Pith. sign in

REVIEW 5 major objections 4 minor 37 references

PATHFinder Agent claims a conversational agent can gather a patient's medical and social context, draft an individualized prenatal care plan under PATH guidelines, add community-resource referrals, and leave approval to a clinician.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-02 11:52 UTC pith:64KYOSSF

load-bearing objection A well-motivated system paper for prenatal-care AI whose end-to-end claims outrun its evaluation — the headline 77.6% score measures LLM recommendations, not the agent's dialogue, tools, or resource matching. the 5 major comments →

arxiv 2607.24768 v1 pith:64KYOSSF submitted 2026-06-09 cs.AI cs.CLcs.CYcs.ET

PATHFinder Agent for Tailored Prenatal Care

Vaibhav Balloli , Carissa Samuel , Samia Abdelnabi , Alex Peahl , Elizabeth Bondi-Kelly This is my paper
classification cs.AI cs.CLcs.CYcs.ET
keywords prenatal caretailored care planslarge language model agentsconversational AIsocial determinants of healthclinician oversighttool usematernal health
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper is trying to establish that a single conversational AI agent can replace the disjoint, multi-party intake process of prenatal care: it collects medical and social context, asks targeted follow-up questions, composes a patient-specific care plan following the PATH tailored-care guidelines, and pulls in community resources for social needs, all under clinician oversight. If this works, tailored prenatal care could be offered at scale without adding to clinicians' workload, and patients with housing, food, or transportation needs would leave with concrete referrals rather than just a visit schedule. The paper also claims that among four frontier large language models used as the agent's core, one performs best on expert-curated rubrics across five clinical dimensions, and that recommending antenatal testing is the hardest dimension for every model. That makes the case both for the agent's promise and for the need to keep human review in the loop.

Core claim

The paper's central claim is that PATHFinder Agent, an end-to-end conversational agentic system, can walk a pregnant person through a standardized intake form, run an adaptive dialogue that elicits unmet social needs and preferences, synthesize a draft prenatal care plan that follows PATH's recommendations for visit frequency, services, testing, timing, and care modality, and attach community-resource referrals—then stop for clinician review and editing before the plan is final. On the system's own evaluation, the strongest of four frontier large language models performs best on expert-curated rubrics covering five care dimensions, while antenatal testing recommendations emerge as the weakes

What carries the argument

The carrying mechanism is a four-stage workflow—intake form, dynamic interaction, plan synthesis, and clinician review—orchestrated by a large language model agent with 13 tools in four categories. Medical tools include a trial-of-labor-after-cesarean calculator and a defer-to-clinician action; resource tools implement a hierarchical query interface to a community-resource directory; a personalization tool calls a second large language model to generate context-aware follow-up questions; and report tools incrementally build patient and clinician summaries and a visit schedule. A roughly 14,000-token system prompt embeds the PATH domain knowledge, tool-use instructions, and safety policies, a

Load-bearing premise

The entire claim rests on the assumption that expert-curated rubrics scored by an LLM judge accurately measure clinical correctness; without validation against human clinicians, the reported performance could be a property of the judge rather than of the plans themselves.

What would settle it

Run a blinded panel of practicing obstetricians over the same synthetic profiles and rubrics: if the clinicians' scores fail to agree with the LLM judge's scores beyond a pre-registered threshold, the claim that PATHFinder drafts appropriate care plans is falsified. A complementary check is a small deployment where clinicians reject or substantially edit more than a preset fraction of draft plans.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • A single conversation could replace the current fragmented intake process, so patients only tell their story once instead of repeating it across providers.
  • Social needs become part of the care plan: the system attaches resource referrals for food, housing, transportation, utilities, or clothing automatically.
  • Clinicians keep final authority because the draft plan must be reviewed, edited, or approved in the final stage, making the agent a drafting aid rather than an autonomous decision-maker.
  • The evaluation points to antenatal testing as a persistent weak spot, meaning real deployments would need extra safeguards or tooling around testing recommendations.
  • The same four-stage pattern could be reused for other guideline-based care planning tasks if the conversational and oversight architecture generalizes.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: a deployment study that tracks whether patients actually use the referred community services would test the system's real-world value more directly than rubric scores.
  • Editorial inference: if the same synthetic profiles were graded by a panel of obstetricians instead of an LLM judge, the model ordering could shift, so clinician validation of the evaluation is the most informative next step.
  • Editorial inference: the agent architecture—structured intake, adaptive questioning, tool-based resource matching, human sign-off—might transfer to postpartum care or chronic disease management, where social context similarly shapes the right care plan.
  • Editorial inference: the paper's own results imply that no tested model should be deployed without a mandatory clinician review step, because even the best model struggles with antenatal testing recommendations.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 4 minor

Summary. The paper presents PATHFinder Agent, a four-stage conversational agent for tailored prenatal care aligned with ACOG's PATH guidelines: (1) structured intake form, (2) LLM-driven adaptive dialogue with tool calls, (3) plan/report synthesis, and (4) clinician oversight. The agent is equipped with 13 tools spanning medical calculators, Michigan 211 resource lookup, personalization, and report building. The evaluation compares four frontier LLMs as the core of the agent on synthetic patient profiles, scoring their recommendations against expert-curated rubrics across five clinical dimensions, and reports GPT-5.2 as best with 77.6% average score, with antenatal testing identified as the weakest dimension. The paper frames future work as formal accuracy guarantees, human participant studies, and randomized controlled trials.

Significance. If the system works as described, PATHFinder Agent could operationalize the ACOG PATH guidelines by combining conversational intake, tailored care planning, and social-resource referral, and the co-design with a board-certified OB/GYN and CNM is a genuine strength. The paper also responsibly includes clinician oversight as a design component. However, the current evidence is component-level LLM scoring only; the distinctive agent capabilities—dialogue, tool use, resource matching, and oversight—are not measured. The significance of the reported 77.6% score therefore depends on an unstated assumption that LLM-as-judge scoring on synthetic profiles is a valid proxy for clinical correctness, which is not established in the manuscript.

major comments (5)
  1. [Abstract; §3.3; §3.5, Table 2] The evaluation does not support the 'end-to-end conversational agentic system' claim. Section 3.5 scores only 'frontier models' recommendations' on five clinical rubric dimensions (visit frequency, services, antenatal testing timing, growth ultrasound timing, modality). None of these dimensions measure social-need identification, Michigan 211 resource-retrieval accuracy, multi-turn dialogue quality, tool-call correctness, or the clinician-edit loop described in Stages 2–4 of §3.3. As written, the 77.6% score is a measure of an underlying LLM's static clinical recommendation quality, not of PATHFinder Agent's end-to-end functionality. The distinctive components of the system remain untested.
  2. [§3.5, 'LLM-as-judge scoring' sentence] The judge model, the judge prompt, and validation against human clinical experts are not reported. With LLMs scoring LLM outputs with no inter-rater reliability or clinician benchmark, the rubric scores cannot be interpreted as 'appropriate' care. This is a load-bearing validity gap: without it, the claim that PATHFinder drafts appropriate care plans is unsupported.
  3. [§3.5 vs. Figure 5] The text says the evaluation covers five clinical dimensions, but Figure 5 plots six rubric dimensions: antenatal_testing, delivery_timing, growth_us, modality, services, and visit_frequency. The extra 'delivery_timing' dimension is not listed in §3.5, so the aggregate in Table 2 is ambiguous and cannot be reconstructed from the figure. The mismatch must be resolved, and the aggregate must be recomputed from a clearly stated dimension set.
  4. [§3.5, Table 2 and Figure 5] No number of synthetic profiles, number of evaluation runs, variance, standard deviation, or stratification by condition is reported. Percentages without dispersion or run count cannot establish that GPT-5.2 is meaningfully better than Gemini 2.5 pro. Also, Figure 5 reports mean scores on a 0–5 scale but Table 2 reports percentages on a 0–100 scale, and the conversion is not described. These details are needed to assess the reliability of the headline result.
  5. [§3.3–§3.4] Clinician oversight and the 'Defer to Clinician' tool are claimed as central safety mechanisms, but no evaluation or trace shows whether the oversight loop changes plans, whether deferral fires appropriately, or whether the ~14,000-token system-prompt safety policies are followed. Given the paper's own framing that 'ensuring oversight and grounding' is a priority, this omission weakens the system-level claim of safe PATH-conformant planning.
minor comments (4)
  1. [Figure 2 caption] 'four-stage overflow' should be 'four-stage workflow'.
  2. [Figure 5] Dimension abbreviations (growth_us, delivery_timing) are not defined in the caption; align labels with the dimension names in §3.5.
  3. [§1 and §2] The paper uses both 'g-AMIE' and 'AMIE' inconsistently; standardize to the cited reference [31].
  4. [§3.5] The synthetic profiles and expert-curated rubrics are not described in enough detail to assess coverage or reproducibility. Please include the profile-generation process, rubric content, and scoring criteria, or an appendix/supplement.

Circularity Check

0 steps flagged

No significant circularity; LLM-as-judge is a methodological limitation, not a circular reduction.

full rationale

PATHFinder Agent is a system paper rather than a derivation; there are no equations, fitted parameters, or quantities computed from the system's own outputs. The central evaluation compares frontier LLM recommendations against expert-curated rubrics (Section 3.5). The rubrics operationalize external ACOG/PATH guidelines, which are cited as published standards (refs [19], [24]), not as outputs of this paper. The only notable self-referential element is LLM-as-judge scoring, where an unspecified LLM rates LLM-generated recommendations; however, the paper does not state that the judge is one of the evaluated models, nor does any equation reduce the score to the model's output, so this is a methodological validity limitation rather than a circular reduction. The paper itself flags the need for human validation ('We discuss future validation through human participant studies and randomized controlled trials') and Section 4 states formal accuracy guarantees are future work. Self-citations by author Peahl to ACOG PATH documents are external, published medical guidelines, not unverified premises. No circular step can be exhibited.

Axiom & Free-Parameter Ledger

0 free parameters · 5 axioms · 0 invented entities

The central claims rest on prompt-encoded clinical guidelines, synthetic evaluation, and LLM-as-judge scoring; none of these are externally grounded in human clinical data. No numerical free parameters are fitted; the load-bearing assumptions are domain assumptions about guideline completeness, prompt faithfulness, judge validity, and data representativeness.

axioms (5)
  • domain assumption ACOG PATH guidelines (including visit frequency, testing, modality recommendations) are an appropriate, complete basis for individualized prenatal care planning.
    The agent's knowledge base and evaluation rubrics are constructed from these guidelines; if the guidelines are incomplete or ambiguous, the plan quality claim fails. Invoked throughout Sections 1 and 3.
  • domain assumption The PATH guidelines can be faithfully represented in ~14,000 tokens of system instructions such that an LLM will follow them through tool orchestration.
    The system relies on prompt-based encoding rather than a formal guideline engine; no compliance test is reported. Section 3.3: 'system instructions consists of ∼14,000 tokens'.
  • ad hoc to paper LLM-as-judge scoring against expert rubrics is a valid proxy for clinical correctness of generated plans.
    The evaluation uses LLM judges for five clinical dimensions with no human validation or inter-rater reliability reported. Section 3.5.
  • domain assumption Synthetic patient profiles are sufficiently diverse and representative to expose clinically meaningful model differences.
    The profiles and their construction are not described; no real patient data is used. Section 3.5 first paragraph.
  • domain assumption Michigan 211 data queried by subcategory and ZIP code provides accurate, current resource matches.
    Resource quality is assumed; no evaluation of resource recommendation accuracy is reported. Section 3.3 'Social needs resources data'.

pith-pipeline@v1.3.0-alltime-deepseek · 6592 in / 11546 out tokens · 116919 ms · 2026-08-02T11:52:16.063713+00:00 · methodology

0 comments
read the original abstract

Prenatal care is an important preventive service designed to improve outcomes for pregnant individuals. The American College of Obstetricians and Gynecologists (ACOG) recently introduced guidelines advocating tailored prenatal care, called PATH (Plan for Tailored Healthcare). We present PATHFinder Agent(Planner for Appropriate Tailored Healthcare), an end-to-end conversational agentic system that gathers patient health and social context through structured dialogue, curates individualized prenatal care plans aligned with PATH guidelines, and surfaces community resources from Michigan 211. The system features a four-stage workflow spanning patient intake, dynamic interaction, plan synthesis, and clinician oversight. We evaluate frontier large language models (LLMs) on expert-curated rubrics across five clinical dimensions, finding that GPT-5.2 achieves the highest average score (77.6\%) while identifying key gaps in antenatal testing recommendations. We discuss future validation through human participant studies and randomized controlled trials.

Figures

Figures reproduced from arXiv: 2607.24768 by Alex Peahl, Carissa Samuel, Elizabeth Bondi-Kelly, Samia Abdelnabi, Vaibhav Balloli.

Figure 1
Figure 1. Figure 1: Illustrations of current practices (left) and our proposed solution (right) [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: PATHFinder interfaces. PATHFinder Agent has a four-stage overflow: 1) Patient information intake (standardized form), 2) LLM-generated UI for structured, open-ended dialogue, 3) Draft report review and clarification for the patient and 4) Clinician review during provision of care [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: System architecture. React frontend communicates [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: PATHFinder Agent interfaces. (a)Structured patient intake form; (b) PATHFinder Agent interviews the user to understand additional medical history details, and social determinants of health factors; (c) Patient reviews the draft report for transparency and clarification; (d)Clinician oversight dashboard showing conversation stage, risk flags, and escalation controls. Model Avg. Rubric Score(%) GPT-5.2 77.60… view at source ↗
Figure 5
Figure 5. Figure 5: Per-dimension rubric scores. Antenatal testing and [PITH_FULL_IMAGE:figures/full_fig_p004_5.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

37 extracted references · 8 linked inside Pith

  1. [1]

    E. Balk, V. Danilack, M. Bhuma, et al . 2023. Reduced Compared With Tradi- tional Schedules for Routine Antenatal Visits: A Systematic Review.Obstetrics & Gynecology142, 1 (2023), 8–18

    2023

  2. [2]

    E. Balk, V. Danilack, W. Cao, et al. 2023. Televisits Compared With In-Person Visits for Routine Antenatal Care: A Systematic Review.Obstetrics & Gynecology 142, 1 (2023), 19–29

    2023

  3. [3]

    Victor Barres, Honghua Dong, Soham Ray, Xujie Si, and Karthik Narasimhan. 2025. 𝜏 2-Bench: Evaluating Conversational Agents in a Dual-Control Environment. arXiv:2506.07982 [cs.AI] https://arxiv.org/abs/2506.07982

    Pith/arXiv arXiv 2025

  4. [4]

    Bellerose, M

    M. Bellerose, M. Rodriguez, and P. Vivier. 2022. A systematic review of the qualitative literature on barriers to high-quality prenatal and postpartum care among low-income women.Health Services Research57, 4 (2022), 775–785

    2022

  5. [5]

    Betron, T

    M. Betron, T. McClair, S. Currie, and J. Banerjee. 2018. Expanding the agenda for addressing mistreatment in maternity care: a mapping review and gender analysis.Reproductive Health15, 1 (2018), 143

    2018

  6. [6]

    Paweł Budzianowski, Tsung-Hsien Wen, Bo-Hsiang Tseng, Iñigo Casanueva, Stefan Ultes, Osman Ramadan, and Milica Gašić. 2018. Multiwoz–a large-scale multi-domain wizard-of-oz dataset for task-oriented dialogue modelling.arXiv preprint arXiv:1810.00278(2018)

    Pith/arXiv arXiv 2018

  7. [7]

    Derek Chen, Howard Chen, Yi Yang, Alexander Lin, and Zhou Yu. 2021. Action- based conversations dataset: A corpus for building more in-depth task-oriented dialogue systems. InProceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Tech- nologies. 3002–3017

    2021

  8. [8]

    Izzeddin Gür, Dilek Hakkani-Tür, Gokhan Tür, and Pararth Shah. 2018. User Modeling for Task Oriented Dialogues. In2018 IEEE Spoken Language Technology Workshop (SLT). 900–906. doi:10.1109/SLT.2018.8639652

    arXiv 2018

  9. [9]

    Yixing Jiang, Kameron C Black, Gloria Geng, Danny Park, James Zou, Andrew Y Ng, and Jonathan H Chen. 2025. MedAgentBench: a virtual EHR environment to benchmark medical LLM agents.NEJM AI2, 9 (2025), AIdbp2500144

    2025

  10. [10]

    Shreya Johri, Jaehwan Jeong, Benjamin A Tran, Daniel I Schlessinger, Shannon Wongvibulsin, Zhuo Ran Cai, Roxana Daneshjou, and Pranav Rajpurkar. 2024. CRAFT-MD: A conversational evaluation framework for comprehensive assess- ment of clinical LLMs. InAAAI 2024 Spring Symposium on Clinical Foundation Models

    2024

  11. [11]

    Kilpatrick, L

    S. Kilpatrick, L. Papile, and G. Macones. 2017.Guidelines for Perinatal Care(8th ed.). American Academy of Pediatrics/The American College of Obstetricians and Gynecologists, Elk Grove Village, IL/Washington, D.C

    2017

  12. [12]

    Jiarui Lu, Thomas Holleis, Yizhe Zhang, Bernhard Aumayer, Feng Nan, Haoping Bai, Shuang Ma, Shen Ma, Mengyu Li, Guoli Yin, et al . 2025. Toolsandbox: A stateful, conversational, interactive evaluation benchmark for llm tool use capabilities. InFindings of the Association for Computational Linguistics: NAACL

    2025

  13. [13]

    Mohamoud, E

    Y. Mohamoud, E. Cassidy, E. Fuchs, et al . 2023. Vital Signs: Maternity Care Experiences - United States, April 2023.MMWR Morbidity and Mortality Weekly Report72, 35 (2023), 961–967

    2023

  14. [14]

    2019.Integrating Social Care into the Delivery of Health Care: Moving Upstream to Improve the Nation’s Health

    National Academies of Sciences, Engineering, and Medicine. 2019.Integrating Social Care into the Delivery of Health Care: Moving Upstream to Improve the Nation’s Health. National Academies Press, Washington, D.C

    2019

  15. [15]

    Nijagal, D

    M. Nijagal, D. Patel, C. Lyles, et al. 2021. Using human centered design to identify opportunities for reducing inequities in perinatal care.BMC Health Services Research21, 1 (2021), 714

    2021

  16. [16]

    Anil Palepu, Valentin Liévin, Wei-Hung Weng, Khaled Saab, David Stutz, Yong Cheng, Kavita Kulkarni, S Sara Mahdavi, Joëlle Barral, Dale R Webster, et al

  17. [17]

    Patil, Huanzhi Mao, Charlie Cheng-Jie Ji, Fanjia Yan, Vishnu Suresh, Ion Stoica, and Joseph E

    Shishir G. Patil, Huanzhi Mao, Charlie Cheng-Jie Ji, Fanjia Yan, Vishnu Suresh, Ion Stoica, and Joseph E. Gonzalez. 2025. The Berkeley Function Calling Leaderboard (BFCL): From Tool Use to Agentic Evaluation of Large Language Models. In Forty-second International Conference on Machine Learning

    2025

  18. [18]

    Shishir G Patil, Tianjun Zhang, Xin Wang, and Joseph E Gonzalez. 2023. Gorilla: Large language model connected with massive apis, 2023.URL https://arxiv. org/abs/2305.15334(2023)

    Pith/arXiv arXiv 2023

  19. [19]

    Alex Peahl, Julia C Phillippi, and Mark A Turrentine. 2025. Tailored prenatal care delivery for pregnant individuals.OBSTETRICS AND GYNECOLOGY145, 5 (2025), 565–577

    2025

  20. [20]

    A. F. Peahl, R. Gourevitch, E. Luo, et al. 2020. Right-Sizing Prenatal Care to Meet Patients’ Needs and Improve Maternity Care Value.Obstetrics & Gynecology135, 5 (2020), 1027–1037

    2020

  21. [21]

    A. F. Peahl, A. Novara, M. Heisler, V. K. Dalton, M. H. Moniz, and R. D. Smith

  22. [22]

    A. F. Peahl, A. Powell, H. Berlin, et al. 2021. Patient and provider perspectives of a new prenatal care model introduced in response to the coronavirus disease 2019 pandemic.American Journal of Obstetrics and Gynecology224, 4 (2021), 384.e1–384.e11

    2021

  23. [23]

    Alex F Peahl, Roger D Smith, and Michelle H Moniz. 2020. Prenatal care redesign: creating flexible maternity care models through virtual care.American journal of obstetrics and gynecology223, 3 (2020), 389–e1

    2020

  24. [24]

    A. F. Peahl, C. Zahn, M. Turrentine, et al. 2021. The Michigan Plan for Appropriate Tailored Health Care in Pregnancy Prenatal Care Recommendations.Obstetrics & Gynecology138, 4 (2021), 593–602

    2021

  25. [25]

    Yujia Qin, Shihao Liang, Yining Ye, Kunlun Zhu, Lan Yan, Yaxi Lu, Yankai Lin, Xin Cong, Xiangru Tang, Bill Qian, et al. 2023. Toolllm: Facilitating large language models to master 16000+ real-world apis.arXiv preprint arXiv:2307.16789(2023)

    Pith/arXiv arXiv 2023

  26. [26]

    Khaled Saab, Jan Freyberg, Chunjong Park, Tim Strother, Yong Cheng, Wei-Hung Weng, David GT Barrett, David Stutz, Nenad Tomasev, Anil Palepu, et al. 2025. Advancing Conversational Diagnostic AI with Multimodal Reasoning.arXiv preprint arXiv:2505.04653(2025)

    Pith/arXiv arXiv 2025

  27. [27]

    Semega, M

    J. Semega, M. Kollar, J. Creamer, and A. Mohanty. 2021.Income and Poverty in the United States: 2018. Technical Report. U.S. Census Bureau

    2021

  28. [28]

    Masami Tabata-Kelly, Xiaochu Hu, Michael J Dill, Philip M Alberti, Karen Bullock, William Crown, Malika Fair, Peter May, Pilar Ortega, and Jennifer Perloff. 2024. Physician engagement in addressing health-related social needs and burnout. JAMA network open7, 12 (2024), e2452152–e2452152

    2024

  29. [29]

    Trost, J

    S. Trost, J. Beauregard, G. Chandra, et al . 2022. Pregnancy-Related Deaths: Data from Maternal Mortality Review Committees in 36 US States, 2017–2019. https://www.cdc.gov/reproductivehealth/maternalmortality/erase- mm/data-mmrc.html

    2022

  30. [30]

    Turrentine

    M. Turrentine. 2023. Prenatal Care Visit Frequency: How Much Is Too Much, and How Little Is Too Little?Obstetrics & Gynecology142, 1 (2023), 6–7

    2023

  31. [31]

    Elahe Vedadi, David Barrett, Natalie Harris, Ellery Wulczyn, Shashir Reddy, Roma Ruparel, Mike Schaekermann, Tim Strother, Ryutaro Tanno, Yash Sharma, et al

  32. [32]

    Ran Xu, Yuchen Zhuang, Yishan Zhong, Yue Yu, Xiangru Tang, Hang Wu, May Dongmei Wang, Peifeng Ruan, Donghan Yang, Tao Wang, et al. 2025. Meda- gentgym: Training llm agents for code-based medical reasoning at scale. InThe Second Workshop on GenAI for Health: Potential, Trust, and Policy Compliance

    2025

  33. [33]

    Shunyu Yao, Noah Shinn, Pedram Razavi, and Karthik Narasimhan. 2024. 𝜏- bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains. arXiv:2406.12045 [cs.AI] https://arxiv.org/abs/2406.12045

    Pith/arXiv arXiv 2024

  34. [34]

    arXiv preprint arXiv:2507.15743(2025)

    Towards physician-centered oversight of conversational diagnostic AI. arXiv preprint arXiv:2507.15743(2025)

    Pith/arXiv arXiv 2025

  35. [37]

    Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik R Narasimhan, and Yuan Cao. 2022. React: Synergizing reasoning and acting in language models. InThe eleventh international conference on learning representations

    2022

  36. [2020]

    Patient Preferences for Prenatal and Postpartum Care Delivery: A Survey of Postpartum Women.Obstetrics & Gynecology135, 5 (2020), 1038–1046

    2020

  37. [2025]

    Towards conversational ai for disease management.arXiv preprint arXiv:2503.06074(2025)

    Pith/arXiv arXiv 2025