Pith. sign in

REVIEW 3 major objections 5 minor 45 references

MedUPS: Towards Diagnostic Assistance in Uncommon Medical Cases with Large Language Models

T0 review · 3 major / 5 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 Mid-stream alignment lifts next-step clinical accuracy up to 11 points

desk verdict Useful new dataset and a clean GRPO-vs-SFT comparison, but the 'alignment over scale' claim is judge-dependent by the paper's own re-scoring (9.2-point swing vs a 5.8-point margin) and stays conditional pending clinician labels. read the letter →

arxiv 2608.01012 v1 pith:QI6WFMWJ submitted 2026-08-02 cs.CL cs.AI

Ofir Ben Shoham , Oriel Perets , Nir Grinberg , Nadav Rappoport This is my paper
classification cs.CLcs.AI
keywords clinicaldecisionsupportlargelanguagemodelsmid-streamreasoningnext-steppredictionreinforcementlearningGRPOuncommoncasesLLM-as-a-judge
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 claims that medical LLM evaluation and training should focus on the next clinical step in a patient's trajectory rather than the final diagnosis. The authors introduce MedUPSQA, 21,874 mid-stream decision points derived from 5,535 real case reports, and MedUPS, an alignment framework that uses GRPO to reward models for predicting what the report records next. Across three backbones, the objective raises next-step accuracy by 6.6 to 11.4 percentage points, and under the paper's judge a 27B aligned model scores above substantially larger frontier models. A fair reader would take away that targeted intermediate supervision can matter more than model scale for uncommon-case reasoning, while noting the authors' own caveat that the metric is judge-produced and not clinician-validated.

What carries the argument

The core mechanism is converting free-text case presentations into chronologically ordered clinical chunks, generating for each prefix a next-step question and a chain-of-thought answer, filtering instances with an LLM judge that checks clinical equivalence to the continuation, and aligning policies with Group Relative Policy Optimization (GRPO) under a four-criterion rubric reward (correct, specific, grounded, complete). This yields dense, trajectory-aware supervision without any final-diagnosis labels.

What would settle it

Take the fixed 500-item evaluation pool and score it with a clinician-validated reference standard, or with a judge family never used in training. If MedUPS-Qwen3.6-27B no longer beats its own base model or GPT-5.6, or if the margin collapses, the central claim is falsified.

Watch

Extended reading notes

Core claim

The central discovery is that dense, label-free supervision at the level of the next action—generated from real case presentations and filtered by an LLM equivalence judge—can be optimized with GRPO to improve mid-stream reasoning on uncommon cases. The paper shows the target carries signal independent of the optimizer (SFT on identical pairs also improves every backbone), that GRPO is the stronger learning signal at the largest tested scale, and that mid-stream alignment moves a 9B model above its unaligned 27B counterpart. The authors state plainly that the reward and primary metric are produced by the same model family and are not validated against clinician labels, and they quantify that

Load-bearing premise

Both the training reward and the evaluation metric come from DeepSeek-family LLM judges that are not validated against clinician labels, so the size of the reported gains and the claim that alignment beats scale are only as strong as the judge's notion of clinical equivalence.

Editorial extensions

If this is right

  • If the central claim holds, medical LLM benchmarks should measure sequential next-step decisions rather than terminal diagnoses.
  • Training on mid-stream decisions could make smaller open models viable for on-premises hospital deployment, easing privacy and audit constraints.
  • The released dataset and checkpoints allow independent, clinician-grounded validation to test whether the gains survive a human-judged reference standard.
  • The paper's error analysis suggests action-question failures are mostly reference under-specification, motivating evaluation with multiple acceptable next steps.
  • The alignment-over-scale result, if it persists under a second judge family, would argue for investing in task-specific alignment rather than model size for uncommon cases.

Reading between the lines

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

  • An implication the paper leaves implicit is that the judge-dependence check was run only on the strongest checkpoint; if a full-table re-scoring shows similar sensitivity, the 'alignment over scale' conclusion could be an artifact of reward hacking rather than a general property of the objective.
  • The same chunk-and-predict framing could transfer to other longitudinal clinical tasks—treatment escalation, discharge planning, ICU management—where the next action is the object of interest.
  • Because the construction filter retains only instances an LLM judge accepts, the training distribution is by construction biased toward judge-recognizable equivalences; this may limit what the models learn about clinically defensible alternatives that the single report did not record.
  • The mixed GRPO-versus-SFT results across backbones suggest the advantage of RL over imitation may depend on base-policy strength; a per-backbone tuned SFT baseline would be needed to separate objective effects from recipe fit.
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 / 5 minor

Summary. The paper introduces MedUPSQA, a dataset of 21,874 mid-stream clinical decision points derived from 5,535 real case reports, and MedUPS, a GRPO-based alignment framework that trains LLMs to predict the next clinical step from an accumulating case prefix. The authors segment case reports into chronological chunks, generate questions and candidate answers with DeepSeek-R1, filter them with an LLM judge, and train with a four-criterion DeepSeek-Chat rubric reward. Evaluation uses a fixed 500-item stratified test pool scored by a binary DeepSeek-Chat equivalence judge. The paper reports that GRPO alignment improves next-step accuracy over the base backbones (e.g., Qwen3.6-27B from 55.2 to 66.7), that SFT on the same pairs also improves all backbones, and that the aligned 27B model outperforms GPT-5.6 variants under the paper's judge. The paper includes a judge-robustness analysis, an error attribution study, and explicit limitations acknowledging judge dependence.

Significance. If the reported gains reflect genuine clinical decision quality, the dataset and the mid-stream alignment objective would be a valuable contribution to medical LLM research, addressing a real gap between terminal-diagnosis benchmarks and the sequential, information-limited nature of clinical care. The paper is commendably transparent: it releases data, code, and checkpoints, reports bootstrap CIs, compares GRPO against SFT on identical pairs, and explicitly quantifies judge disagreement. However, the central 'alignment over scale' claim rests on an LLM judge from the same model family as the training reward and the data-construction filter, and the paper's own alternate-judge rescoring shows a 9.2-point swing that exceeds the claimed 5.8-point margin over a frontier model. The contribution is therefore promising but not yet established as a statement about clinical competence or model ordering.

major comments (3)
  1. [§5, 'How much depends on the judge'] The paper reports that rescoring MedUPS-Qwen3.6-27B with GPT-5.6 lowers accuracy from 67.2 to 58.0, a 9.2-point change larger than the 5.8-point margin by which this model leads GPT-5.6-sol under the DeepSeek judge. Only the strongest checkpoint was rescored, so the full Table 3 ordering under a second judge family is unknown. The abstract claim that 'smaller models surpassing larger, frontier models we evaluate' is therefore not supported by the evidence. The paper explicitly says 'we do not claim that the open-versus-frontier ordering is judge-independent,' but the abstract and introduction still present the scale comparison without the same caveat. This load-bearing claim needs to be either supported by rescoring all models in Table 3 or downgraded to a judge-conditional observation.
  2. [§3.3, §4, and Limitations] The GRPO reward and the primary evaluation metric are produced by the same judge family (DeepSeek-Chat), and the construction-time filter also uses a DeepSeek-family judge. Thus the retained supervision is, by construction, what a DeepSeek-family judge considers equivalent to the reference continuation, and an aligned policy can increase its score by conforming to that judge's equivalence notion without changing clinical decisions. The 30-instance manual prompt-development check is explicitly not an independent validation. This design makes it impossible to interpret the reported gains as clinical improvement without external validation. The claim that 'the target carries signal independently of the optimizer' is weakened because the target itself is filtered by the same judge family. A clinician-scored subset, even a few hundred items, is needed to anchor the metric.
  3. [§5, 'Error analysis'] The error attribution is performed by an LLM rater instructed to 'resolve doubt against calling a failure clinical,' so the resulting 2.0% category-E rate is a lower bound by construction and is not clinician-verified. The paper uses this analysis to argue that the headline accuracies 'understate clinical competence,' but that inference is not supported: an unvalidated rater cannot establish that the LLM judge's verdicts are conservative. This is important because it is used to soften the judge-dependence concern. External clinician review of a stratified failure sample would be needed to support this interpretation.
minor comments (5)
  1. [Abstract] Typo: 'target framwork carries signal' should be 'target framework carries signal.' Also, 'with 95% CI' in the abstract is vague; the CIs are bootstrap intervals over a fixed 500-item pool, not over training runs, as correctly explained in Section 4.
  2. [Appendix C, Figure A3] The prompt text contains typos ('wether', 'wich') that are acknowledged as reproduced verbatim. This is acceptable, but please note in a caption that typos are intentional and not OCR artifacts, since readers may otherwise file correction requests.
  3. [Limitations] Typo: 'strenghten' should be 'strengthen.' Also, the sentence beginning 'however, due to the open, worldwide access...' is grammatically broken and should be revised for clarity.
  4. [§4, Evaluation protocol] The bootstrap procedure uses only 5 resamples (seeds 1001–1005). This is a small number; 1000 would give more stable intervals. Please justify or increase the number of resamples.
  5. [§3.2] The dataset construction uses GPT-4o for chunking and DeepSeek-R1 for generation. Please clarify whether the chunking and generation prompts are released with the dataset, as stated, and specify the exact model versions/dates for reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: the training reward, construction-time filter, and evaluation judge are distinct operational stages, and the paper's own limitations section disclaims judge-independence rather than relying on it.

full rationale

The derivation chain is not circular. MedUPSQA is built by chronological chunking of case reports, DeepSeek-R1 question/answer generation, and a construction-time equivalence filter; GRPO optimizes the four-criterion rubric reward (Eq. 1, weights 3,1,2,1) from DeepSeek-Chat; accuracy is measured with a separate, stricter binary equivalence judge (Appendix C, Figure A2) on a held-out 500-item pool drawn from a case-disjoint test split. These are not the same function: the reward is a weighted multicriteria score, the evaluation is a binary verdict, and the test instances were never used for training or checkpoint selection. No parameter is fitted to the evaluation pool and then reported as a prediction; the SFT baselines on identical pairs are a controlled contrast, and the data-filter/reward/evaluation overlap is explicitly presented as a limitation ('Our reward and our primary metric are produced by language models, and neither is validated against clinician labels', Section 6; 'Every stage of dataset construction, filtering, reward assignment, evaluation, and error attribution runs through LLMs, so the retained supervision is by definition what an automated judge considered equivalent'). The self-citations (CUPCase, Cpllm) point to external corpora and prior methods, not to unverified results that carry the argument. The judge-family overlap is a genuine validity threat for the clinical and cross-scale conclusions, and the paper quantifies it (0.672 to 0.580 under GPT-5.6, Section 5), but a threat to external validity is not circularity by construction: the reported claim is explicitly conditional on the evaluation judge, and the paper disclaims judge-independence of the open-vs-frontier ordering. Hence no enumerated circularity step is present.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central empirical claims rest on a self-referential evaluation loop: an LLM judge provides both the training reward and the primary metric, and the paper does not supply clinician labels. The free parameters are the judge identity, its rubric weights, and the sampling design. The domain assumptions about proxy validity are explicitly flagged by the authors.

free parameters (2)
  • Rubric weights = (3,1,2,1)
    Weights for Correct, Specific, Grounded, Complete criteria in the GRPO reward (Eq. 1); chosen a priori, not ablated, sensitivity unknown (Section 3.3).
  • Judge model and temperature = DeepSeek-Chat at 0.1
    The reward and evaluation judge family is a free choice that affects all reported accuracies; the paper shows a 9.2-point shift under an alternate judge.
assumptions (4)
  • domain assumption Next-step prediction is a valid proxy for clinical decision quality
    The paper evaluates models on predicting the next step recorded in a single case report, which may be one of several defensible actions (Section 6 limitations).
  • domain assumption LLM judge equivalence reflects clinical correctness
    No clinician labels exist; the paper states 'we cannot establish which judge is closer to clinical assessment' (Section 4).
  • domain assumption Automated chunking and generation preserve clinical trajectory structure
    Chunking and question generation are fully automated and 'not verified by clinicians' (Section 3.2).
  • standard math Bootstrap resampling adequately quantifies uncertainty
    The reported 95% CI covers only sampling variation within the evaluation pool, not training or decoding variance (Section 4).

how reviews work

0 comments
Cite this review

Pith. "Pith review of MedUPS: Towards Diagnostic Assistance in Uncommon Medical Cases with Large Language Models." pith.science (2026). https://pith.science/paper/QI6WFMWJ

@misc{pith2026260801012,
  author       = {Pith},
  title        = {Pith review of: MedUPS: Towards Diagnostic Assistance in Uncommon Medical Cases with Large Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QI6WFMWJ}},
  note         = {Machine review of arXiv:2608.01012}
}
read the original abstract

Uncommon and off-guideline cases are difficult for clinical decision support, because physicians must make a series of management decisions under diagnostic uncertainty and rarely see the full case at once. Most large language model (LLM) benchmarks for medicine score only the final diagnosis, yet much of clinical care turns on the next appropriate action: the next test to order, the imaging study to obtain, the specialist to involve, or the differential to pursue. We introduce MedUPSQA, a dataset of 21,874 mid-stream clinical decision points built from 5,535 real case reports, and MedUPS, an alignment framework that supervises models on these intermediate decisions as they unfold along a patient's trajectory. We segment free-text case presentations into chronologically ordered, accumulating clinical chunks and align models to predict the next step with reinforcement learning (GRPO), using an external LLM-as-a-Judge reward. This objective mirrors how clinicians actually meet patients, reasoning forward from accumulating evidence toward the next decision, rather than committing to a final label. Across three backbones, mid-stream alignment raises next-step accuracy from 55.2 to 66.7 for Qwen3.6-27B, from 47.2 to 57.8 for Qwen3.5-9B, and from 37.8 to 44.4 for HuatuoGPT-3-8B, with 95% CI. In several model scales we test the objective improves accuracy more than scale, with smaller models surpassing larger, frontier models we evaluate. We further train supervised fine-tuning (SFT) baselines on the mid-stream task, SFT improves all backbones above base, indicating the target framwork carries signal independently of the optimizer. We release the dataset, code, and aligned checkpoints.

Figures

Figures reproduced from arXiv: 2608.01012 by the authors.

Figure 1
Figure 1. MedUPSQA construction and mid-stream alignment. Case reports are segmented into chronologically ordered chunks c1, . . . , cT . For each prefix, a question qi answerable from ci+1 is generated with an answer ai and rationale ri , and a judge keeps only instances clinically equivalent to ci+1. The retained instances align policy models with GRPO. Every stage of construction is automated; no clinician adjudication is … view at source ↗
Figure 2
Figure 2. Mid-stream accuracy against model scale (0- shot). Circles mark base models, triangles their MedUPS￾aligned counterparts, and arrows the alignment gain. Open models (green) sit on the parameter axis. Frontier models (blue) appear in a separate panel sharing only the accuracy axis: their parameter counts are undisclosed, so no horizon￾tal position should be read from them. Error bars are 95% bootstrap CIs over the ev… view at source ↗
Figure 3
Figure 3. Few-shot ablation. Mid-stream accuracy against the number of answer-only in-context demonstrations. Each aligned checkpoint stays above its own backbone at every shot count. tems on a separate axis because their parameter counts are undisclosed. This comparison is also the one most exposed to judge identity, which we quantify next. In-context demonstrations do not substitute for align￾ment. The 5-shot column of [PI… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

45 extracted references · 19 canonical work pages

  1. [1]

    Communication, Simulation, and Intelligent Agents: Implications of Personal Intelligent Machines for Medical Education

    Clancey, William J. Communication, Simulation, and Intelligent Agents: Implications of Personal Intelligent Machines for Medical Education. Proceedings of the Eighth International Joint Conference on Artificial Intelligence (IJCAI-83)

  2. [2]

    Classification Problem Solving

    Clancey, William J. Classification Problem Solving. Proceedings of the Fourth National Conference on Artificial Intelligence

  3. [3]

    , title =

    Robinson, Arthur L. , title =. 1980 , doi =. https://science.sciencemag.org/content/208/4447/1019.full.pdf , journal =

    1980

  4. [4]

    New Ways to Make Microcircuits Smaller---Duplicate Entry

    Robinson, Arthur L. New Ways to Make Microcircuits Smaller---Duplicate Entry. Science

  5. [5]

    Clancey and Glenn Rennels , abstract =

    Diane Warner Hasling and William J. Clancey and Glenn Rennels , abstract =. Strategic explanations for a diagnostic consultation system , journal =. 1984 , issn =. doi:https://doi.org/10.1016/S0020-7373(84)80003-6 , url =

    doi:10.1016/s0020-7373(84)80003-6 1984

  6. [6]

    and Rennels, Glenn R

    Hasling, Diane Warner and Clancey, William J. and Rennels, Glenn R. and Test, Thomas. Strategic Explanations in Consultation---Duplicate. The International Journal of Man-Machine Studies

  7. [7]

    Poligon: A System for Parallel Problem Solving

    Rice, James. Poligon: A System for Parallel Problem Solving

  8. [8]

    Transfer of Rule-Based Expertise through a Tutorial Dialogue

    Clancey, William J. Transfer of Rule-Based Expertise through a Tutorial Dialogue

Show all 45 references
  1. [9]

    The Engineering of Qualitative Models

    Clancey, William J. The Engineering of Qualitative Models

  2. [10]

    2023 , eprint=

    Attention Is All You Need , author=. 2023 , eprint=

    2023

  3. [11]

    Pluto: The 'Other' Red Planet

    NASA. Pluto: The 'Other' Red Planet

  4. [12]

    Nature Medicine , year =

    Toward expert-level medical question answering with large language models , author =. Nature Medicine , year =. doi:10.1038/s41591-024-03423-7 , url =

    doi

  5. [13]

    European Journal of Radiology , year =

    Aligning large language models with radiologists by reinforcement learning from AI feedback for chest CT reports , author =. European Journal of Radiology , year =

  6. [14]

    MMedPO: Aligning Medical Vision

    Zhu, Kangyu and Xia, Peng and Li, Yun and Zhu, Hongtu and Wang, Sheng and Yao, Huaxiu , journal =. MMedPO: Aligning Medical Vision. 2024 , url =

    2024

  7. [15]

    MMedPO: Aligning Medical Vision

    Zhu, Kangyu and Xia, Peng and Li, Yun and Zhu, Hongtu and Wang, Sheng and Yao, Huaxiu , booktitle =. MMedPO: Aligning Medical Vision. 2025 , note =

    2025

  8. [16]

    Beyond Distillation: Pushing the Limits of Medical LLM Reasoning with Minimalist Rule

    Liu, Che and Wang, Haozhe and Pan, Jiazhen and Wan, Zhongwei and Dai, Yong and Lin, Fangzhen and Bai, Wenjia and Rueckert, Daniel and Arcucci, Rossella , journal =. Beyond Distillation: Pushing the Limits of Medical LLM Reasoning with Minimalist Rule. 2025 , url =

    2025

  9. [17]

    Towards Adapting Open

    Wang, Hanyin and Gao, Chufan and Liu, Bolun and Xu, Qiping and Hussein, Guleid and El Labban, Mohamad and Iheasirim, Kingsley and Korsapati, Hariprasad and Sun, Jimeng , journal =. Towards Adapting Open. 2024 , url =

    2024

  10. [18]

    arXiv preprint arXiv:2502.14860 , year =

    Aligning LLMs to Ask Good Questions: A Case Study in Clinical Reasoning , author =. arXiv preprint arXiv:2502.14860 , year =

    arXiv

  11. [19]

    Reason Like a Radiologist: Chain

    Jing, Peiyuan and Lee, Kinhei and Zhang, Zhenxuan and Zhou, Huichi and Yuan, Zhengqing and Gao, Zhifan and Zhu, Lei and Papanastasiou, Giorgos and Fang, Yingying and Yang, Guang , journal =. Reason Like a Radiologist: Chain. 2025 , doi =

    2025

  12. [20]

    Dai, Wei and Chen, Peilin and Ekbote, Chanakya and Liang, Paul Pu , journal =. QoQ. 2025 , url =

    2025

  13. [21]

    Proceedings of the AAAI Conference on Artificial Intelligence , number=

    CUPCase: Clinically Uncommon Patient Cases and Diagnoses Dataset , author=. Proceedings of the AAAI Conference on Artificial Intelligence , number=

  14. [22]

    Computers in Biology and Medicine , volume=

    MedConceptsQA: Open source medical concepts QA benchmark , author=. Computers in Biology and Medicine , volume=. 2024 , publisher=

    2024

  15. [23]

    arXiv preprint arXiv:2502.19655 , year=

    Med-rlvr: Emerging medical reasoning from a 3b base model via reinforcement learning , author=. arXiv preprint arXiv:2502.19655 , year=

    arXiv

  16. [24]

    arXiv preprint arXiv:2503.23829 , year=

    Crossing the Reward Bridge: Expanding RL with Verifiable Rewards Across Diverse Domains , author=. arXiv preprint arXiv:2503.23829 , year=

    arXiv

  17. [25]

    Nature Communications , volume=

    Quantifying the reasoning abilities of LLMs on clinical cases , author=. Nature Communications , volume=. 2025 , publisher=

    2025

  18. [26]

    arXiv preprint arXiv:2505.11733 , year=

    MedCaseReasoning: Evaluating and learning diagnostic reasoning from clinical case reports , author=. arXiv preprint arXiv:2505.11733 , year=

    arXiv

  19. [27]

    arXiv preprint arXiv:2412.18925 , year=

    Huatuogpt-o1, towards medical complex reasoning with llms , author=. arXiv preprint arXiv:2412.18925 , year=

    arXiv

  20. [28]

    arXiv preprint arXiv:2504.00993 , year=

    Medreason: Eliciting factual medical reasoning steps in llms via knowledge graphs , author=. arXiv preprint arXiv:2504.00993 , year=

    arXiv

  21. [29]

    PLOS Digital Health , volume=

    Cpllm: Clinical prediction with large language models , author=. PLOS Digital Health , volume=. 2024 , publisher=

    2024

  22. [30]

    arXiv preprint arXiv:2307.08922 , year=

    Large language models perform diagnostic reasoning , author=. arXiv preprint arXiv:2307.08922 , year=

    arXiv

  23. [31]

    Patterns , volume=

    Can large language models reason about medical questions? , author=. Patterns , volume=. 2024 , publisher=

    2024

  24. [32]

    Japanese Journal of Radiology , volume=

    Structured clinical reasoning prompt enhances LLM’s diagnostic capabilities in diagnosis please quiz cases , author=. Japanese Journal of Radiology , volume=. 2025 , publisher=

    2025

  25. [33]

    arXiv preprint arXiv:1707.06347 , year=

    Proximal policy optimization algorithms , author=. arXiv preprint arXiv:1707.06347 , year=

    arXiv

  26. [34]

    Nature medicine , volume=

    Toward expert-level medical question answering with large language models , author=. Nature medicine , volume=. 2025 , publisher=

    2025

  27. [35]

    2015 , publisher=

    Improving diagnosis in health care , author=. 2015 , publisher=

    2015

  28. [36]

    European Journal of Human Genetics , volume=

    Time to diagnosis and determinants of diagnostic delays of people living with a rare disease: results of a Rare Barometer retrospective patient survey , author=. European Journal of Human Genetics , volume=. 2024 , publisher=

    2024

  29. [37]

    New England Journal of Medicine , volume=

    Educational strategies to promote clinical diagnostic reasoning , author=. New England Journal of Medicine , volume=. 2006 , publisher=

    2006

  30. [38]

    arXiv preprint arXiv:2402.03300 , year=

    Deepseekmath: Pushing the limits of mathematical reasoning in open language models , author=. arXiv preprint arXiv:2402.03300 , year=

    arXiv

  31. [39]

    Educational and psychological measurement , volume=

    A coefficient of agreement for nominal scales , author=. Educational and psychological measurement , volume=. 1960 , publisher=

    1960

  32. [40]

    biometrics , pages=

    The measurement of observer agreement for categorical data , author=. biometrics , pages=. 1977 , publisher=

    1977

  33. [41]

    and Pierson, Emma and Koh, Pang Wei and Tsvetkov, Yulia , booktitle =

    Li, Shuyue Stella and Balachandran, Vidhisha and Feng, Shangbin and Ilgen, Jonathan S. and Pierson, Emma and Koh, Pang Wei and Tsvetkov, Yulia , booktitle =

  34. [42]

    Fansi Tchango, Arsene and Goel, Rishab and Wen, Zhi and Martel, Julien and Ghosn, Joumana , booktitle =

  35. [43]

    Nature Medicine , volume =

    The Artificial Intelligence Clinician Learns Optimal Treatment Strategies for Sepsis in Intensive Care , author =. Nature Medicine , volume =

  36. [44]

    Schmidgall, Samuel and Ziaei, Rojin and Harris, Carl and Reis, Eduardo and Jopling, Jeffrey and Moor, Michael , journal =

  37. [45]

    Nature Medicine , volume =

    An Evaluation Framework for Clinical Use of Large Language Models in Patient Interaction Tasks , author =. Nature Medicine , volume =

Pith tools

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