abstract and Section 1 state that the simulator is open-source, but no repository link or availability statement is provided in the manuscript; please add the link or state where the code can be obtained.
  5. [Figure 4] In Figure 4, the word 'Forgetfulness' appears with a formatting artifact (broken text); please fix the figure rendering.

Circularity Check

2 steps flagged · score 6.0 of 10

Main 'effectiveness' result is generated inside the authors' own simulator: the learned policy and the 140-vs-70 return gap follow from hand-set transition probabilities and reward weights, and the final policy is selected as the best of several stochastic runs.

  1. self definitional [Sections 2.1.2, 3.2.3 (Eq. 1, Table 4), and 4.1]
    "Regarding the statuses Forgetfulness and Confusion, if their current status is No, the probability of following status being Yes will be 0% and 0%, correspondingly. Comparatively, if their current status is Yes, the probability of following status being Yes will be 40% and 60%, correspondingly, given verbal non-directive assistance, and 5% and 5%, correspondingly, given verbal directive assistance."

    These hand-set probabilities make directive assistance almost certainly clear Forgetfulness/Confusion, while Table 4 penalizes negative states (forgetfulness -1, confusion -1, anger -5) and rewards subtask completion +50. A Q-learner trained on this environment therefore 'discovers' exactly the policy Section 4.1 reports, e.g., directive assistance in [YFor, YCon, NAng, NDis], and the ~140-vs-70 return gap is a consequence of the authors' chosen dynamics and reward, not independent evidence about real PLWDs. The conclusion that the RL system effectively responds to PLWD needs is equivalent to the reward/dynamics assumptions used to define 'effective.'

  2. fitted input called prediction [Sections 3.2.5 and 4.1 (final policy selection and reported returns)]
    "we record all the optimal policies suggested in the last 100 episodes and use the top five policies most frequently suggested to run 10000 experiments and choose the policy which enables the maximum return as the final policy... These policies were evaluated by running 1,000 experiments... the policies... achieved average returns of 144.3... Thus, the final optimal policy, π′∗, is the one recommending verbal non-directive assistance (a2)... These findings validate the effectiveness of the RL approach in selecting optimal, personalized strategies for assisting PLWDs."

    The policy is explicitly selected as the one with maximum return on the evaluation runs, and the same selected maximum (144.3) is then cited as evidence that the final policy is optimal. Because the evaluation metric is the selection criterion, the reported return is a maximum over candidate policies under stochastic transitions, not an independent held-out estimate; quoting it as 'validation of the effectiveness of the RL approach' is a fitted-selection-called-prediction. This inflates the headline performance beyond what a single pre-specified policy would be expected to achieve.

full rationale

Score 6 because the central quantitative claim is not independent of the environment in which it is produced: the transition probabilities and reward weights are authored by the same team, and the final policy's reported return is the maximum of evaluated candidates in that same simulator. This is a partial circularity: the 'effectiveness' of the learned assistance is defined, not measured, by the reward function and dynamics. No load-bearing self-citation chain is present: PASS [21] is external, and the self-citation [24] only supplies qualitative background for the hand-set probabilities. The paper's own Section 4.3 concedes 'Improvements in RL returns do not necessarily translate to enhanced caregiving outcomes for PLWDs,' which corroborates that the headline result is an internal-consistency demonstration rather than external validation. The simulator-building and integration contributions remain substantive, so the analysis stops at 6 rather than 8-10.

Assumptions & free parameters 6 free parameters · 6 assumptions · 0 invented entities

The framework is built entirely from hand-set probabilities, hand-designed rewards, and LLM prompts, with no real-world data or external validation in the loop. The central results are therefore contingent on these authored parameters; every score and policy in the paper is a function of the authors' choices, not of independent measurement. This is a heavy burden for a paper that claims effective interpretation of complex PLWD needs.

free parameters (6)
  • Base transition probabilities (Table 3). = Forgetfulness 30%, Confusion 30%, Anger 5%, Disengagement 20%, self-persistence 99-100%, mutual influence +2% to +20%.
    Hand-chosen by the authors as simplifications of qualitative clinical observations. They determine the entire simulated dynamics, and no real patient data is used to fit them.
  • Assistance-conditioned transition probabilities (Section 2.1.2). = Verbal supportive: Anger/Disengagement persist 5%; non-directive: Forgetfulness 40%, Confusion 60% if currently Yes…
    Ad hoc numbers selected to encode the expected effect of each assistance type; no empirical calibration.
  • Subtask-skipping persistence probabilities (Section 2.1.3). = 0.5 for persistence of Forgetfulness/Confusion when both present, 0.2 when only one present; 0.5 for…
    Hand-chosen to reflect the idea that skipping a subtask only partially resets emotional and cognitive states.
  • Reward weights (Table 4). = wf=-1, wc=-1, wa=-5, wd=-1, cost per trial=-1, cost per timestep=-1, wsubtaskComplete=+50, wsubtaskSkip=-10…
    Designed by the authors to enforce the minimal-assistance and emotion-prioritization principles; these weights define what 'good' means in this simulation.
  • MaxTrials per subtask. = 5.
    Set in Section 3.2.2 without an empirical basis; it controls when the system skips a subtask.
  • Q-learning hyperparameters and epsilon schedule. = alpha=0.05, gamma=0.95, 6000 epochs x 30 episodes, epsilon_min=0.03, epsilon_max=1, lambda set so epsilon=0.8 at epoch…
    Standard RL hyperparameters chosen by hand; the epsilon decay schedule is tuned to the training length.
assumptions (6)
  • ad hoc to paper PLWD cognitive and affective states are binary and evolve as a Markov chain with the given hand-set transition probabilities.
    Section 2.1 explicitly lists these as assumptions ('Here are the four assumptions to simulate...'), with no empirical calibration.
  • ad hoc to paper Without external assistance, a PLWD almost never recovers from a negative state (probability of Yes->Yes near 1).
    Section 2.1 states the no-recovery assumption; it is a strong claim about disease dynamics.
  • domain assumption The four PASS-based assistance levels (none, supportive, non-directive, directive) are sufficient to capture effective caregiver help for mild-to-moderate dementia.
    Adopted from PASS [21]; the paper restricts to four of the ten PASS levels without demonstrating sufficiency.
  • domain assumption GPT-4o can produce realistic verbal and nonverbal PLWD behaviors from a latent binary state vector.
    Section 2.2 assumes LLM faithfulness; Section 4.3 flags hallucination and stereotyping risks.
  • domain assumption The reward function in Equation 1 encodes the principles that keeping PLWD positive and providing minimal assistance lead to better long-term outcomes.
    Section 3.2.3 states the design principles; the weights are not derived from clinical outcome data.
  • domain assumption An LLM can infer the PLWD state vector from behavior text accurately enough for good action selection.
    Section 3.1 defines the perception module; Section 4.2 shows a concrete misperception, and the paper acknowledges the limitation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Integrating Reinforcement Learning and AI Agents for Adaptive Robotic Interaction and Assistance in Dementia Care." pith.science (2026). https://pith.science/paper/DEYQLGJC

@misc{pith2026250117206,
  author       = {Pith},
  title        = {Pith review of: Integrating Reinforcement Learning and AI Agents for Adaptive Robotic Interaction and Assistance in Dementia Care},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DEYQLGJC}},
  note         = {Machine review of arXiv:2501.17206}
}
read the original abstract

This study explores a novel approach to advancing dementia care by integrating socially assistive robotics, reinforcement learning (RL), large language models (LLMs), and clinical domain expertise within a simulated environment. This integration addresses the critical challenge of limited experimental data in socially assistive robotics for dementia care, providing a dynamic simulation environment that realistically models interactions between persons living with dementia (PLWDs) and robotic caregivers. The proposed framework introduces a probabilistic model to represent the cognitive and emotional states of PLWDs, combined with an LLM-based behavior simulation to emulate their responses. We further develop and train an adaptive RL system enabling humanoid robots, such as Pepper, to deliver context-aware and personalized interactions and assistance based on PLWDs' cognitive and emotional states. The framework also generalizes to computer-based agents, highlighting its versatility. Results demonstrate that the RL system, enhanced by LLMs, effectively interprets and responds to the complex needs of PLWDs, providing tailored caregiving strategies. This research contributes to human-computer and human-robot interaction by offering a customizable AI-driven caregiving platform, advancing understanding of dementia-related challenges, and fostering collaborative innovation in assistive technologies. The proposed approach has the potential to enhance the independence and quality of life for PLWDs while alleviating caregiver burden, underscoring the transformative role of interaction-focused AI systems in dementia care.

Figures

Figures reproduced from arXiv: 2501.17206 by the authors.

Figure 1
Figure 1. Illustration of the system architecture for our two AI agents, [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. The flowchart illustrates the dynamic interaction between the two AI agents, a person living with dementia [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. A high-level illustration of the state transition model for agent [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: A graph illustration of how the dynamics of PLWD’s status [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: The prompt used in our PLWD behavior simulation model to generate PLWD’s realistic, detailed behavior. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: The prompt used in the robot perception module. [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: The prompt used in the robot action execution module, with a prompt of including a brief assistance guidance [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]
Figure 8
Figure 8. Figure 8: The prompt used in the robot action execution module, with a prompt of including a detailed assistance [PITH_FULL_IMAGE:figures/full_fig_p010_8.png]
Figure 9
Figure 9. Figure 9: Average Return from 40 Experiments Using Temporal Optimal Policy [PITH_FULL_IMAGE:figures/full_fig_p011_9.png]
Figure 10
Figure 10. Figure 10: The optimal policy learned by the constant-epsilon greedy RL agent (upper lane) and the decaying epsilon [PITH_FULL_IMAGE:figures/full_fig_p012_10.png]
Figure 11
Figure 11. Figure 11: Screenshot of experimental results recording in txt. [PITH_FULL_IMAGE:figures/full_fig_p013_11.png]
Figure 12
Figure 12. Figure 12: Example of the interaction between our AI agent [PITH_FULL_IMAGE:figures/full_fig_p014_12.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Beyond Prediction: Reinforcement Learning as the Defining Leap in Healthcare AI

    cs.LG 2025-08 reject novelty 3.0 of 10

    A survey of reinforcement learning in healthcare that frames RL as a paradigm shift from prediction to agentive clinical intelligence.

Reference graph

Works this paper leans on

34 extracted references · 32 canonical work pages · cited by 1 Pith paper

  1. [1]

    2024 alzheimer’s disease facts and figures.Alzheimer’s & Association, 20(5):3708–3821, May 2024

    Alzheimer’s Association. 2024 alzheimer’s disease facts and figures.Alzheimer’s & Association, 20(5):3708–3821, May 2024

    work page 2024

  2. [2]

    Discovering the person with alzheimer’s disease: cognitive, emotional and behavioural aspects

    RT Woods. Discovering the person with alzheimer’s disease: cognitive, emotional and behavioural aspects. Aging & Mental Health, 5(sup1):7–16, 2001

    work page 2001

  3. [3]

    Recommendations to address key recruitment challenges of alzheimer’s disease clinical trials

    Jessica B Langbaum, Julie Zissimopoulos, Rhoda Au, Niranjan Bose, Chris J Edgar, Evan Ehrenberg, Howard Fillit, Carl V Hill, Lynne Hughes, Michael Irizarry, et al. Recommendations to address key recruitment challenges of alzheimer’s disease clinical trials. Alzheimer’s & Dementia, 19(2):696–707, 2023

    work page 2023

  4. [4]

    The public health road map to respond to the growing dementia crisis

    Benjamin S Olivari, Molly E French, and Lisa C McGuire. The public health road map to respond to the growing dementia crisis. Innovation in Aging, 4(1):igz043, 2020

    work page 2020

  5. [5]

    The promise of technology in the future of dementia care

    Wendy Moyle. The promise of technology in the future of dementia care. Nature Reviews Neurology, 15(6):353– 359, 2019

    work page 2019

  6. [6]

    A systematic review of robotic rehabilitation for cognitive training

    Fengpei Yuan, Elizabeth Klavon, Ziming Liu, Ruth Palan Lopez, and Xiaopeng Zhao. A systematic review of robotic rehabilitation for cognitive training. Frontiers in Robotics and AI, 8:605715, 2021

    work page 2021

  7. [7]

    Social Robots for Older Adults with Dementia: A Narrative Review on Challenges & Future Directions

    Daniel Woods, Fengpei Yuan, Ying-Ling Jao, and Xiaopeng Zhao. Social Robots for Older Adults with Dementia: A Narrative Review on Challenges & Future Directions. In International Conference on Social Robotics, pages 411–420, Cham, 2021. Springer

    work page 2021

  8. [8]

    Cognitive exercise for persons with Alzheimer’s disease and related dementia using a social robot

    Fengpei Yuan, Marie Boltz, Dania Bilal, Ying-Ling Jao, Monica Crane, Joshua Duzan, Abdurhman Bahour, and Xiaopeng Zhao. Cognitive exercise for persons with Alzheimer’s disease and related dementia using a social robot. IEEE Transactions on Robotics, 39(4):3332–3346, 2023. ISBN: 1552-3098 Publisher: IEEE

    work page 2023

Show all 34 references
  1. [9]

    Assistive technologies in dementia care: an updated analysis of the literature

    Alessandro Pappadà, Rabih Chattat, Ilaria Chirico, Marco Valente, and Giovanni Ottoboni. Assistive technologies in dementia care: an updated analysis of the literature. Frontiers in psychology, 12:644587, 2021

    2021

  2. [10]

    Kwan Min Lee, Younbo Jung, Jaywoo Kim, and Sang Ryong Kim. Are physically embodied social agents better than disembodied social agents?: The effects of physical embodiment, tactile interaction, and people’s loneliness in human–robot interaction. International journal of human-...

    2006

  3. [11]

    Social robots for the care of persons with dementia: a systematic review

    Moojan Ghafurian, Jesse Hoey, and Kerstin Dautenhahn. Social robots for the care of persons with dementia: a systematic review. ACM Transactions on Human-Robot Interaction (THRI), 10(4):1–31, 2021

    2021

  4. [12]

    The benefits of and barriers to using a social robot PARO in care settings: a scoping review

    L Hung, C Liu, E Woldum, A Au-Yeung, A Berndt, C Wallsworth, N Horne, M Gregorio, J Mann, and H Chaudhury. The benefits of and barriers to using a social robot PARO in care settings: a scoping review. BMC Geriatr, 19:232, 2019

    2019

  5. [13]

    Design and implementation framework of social assistive robotics for people with dementia-a scoping review

    Deniz Ozdemir, Jaroslav Cibulka, Olga Stepankova, and Iva Holmerova. Design and implementation framework of social assistive robotics for people with dementia-a scoping review. Health and Technology, 11(2):367–378, 2021

    2021

  6. [14]

    ENRICHME: Perception and Interaction of an Assistive Robot for the Elderly at Home

    Serhan Co¸ sar, Manuel Fernandez-Carmona, Roxana Agrigoroaie, Jordi Pages, François Ferland, Feng Zhao, Shigang Yue, Nicola Bellotto, and Adriana Tapus. ENRICHME: Perception and Interaction of an Assistive Robot for the Elderly at Home. International Journal of Social Robotics...

    2020

  7. [15]

    A simulated experiment to explore robotic dialogue strategies for people with dementia.Journal of Rehabilitation and Assistive Technologies Engineering, 9:20556683221105768, 2022

    Fengpei Yuan, Amir Sadovnik, Ran Zhang, Devin Casenhiser, Eun Jin Paek, and Xiaopeng Zhao. A simulated experiment to explore robotic dialogue strategies for people with dementia.Journal of Rehabilitation and Assistive Technologies Engineering, 9:20556683221105768, 2022. ISBN: ...

    2022

  8. [16]

    SocialRobot: An interactive mobile robot for elderly home care

    David Portugal, Luís Santos, Paulo Alvito, Jorge Dias, George Samaras, and Eleni Christodoulou. SocialRobot: An interactive mobile robot for elderly home care. In 2015 IEEE/SICE International Symposium on System Integration (SII), pages 811–816, Nagoya, Japan, 2015. IEEE

    2015

  9. [17]

    Socially Assistive Robots for Dementia Care: Exploring Caregiver Perceptions of Use Cases and Acceptance

    Xian Wu, Anne E Adams, Jane C Komsky, Sarah E Saint, Taylor E Mackin, Jason P Zamer, Daniel S Hedin, Robert J Dahlstrom, and Jenay M Beer. Socially Assistive Robots for Dementia Care: Exploring Caregiver Perceptions of Use Cases and Acceptance. In Proceedings of the Human Fact...

    2021

  10. [18]

    A social robot-facilitated performance assessment of self-care skills for people with alzheimer’s: A preliminary study

    Fengpei Yuan, Robert Bray, Michael Oliver, Joshua Duzan, Monica Crane, and Xiaopeng Zhao. A social robot-facilitated performance assessment of self-care skills for people with alzheimer’s: A preliminary study. International Journal of Social Robotics, 16(6):2065–2078, 2024

    2024

  11. [19]

    Assessing the acceptability of a humanoid robot for alzheimer’s disease and related dementia care using an online survey

    Fengpei Yuan, Joel G Anderson, Tami H Wyatt, Ruth Palan Lopez, Monica Crane, Austin Montgomery, and Xiaopeng Zhao. Assessing the acceptability of a humanoid robot for alzheimer’s disease and related dementia care using an online survey. International Journal of Social Robotics...

    2022

  12. [20]

    Robots to assist daily activities: views of older adults with alzheimer’s disease and their caregivers

    Rosalie H Wang, Aishwarya Sudhama, Momotaz Begum, Rajibul Huq, and Alex Mihailidis. Robots to assist daily activities: views of older adults with alzheimer’s disease and their caregivers. International psychogeriatrics, 29(1):67–79, 2017

    2017

  13. [21]

    Evaluating capacity to live independently and safely in the community: Performance assessment of self-care skills

    Denise Chisholm, Pamela Toto, Ketki Raina, Margo Holm, and Joan Rogers. Evaluating capacity to live independently and safely in the community: Performance assessment of self-care skills. British Journal of Occupational Therapy, 77(2):59–63, 2014

    2014

  14. [22]

    Functional competence and cognition in individuals with amnestic mild cognitive impairment

    Pallavi Dham, Kathleen S Bingham, Christopher R Bowie, Meryl A Butters, Corinne E Fischer, Alastair Flint, Nathan Herrmann, Sanjeev Kumar, Linda Mah, Benoit H Mulsant, et al. Functional competence and cognition in individuals with amnestic mild cognitive impairment. Journal of...

    2020

  15. [23]

    Amyloid accumulation associated with worse performance of complex task

    Audrey A Keleman, Rebecca M Bollinger, Beau M Ances, and Susan L Stark. Amyloid accumulation associated with worse performance of complex task. Alzheimer’s & Dementia, 17:e054605, 2021

    2021

  16. [24]

    Assessment of a humanoid partner for older adults and persons with dementia during home-based activities

    Fengpei Yuan, Marie Boltz, Ying-Ling Jao, Arowyn Casenhiser, Aidan Siddiqi, Robert Bray, Joshua Duzan, Monica Crane, and Xiaopeng Zhao. Assessment of a humanoid partner for older adults and persons with dementia during home-based activities. In International Conference on Soci...

    2022

  17. [25]

    Hallucinations in llms: Understanding and addressing challenges

    Gabrijela Perkovi´c, Antun Drobnjak, and Ivica Botiˇcki. Hallucinations in llms: Understanding and addressing challenges. In 2024 47th MIPRO ICT and Electronics Convention (MIPRO), pages 2084–2088, Opatija, Croatia,

    2024

  18. [26]

    Same task, more tokens: the impact of input length on the reasoning performance of large language models, 2024

    Mosh Levy, Alon Jacoby, and Yoav Goldberg. Same task, more tokens: the impact of input length on the reasoning performance of large language models, 2024. Accepted to ACL 2024

    2024

  19. [27]

    Assessment and management of behavioral and psychological symptoms of dementia

    Helen C Kales, Laura N Gitlin, and Constantine G Lyketsos. Assessment and management of behavioral and psychological symptoms of dementia. BMJ, 350:h369, 2015

    2015

  20. [28]

    Technical challenges for smooth interaction with seniors with dementia: Lessons from humanitude ™

    Hidenobu Sumioka, Masahiro Shiomi, Miwako Honda, and Atsushi Nakazawa. Technical challenges for smooth interaction with seniors with dementia: Lessons from humanitude ™. Frontiers in Robotics and AI, 8:650906, 2021

    2021

  21. [29]

    Ai and human-robot interaction: A review of recent advances and challenges

    Alexander Obaigbena, Oluwaseun Augustine Lottu, Ejike David Ugwuanyi, Boma Sonimitiem Jacks, Enoch Oluwademilade Sodiya, and Obinna Donald Daraojimba. Ai and human-robot interaction: A review of recent advances and challenges. GSC Advanced Research and Reviews, 18(2):321–330, 2024

    2024

  22. [30]

    Machine learning for synthetic data generation: a review

    Yingzhou Lu, Minjie Shen, Huazheng Wang, Xiao Wang, Capucine van Rechem, Tianfan Fu, and Wenqi Wei. Machine learning for synthetic data generation: a review. arXiv preprint arXiv:2302.04062, 2023

    2023 arXiv

  23. [31]

    Artificial intelligence and technology collaboratories: Innovating aging research and alzheimer’s care

    Peter Abadir, Esther Oh, Rama Chellappa, Niteesh Choudhry, George Demiris, Deepak Ganesan, Jason Karlawish, Benjamin Marlin, Rose M Li, Najim Dehak, et al. Artificial intelligence and technology collaboratories: Innovating aging research and alzheimer’s care. Alzheimer’s & Dem...

    2024

  24. [32]

    A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions

    Lei Huang, Weijiang Yu, Weitao Ma, Weihong Zhong, Zhangyin Feng, Haotian Wang, Qianglong Chen, Weihua Peng, Xiaocheng Feng, Bing Qin, et al. A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions. ACM Transactions on Informatio...

    2023

  25. [33]

    A comprehen- sive survey of hallucination mitigation techniques in large language models

    SM Tonmoy, SM Zaman, Vinija Jain, Anku Rani, Vipula Rawte, Aman Chadha, and Amitava Das. A comprehen- sive survey of hallucination mitigation techniques in large language models. arXiv preprint arXiv:2401.01313, 2024

    2024 arXiv

  26. [34]

    Toward a collaborative ai framework for assistive dementia care

    Tze-Yun Leong. Toward a collaborative ai framework for assistive dementia care. InWorkshops at the Thirty-First AAAI Conference on Artificial Intelligence, 2017. 18

    2017

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.