Pith. sign in

REVIEW 3 major objections 4 minor 37 references

Personalised Explanations in Long-term Human-Robot Interactions

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a robot can tailor explanations to each user's prior knowledge by retrieving stored concepts from a persistent user memory, and that a two-stage generate-then-personalise architecture shortens explanations only when…

desk verdict A useful architecture comparison with an overstated abstract: the 'only when' claim depends entirely on the retrieval threshold. read the letter →

arxiv 2507.03049 v1 pith:4OSEL5M4 submitted 2025-07-03 cs.RO cs.AIcs.HC

Ferran Gebellí , Anaís Garrell , Jan-Gerrit Habekost , Séverin Lemaignan , Stefan Wermter , Raquel Ros This is my paper · ORCID
classification cs.ROcs.AIcs.HC
keywords human-robotinteractionexplainableroboticspersonalisedexplanationslargelanguagemodelsusermemorymodelretrieval-augmentedgenerationknowledgetracinglong-term
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 is trying to establish that a robot can personalise the level of detail in its explanations by keeping a persistent memory of the concepts each user has learned, and that the best way to use that memory is a two-stage process: first generate a full explanation, then shorten it only if the user's stored knowledge is relevant. The authors propose a framework with four modules, including persistent user memory, knowledge extraction, retrieval of relevant knowledge, and post-processing, and compare three architectures that differ in what is used to query the memory: the user's question, the robot context, or the generated explanation. In controlled synthetic experiments across a hospital patrolling robot and a kitchen assistant robot, the explanation-query architecture with a retrieval threshold retrieves relevant concepts most accurately and produces length reductions for known topics while leaving unknown topics nearly untouched. If correct, the framework gives a way to avoid redundant detail in long-term interactions, but the level-of-detail results are measured against synthetic topic labels, not against what real users actually remember.

What carries the argument

The load-bearing object is a persistent user memory: a dynamic list of short string concepts that the user is estimated to know, each stored with a timestamp and an embedding vector. A knowledge-extraction module uses an LLM to populate this memory from previous explanations; a retrieval module selects relevant concepts by cosine similarity between the query and the stored embeddings, gated by a threshold $\tau_{\text{retr}}$ so that no user context is injected when nothing related is found; and a post-processing module applies an exponential forgetting curve $P = e^{-\lambda t}$ to turn elapsed time and similarity into qualitative probability descriptors and one of four detail-level instructions: full detail, moderately short, concise, or very short summary. The architecture comparison is what carries the argument: the same framework with a different query for retrieval behaves differently, and the two-stage explanation-query version, where the first LLM stage generates a full explanation that is then used as the query for the second personalisation stage, is the one that systematically retrieves relevant concepts and shortens explanations only for known topics.

What would settle it

Give participants a realistic interaction with the robot, wait one week and one month, measure what they actually remember with a recall test, then compare the system's memory retrieval and length reduction against their measured knowledge: the framework's claim predicts that explanations should be shortened only for participants who demonstrably recall the retrieved concept, and that unknown topics should keep their full length; observing shortening when recall is absent, or full explanations when recall is present, would falsify the claim.

Watch

Extended reading notes

Core claim

On the terms of the paper, the central discovery is that retrieval quality determines personalisation quality, and the full explanation is the best retrieval query. The authors show that when the user's question is used as the query, the same question can map to different underlying causes, so retrieval is noisy; when the robot context is used, retrieval improves; and when the generated explanation itself is used, the retrieved concepts are closest in semantic similarity to what the user previously learned. In the two-stage explanation-query architecture, an initial LLM produces a complete explanation, and a second LLM adapts it to the user context only when at least one memory concept is retrieved. With a mid-level similarity threshold ($\tau_{\text{retr}} = 0.45$), this architecture delivers the best precision and false-positive rate, reduces explanation length by roughly 26 to 39 percent for topics known from a week or a month earlier, and reduces length by only a few percent for unknown topics; with a zero threshold, irrelevant memories are retrieved and explanations are shortened even without related user knowledge, which the authors take as confirmation that the retrieval threshold is doing the essential work.

Load-bearing premise

The load-bearing premise is that a user's knowledge can be captured as a small set of discrete string concepts automatically extracted from prior explanations, and that the cosine similarity between the embeddings of those strings and the current query reliably indicates whether the user actually holds related knowledge.

Editorial extensions

If this is right

  • A robot using this framework can shorten an explanation about a topic the user once discussed, and can reference that earlier conversation when doing so.
  • When no related concept is retrieved, the two-stage architecture forwards the full explanation, so users who have never encountered a topic get complete information by default.
  • The two-stage design can be attached to any existing explanation generator, including non-LLM systems, because personalisation acts as a separate post-processing step.
  • Retrieval thresholds can be tuned for the application, trading fewer false positives against missing relevant memories.
  • In long-term interactions, the system should avoid repeating previously explained information, reducing cognitive load without hiding new content.

Reading between the lines

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

  • An untested consequence is that the framework's detail-level instructions will be correct only if the forgetting curve $P=e^{-\lambda t}$ matches the target population; the paper itself leaves the decay parameter to be set per user type, so a longitudinal user study could estimate $\lambda$ from measured recall and expose a mismatch.
  • Because memory entries are plain strings with embeddings, the same retrieval machinery could be populated from multimodal sources such as GUI interactions or observation systems, which the paper identifies as future work; that extension would make personalisation depend less on verbal explanations.
  • The comparison suggests a general design principle beyond this setting: the quality of a memory-based personalisation system is bounded by how semantically close the retrieval query is to the stored knowledge, so anything that moves the query closer to the eventual explanation should improve retrieval.
  • A practical implication not tested here is that the threshold itself could be made adaptive per user, since the reported precision-recall trade-off means a user who prefers redundancy could use a low threshold while one who wants brevity could use a high one.
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

3 major / 4 minor

Summary. The paper proposes a framework for personalised explanations in human-robot interaction that maintains a per-user memory of concepts extracted from prior explanations, retrieves relevant concepts via embedding similarity with a tunable threshold, and uses an Ebbinghaus-style forgetting curve to estimate the user's current knowledge. Three architectures are compared: question-query, robot-context-query, and a two-stage explanation-query that first generates a full explanation and then adapts its level of detail. The framework is evaluated in two synthetic scenarios, a kitchen assistant robot and a hospital patrolling robot, across known/unknown topics and one-week/one-month delays, with length reduction and retrieval metrics as the main outcomes. The paper claims that the two-stage architecture effectively reduces the level of detail only when there is related user knowledge.

Significance. If the central claim were established, the paper would be a useful step toward long-term personalised XHRI, because it introduces a modular architecture that can attach personalisation to arbitrary explanation generators and it ships reproducible code and experiments. The retrieval results across thresholds provide informative evidence about how the choice of query content affects precision and recall in a RAG-style user-memory design. However, the experiments are entirely synthetic, the headline conclusion overreaches the evidence, and the main length-reduction outcome is partly forced by the system's skip rule rather than by measured user knowledge.

major comments (3)
  1. [Sec III-E, Table I] The abstract's claim that the two-stage architecture 'effectively reduces the level of detail only when there is related user knowledge' is not established. The second adaptation stage is executed only if at least one concept is retrieved; otherwise the full explanation is forwarded unchanged. Therefore, for unknown topics at mid/high thresholds, near-zero length reduction is guaranteed by construction whenever no concept passes the threshold. Table I confirms this: the unknown explanation-query rows at mid/high thresholds show 2.51% and 1.63% length reduction for the one-week condition. Moreover, length reduction is triggered by any retrieved concept regardless of relevance: at tau_retr=0, unknown topics receive 38.35% length reduction, and at tau_retr=low, 23.86%. To support the 'only when related' qualifier, the authors need to condition length reduction on retrieval presence and on false-positive retrievals, not only on the known/unknown topic label.
  2. [Sec III-B/C and Sec IV-C] The link between retrieved concepts and actual user knowledge is assumed, not measured. The persistent user memory is populated by LLM extraction from explanations, and 'known' versus 'unknown' is defined by the experimenter's topic assignment in synthetic two-interaction scenarios. No data establish that the automatically extracted concepts, or their cosine similarities to a query, correspond to what a user actually knows or remembers. Consequently, the conclusion that the framework 'effectively leverages the user's previous knowledge' is not supported by the experiments, which measure only text-length reduction in a simulated setting. The paper should either add a human evaluation of knowledge and comprehension, or substantially weaken the claims to state that the system reduces length when its memory retrieval succeeds, not when the user demonstrably holds related knowledge.
  3. [Section V, Table I and Fig. 11] The comparative claims are made without any measure of variability or statistical testing. The reported means are averages over 140 runs, but no confidence intervals, standard deviations, or significance tests are provided; for instance, the difference between known one-week explanation-query at low and mid thresholds (43.89% vs 38.98%) may be within sampling noise. Without such information, the statements in Sec VI-B that one ablation 'delivers the broadest difference' and that tau_retr=mid shows 'the best tradeoff' are not quantitatively supported. The authors should report error bars and, where appropriate, paired comparisons across the shared random draws.
minor comments (4)
  1. [Sec VI] The manual inspection of 10% of outputs is described without selection criteria, annotation protocol, or inter-rater reliability; please specify what 'truthful and adequately addressed' means and how the 10% were sampled.
  2. [Sec III-D] The four detail-level instructions are listed, but the thresholds that map the probability-similarity product to each instruction are not specified; please state the values or explain how they were chosen.
  3. [Fig. 11 and Table I] The labels such as 'quest./r. context query' are cramped and hard to read; consider using the full architecture names or a legend with defined abbreviations.
  4. [References and Fig. 10] Reference [13] contains a typo ('navigatior's'), and the timeline in Fig. 10 would be clearer if the four test-case conditions were labeled directly in the main text rather than only in the figure.

Circularity Check

2 steps flagged · score 6.0 of 10

Headline claim that the two-stage architecture reduces detail 'only when there is related user knowledge' is largely a restatement of the retrieval-gating rule: with no retrieved concepts the second stage is skipped, so low length reduction for unknown topics is by construction, not evidence of personalisation.

  1. self definitional [Sec. III-E (Explanation Generation); Sec. VI-B (Length reduction); Table I]
    "Importantly, the second step is only executed if there are retrieved concepts; otherwise, the full explanation is forwarded. ... the length reduction difference between the known and unknown settings should be as high as possible, indicating that the system effectively reduces detail only when the user has related knowledge."

    Because the adaptation stage is skipped whenever no concepts are retrieved, a zero or near-zero length reduction for unknown topics at mid/high thresholds is guaranteed by the if-statement, not by personalisation. The paper operationally defines 'related user knowledge' as 'retrieved concepts' in Sec. III-C, so the headline conclusion 'reduces only when there is related user knowledge' is definitionally equivalent to 'reduces only when at least one concept is retrieved'. Table I confirms this: at tau_retr = 0, where all concepts are retrieved, unknown topics still receive 37-38% length reduction, showing that the known/unknown gap is a retrieval-presence artifact rather than a measured effect of user knowledge.

  2. self definitional [Sec. III-D (Post-processing of Retrieved Knowledge); Sec. V (Results)]
    "The probabilities are weighted by the retrieval similarity scores, and the obtained values are thresholded into one of the following instructions that specify detail level target: (1) Provide a complete explanation with all the details, (2) The explanation should be moderately short, (3) The explanation should be concise, consisting of one or two sentences or (4) Provide a very short summary in a single sentence."

    The length reduction that is later reported as evidence of personalisation is directly generated by thresholding the same retrieval similarity scores that are used to define whether the user has related knowledge. A higher retrieval score produces a shorter instruction, and a lower score produces a complete explanation; therefore the observed 'length reduction' is not an independent outcome validating the personalisation claim but a mechanical output of the post-processing rule. The architecture comparison retains empirical content in retrieval precision/recall, but the central known/unknown length-reduction gap is built into the prompt-generation step.

full rationale

The paper's strongest claim, that the two-stage architecture 'effectively reduces the level of detail only when there is related user knowledge', is not independently established. The mechanism that produces the effect is the retrieval gate: the second stage is executed only if concepts are retrieved, otherwise the full explanation is forwarded unchanged. Consequently, low reduction for unknown topics at mid/high retrieval thresholds is guaranteed whenever no concept passes the threshold, and the paper's own zero-threshold results show that irrelevant retrieved concepts still trigger 37-38% length reduction. The detail-level instructions are also generated by thresholding the same similarity scores used to label knowledge as related, so the reported length reduction is in part a direct consequence of the instruction-selection rule. The retrieval precision/recall comparisons and the token-length analysis are genuine empirical contributions and are not circular. There is a self-citation to the authors' prior framework [8], but it is not load-bearing for the specific numerical claims, so it does not independently raise the score. Overall, the central claim partially reduces by construction, giving a circularity score of 6.

Assumptions & free parameters 4 free parameters · 6 assumptions · 1 invented entities

The central claim rests on the assumption that user knowledge is representable as discrete concepts and that retrieval by embedding similarity captures it; both are untested with human participants. The prompt instructions and thresholds are hand-set, so the detail reduction outcome is partly a design property.

free parameters (4)
  • forgetting decay parameter lambda = not reported
    Defines P=e^{-lambda t} in Sec. III-D; controls how quickly retrieved concepts are judged forgotten and therefore the detail-level instruction; can be adjusted by user type, but no value or calibration is given.
  • retrieval similarity threshold tau_retr = 0, 0.35, 0.45, 0.55 in ablations
    Sec. III-C: determines whether any user context is retrieved; the headline 'reduces only when related knowledge' only holds for the mid/high values.
  • concept deduplication threshold = predefined threshold, value not given
    Sec. III-B: used to replace memory entries whose embedding similarity is high; affects memory content.
  • probability-to-detail-level thresholds = not specified
    Sec. III-D: the weighted probabilities are thresholded into four detail instructions; thresholds are hand-set and not reported.
assumptions (6)
  • domain assumption Relevant robot internal information is available in the robot context for explanation generation.
    Stated in Sec. III: 'our framework assumes that the relevant robot internal information to generate the explanation is available within the robot context'. If logs lack root causes, explanations cannot be truthful.
  • domain assumption User knowledge can be represented as a discrete list of string concepts extracted from prior explanations.
    Sec. III-B defines the persistent user memory as a dynamic list of concepts; no user study tests whether this captures what users actually know.
  • domain assumption Ebbinghaus exponential forgetting curve P=e^{-lambda t} models user knowledge retention.
    Sec. III-D applies the curve with an adjustable lambda; this psychological model is treated as valid for robot explanation knowledge.
  • domain assumption Embedding cosine similarity between query and stored concepts corresponds to semantic relatedness of user knowledge.
    Sec. III-C uses OpenAI text-embedding-3-small cosine similarity for retrieval; no ground-truth evaluation of similarity quality beyond topic-based precision/recall.
  • domain assumption A privacy-aware user re-identification system selects the correct memory model.
    Footnote 2 in Sec. III assumes such a system exists; otherwise personalisation cannot start.
  • domain assumption The synthetic test scenarios are representative of real long-term human-robot interactions.
    Sec. IV says evaluation is 'conducted based on synthetic data'; no human participants, so generalisation to real users is assumed.
invented entities (1)
  • persistent user memory (concept list with embeddings and timestamps)
    purpose: Stores estimated user knowledge so explanations can be personalised
    The memory is an internal representation introduced by the paper; its accuracy is never validated against measured user knowledge, and there is no external falsifiable handle.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Personalised Explanations in Long-term Human-Robot Interactions." pith.science (2026). https://pith.science/paper/4OSEL5M4

@misc{pith2026250703049,
  author       = {Pith},
  title        = {Pith review of: Personalised Explanations in Long-term Human-Robot Interactions},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4OSEL5M4}},
  note         = {Machine review of arXiv:2507.03049}
}
read the original abstract

In the field of Human-Robot Interaction (HRI), a fundamental challenge is to facilitate human understanding of robots. The emerging domain of eXplainable HRI (XHRI) investigates methods to generate explanations and evaluate their impact on human-robot interactions. Previous works have highlighted the need to personalise the level of detail of these explanations to enhance usability and comprehension. Our paper presents a framework designed to update and retrieve user knowledge-memory models, allowing for adapting the explanations' level of detail while referencing previously acquired concepts. Three architectures based on our proposed framework that use Large Language Models (LLMs) are evaluated in two distinct scenarios: a hospital patrolling robot and a kitchen assistant robot. Experimental results demonstrate that a two-stage architecture, which first generates an explanation and then personalises it, is the framework architecture that effectively reduces the level of detail only when there is related user knowledge.

Figures

Figures reproduced from arXiv: 2507.03049 by the authors.

Figure 1
Figure 1. Illustrative personalisation of an explanation’s level of detail in a [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of the proposed architectures that employ either the [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Example prompt to extract the concepts from new conversations. [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Prompt structure for the one-stage explanation generation process. [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 6
Figure 6. Figure 6: Use case of a kitchen assistant robot. B. Hospital patrolling robot use case In this use case, a robot is in charge of autonomously patrolling the rehabilitation unit of an intermediate healthcare facility. The robot is tasked to alert the nursing staff in case of haza…
Figure 7
Figure 7. Figure 7: Topics and sample questions for the kitchen assistant use case. [PITH_FULL_IMAGE:figures/full_fig_p005_7.png]
Figure 8
Figure 8. Figure 8: Use case of a robot patrolling a healthcare facility. [PITH_FULL_IMAGE:figures/full_fig_p005_8.png]
Figure 11
Figure 11. Figure 11: Length reduction of the personalised explanations compared to the baseline explanation length. [PITH_FULL_IMAGE:figures/full_fig_p006_11.png]
Figure 12
Figure 12. Figure 12: Processed tokens in the prompt and response. [PITH_FULL_IMAGE:figures/full_fig_p007_12.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

37 extracted references · 31 canonical work pages

  1. [8]

    Gebell ´ı, L

    F. Gebell ´ı, L. Hriscu, R. Ros, S. Lemaignan, A. Sanfeliu, and A. Gar- rell. Personalised explainable robots using llms. In Companion of the ACM/IEEE International Conference on Human-Robot Interaction, 2025

    work page 2025

  2. [1]

    Andriella, C

    A. Andriella, C. Torras, and G. Aleny `a. Learning robot policies using a high-level abstraction persona-behaviour simulator. In 2019 28th IEEE International Conference on Robot and Human Interactive Communication (RO-MAN), pages 1–8, 2019

    work page 2019

  3. [2]

    Anjomshoae, A

    S. Anjomshoae, A. Najjar, D. Calvaresi, and K. Fr ¨amling. Explain- able agents and robots: Results from a systematic literature review. In International Conference on Autonomous Agents and Multiagent Systems, pages 1078–1088, 2019

    work page 2019

  4. [3]

    Belpaeme

    T. Belpaeme. Advice to new human-robot interaction researchers. Human-Robot Interaction: Evaluation Methods and their Standard- ization, pages 355–369, 2020

    work page 2020

  5. [4]

    Towards Personalized Explanation of Robot Path Planning via User Feedback

    K. Boggess, S. Chen, and L. Feng. Towards personalized expla- nation of robot path planning via user feedback. arXiv preprint arXiv:2011.00524, 2020

    work page Pith review arXiv 2011

  6. [5]

    H. Cen, K. Koedinger, and B. Junker. Learning factors analysis–a general method for cognitive model evaluation and improvement. In International Conference on Intelligent Tutoring Systems , pages 164–

    work page

  7. [6]

    A. T. Corbett and J. R. Anderson. Knowledge tracing: Modeling the acquisition of procedural knowledge. User Modeling and User- adapted Interaction, 4:253–278, 1994

    work page 1994

  8. [7]

    Ebbinghaus

    H. Ebbinghaus. ¨Uber das ged ¨achtnis: untersuchungen zur experi- mentellen psychologie. Duncker & Humblot , 1885

    work page

Show all 37 references
  1. [9]

    Hellstr ¨om and S

    T. Hellstr ¨om and S. Bensch. Understandable robots-what, why, and how. Paladyn, Journal of Behavioral Robotics , 9(1):110–123, 2018

    2018

  2. [10]

    D. Hillson. Describing probability: The limitations of natural language. Project Management Institute, 2005

    2005

  3. [11]

    Irfan, S.-M

    B. Irfan, S.-M. Kuoppam ¨aki, and G. Skantze. Between reality and delusion: challenges of applying large language models to companion robots for open-domain dialogues with older adults. 2023

    2023

  4. [12]

    H. Joko, S. Chatterjee, A. Ramsay, A. P. De Vries, J. Dalton, and F. Hasibi. Doing personal laps: Llm-augmented dialogue construction for personalized multi-session conversational search. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development...

    2024

  5. [13]

    Korpan, D

    R. Korpan, D. Tiourine, S. Chen, and S. Epstein. Evaluation of a robot navigatior’s explanations. In Companion of the ACM/IEEE International Conference on Human-Robot Interaction , 2025

    2025

  6. [14]

    Kulesza, S

    T. Kulesza, S. Stumpf, M. Burnett, and I. Kwan. Tell me more? the effects of mental model soundness on personalizing an intelligent agent. In Proceedings of the Sigchi Conference on Human Factors in Computing Systems, pages 1–10, 2012

    2012

  7. [15]

    LeMasurier, C

    G. LeMasurier, C. Tagliamonte, J. Breen, D. Maccaline, and H. A. Yanco. Templated vs. generative: Explaining robot failures. In 33rd IEEE International Conference on Robot and Human Interactive Communication, pages 1346–1353. IEEE, 2024

    2024

  8. [16]

    J. Liu, Y . Zhu, S. Wang, X. Wei, E. Min, Y . Lu, S. Wang, D. Yin, and Z. Dou. Llms+ persona-plug= personalized llms. arXiv preprint arXiv:2409.11901, 2024

    2024 arXiv

  9. [17]

    Z. Liu, A. Bahety, and S. Song. Reflect: Summarizing robot experiences for failure explanation and correction. arXiv preprint arXiv:2306.15724, 2023

    2023 arXiv

  10. [18]

    Madaan, N

    A. Madaan, N. Tandon, P. Clark, and Y . Yang. Memory-assisted prompt editing to improve gpt-3 after deployment. arXiv preprint arXiv:2201.06009, 2022

    2022 arXiv

  11. [19]

    T. Miller. Explanation in artificial intelligence: Insights from the social sciences. Artificial Intelligence, 267:1–38, 2019

    2019

  12. [20]

    Mualla, I

    Y . Mualla, I. Tchappi, T. Kampik, A. Najjar, D. Calvaresi, A. Abbas- Turki, S. Galland, and C. Nicolle. The quest of parsimonious xai: A human-agent architecture for explanation formulation. Artificial Intelligence, 302:103573, 2022

    2022

  13. [21]

    Mualla, I

    Y . Mualla, I. H. Tchappi, A. Najjar, T. Kampik, S. Galland, and C. Nicolle. Human-agent explainability: An experimental case study on the filtering of explanations. In ICAART, pages 378–385, 2020

    2020

  14. [22]

    Onorati, ´A

    T. Onorati, ´A. Castro-Gonz ´alez, J. C. del Valle, P. D ´ıaz, and J. C. Castillo. Creating personalized verbal human-robot interactions using llm with the robot mini. In International Conference on Ubiquitous Computing and Ambient Intelligence , pages 148–159. Springer, 2023

    2023

  15. [23]

    Piech, J

    C. Piech, J. Bassen, J. Huang, S. Ganguli, M. Sahami, L. J. Guibas, and J. Sohl-Dickstein. Deep knowledge tracing. Advances in Neural Information Processing Systems , 28, 2015

    2015

  16. [24]

    Richardson, Y

    C. Richardson, Y . Zhang, K. Gillespie, S. Kar, A. Singh, Z. Raeesy, O. Z. Khan, and A. Sethy. Integrating summarization and retrieval for enhanced personalization via large language models. arXiv preprint arXiv:2310.20081, 2023

    arXiv

  17. [25]

    Rossi, F

    S. Rossi, F. Ferland, and A. Tapus. User profiling and behavioral adaptation for HRI. Pattern Recognition Letters, 99:3–12, 2017

    2017

  18. [26]

    Salemi, S

    A. Salemi, S. Kallumadi, and H. Zamani. Optimization methods for personalizing large language models through retrieval augmentation. In International ACM SIGIR Conference on Research and Develop- ment in Information Retrieval , pages 752–762, 2024

    2024

  19. [27]

    Setchi, M

    R. Setchi, M. B. Dehkordi, and J. S. Khan. Explainable robotics in human-robot interactions. Procedia Computer Science , 176:3057– 3066, 2020

    2020

  20. [28]

    Silva, P

    A. Silva, P. Tambwekar, M. Schrum, and M. Gombolay. Towards balancing preference and performance through adaptive personalized explainability. In Proceedings of the ACM/IEEE International Con- ference on Human-Robot Interaction , pages 658–668, 2024

    2024

  21. [29]

    Sobr ´ın-Hidalgo, M

    D. Sobr ´ın-Hidalgo, M. A. Gonz ´alez-Santamarta, ´A. M. Guerrero- Higueras, F. J. Rodr ´ıguez-Lera, and V . Matell ´an-Olivera. Explain- ing autonomy: Enhancing human-robot interaction through expla- nation generation with large language models. arXiv preprint arXiv:2402.04206, 2024

    2024 arXiv

  22. [30]

    U. Soni, S. Sreedharan, and S. Kambhampati. Not all users are the same: Providing personalized explanations for sequential decision making problems. In IEEE/RSJ International Conference on Intelligent Robots and Systems , pages 6240–6247. IEEE, 2021

    2021

  23. [31]

    Sreedharan, T

    S. Sreedharan, T. Chakraborti, and S. Kambhampati. Foundations of explanations as model reconciliation. Artificial Intelligence , 301:103558, 2021

    2021

  24. [32]

    Z. Tan, Z. Liu, and M. Jiang. Personalized pieces: Efficient per- sonalized large language models through collaborative efforts. arXiv preprint arXiv:2406.10471, 2024

    2024 arXiv

  25. [33]

    R. S. Verhagen, M. A. Neerincx, C. Parlar, M. V ogel, and M. L. Tielman. Personalized agent explanations for human-agent teamwork: Adapting explanations to user trust, workload, and performance. In Proceedings of the International Conference on Autonomous Agents and Multiagent...

    2023

  26. [34]

    R. S. Verhagen, M. A. Neerincx, and M. L. Tielman. A two- dimensional explanation framework to classify AI as incomprehen- sible, interpretable, or understandable. In International Workshop on Explainable, Transparent Autonomous Agents and Multi-Agent Systems, pages 119–138. S...

    2021

  27. [35]

    Z. Wang, B. Liang, V . Dhat, Z. Brumbaugh, N. Walker, R. Krishna, and M. Cakmak. I can tell what i am doing: Toward real-world natural language grounding of robot experiences. In Annual Conference on Robot Learning, 2024

    2024

  28. [36]

    Wo ´zniak, B

    S. Wo ´zniak, B. Koptyra, A. Janz, P. Kazienko, and J. Koco ´n. Per- sonalized large language models. arXiv preprint arXiv:2402.09269 , 2024

    2024 arXiv

  29. [37]

    J. Wu, R. Antonova, A. Kan, M. Lepert, A. Zeng, S. Song, J. Bohg, S. Rusinkiewicz, and T. Funkhouser. Tidybot: Personalized robot as- sistance with large language models. Autonomous Robots, 47(8):1087– 1102, 2023

    2023

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.