Pith. sign in

REVIEW 5 major objections 7 minor 18 references

Leaps Beyond the Seen: Reinforced Reasoning Augmented Generation for Clinical Notes

T0 review · 5 major / 7 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A reinforcement-trained retriever that leaps between semantic clusters in a medical knowledge graph can draft discharge instructions from pre-admission data alone, improving clinical recall by 12% over a vanilla LLM and by 6% over the…

desk verdict The task is new and the RL retrieval recipe is worth studying, but Table 2's Jaccard numbers are inconsistent with its own P/R/F1, so the headline empirical claim is not yet supported. read the letter →

arxiv 2506.05386 v3 pith:TDXMSYZZ submitted 2025-06-03 cs.CL cs.AI

Lo Pang-Yun Ting , Chengshuai Zhao , Yu-Hua Zeng , Yuan Jee Lim , Kun-Ta Chuang , Huan Liu This is my paper
classification cs.CLcs.AI
keywords clinicalnotegenerationdischargeinstructionsretrieval-augmentedmedicalknowledgegraphreinforcementlearningUMLSreasoningleapspre-admissioninformation
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

The paper takes on a specific clinical writing problem: generating a patient's discharge instructions from only the information available before admission, such as allergies, chief complaint, and history of present illness. It argues that ordinary LLMs and standard retrieval-augmented methods fail at this because they stick close to the input's wording, while discharge instructions depend on events and decisions from the hospital stay that are absent from pre-admission data. To close that gap, ReinRAG trains a retriever with reinforcement learning to walk paths through a medical knowledge graph, sometimes leaping between broad semantic clusters such as Disorders and Procedures. The paper's central claim is that these leaping reasoning paths give the LLM the semantic guidance it needs, lifting recall of correct medical concepts by at least 12% over a vanilla LLM and by at least 6% over the best competing baseline while keeping precision comparable. If true, this means an LLM could produce a useful early draft of discharge instructions at admission time, before in-hospital records accumulate.

What carries the argument

The load-bearing component is the reinforcement-learning-trained two-level retrieval policy over a medical knowledge graph whose concepts are organized into semantic groups. The policy network sees a cluster state that appends the embedding of the currently selected semantic cluster with the embedding of the most under-represented cluster among the patient's pre-admission keywords, plus a concept state that averages the embeddings of all explored concepts; it then scores possible next-cluster actions. A concept-level step connects the new cluster to the existing path at its most similar concept and extends the path to a neighbor concept that balances similarity to the patient's keywords and to the path so far. This machinery makes 'reasoning leaps' a learnable action: the retriever learns when to jump to a distant semantic cluster, and a group-normalized episodic reward, based on overlap with ground-truth discharge concepts and embedding similarity to them, tells it whether the whole path helped the LLM produce correct long-form content.

What would settle it

Replace ReinRAG's retrieved paths on the 1,000 test patients with random knowledge-graph paths matched for length and number of cluster jumps. If concept recall and F1 do not fall, the improvement is attributable to added context, not to learned reasoning leaps; if performance collapses, the leaps are doing the claimed work.

Watch

Extended reading notes

Core claim

ReinRAG's central claim is that the information gap between sparse pre-admission data and complete discharge instructions can be closed by adaptive retrieval that controls reasoning granularity. The retriever operates at two levels: a policy network scores which semantic cluster to visit next, and a concept-level step then selects concrete concepts inside that cluster, inserting an explicit 'cluster leap' marker when the path moves to a new cluster. The state fed to the policy includes the currently selected cluster and the most under-represented cluster among the patient's pre-admission keywords, so the model can decide to stay in familiar semantics or jump to a sparse one. The retrieved paths are passed to a 7B LLM as explicit reasoning guidance. Training uses a group-based optimization in which several paths are retrieved per patient, each path receives a reward combining overlap with ground-truth discharge concepts and embedding similarity to those concepts, and rewards are normalized within the group before a policy-gradient update. On a 1,000-patient held-out test set from a real-world critical-care note collection, the paper reports that ReinRAG outperforms all compared baselines on both clinical-efficacy and language-generation metrics, with human evaluators rating its outputs as more focused and less prone to irrelevant or incorrect medical content.

Load-bearing premise

The load-bearing premise is that jumping between broad semantic categories in the medical knowledge graph, such as from a symptom category to a treatment category, mirrors how a clinician actually reasons about a patient, so the retrieved paths are safe guides to diagnoses and treatments rather than arbitrary graph associations.

Editorial extensions

If this is right

  • Discharge-instruction drafting can start at admission: with only allergies, chief complaint, and history of present illness, ReinRAG produces instructions whose medical-concept recall beats a vanilla 7B LLM by at least 12% and the best tested baseline by at least 6%.
  • The retrieved paths cut irrelevant content: the count of generated medical concepts that miss the ground-truth instructions drops by at least 12% versus a vanilla LLM and 5% versus the best baseline tested, and clinician reviewers noted fewer inappropriate medications and diagnostic errors.
  • Reasoning leaps are the load-bearing design choice: removing leaps degrades both clinical-efficacy and language-generation scores substantially more than removing exploration.
  • The gains appear exactly where the information gap is widest: recall improves most in clusters such as Anatomy, Procedures, Physiology, and Phenomena, while clusters already well covered by the input change little.

Reading between the lines

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

  • Beyond the paper: the same group-normalized reward over ground-truth content could be recycled for other sparse-input clinical writing tasks, such as pre-operative plans or early consult notes, by changing only the target document type.
  • Beyond the paper: replacing the fixed path length with an adaptive stopping rule is a natural next test; the paper flags fixed-length retrieval as a limitation, and uneven patient complexity suggests varied lengths could improve both precision and coverage.
  • Beyond the paper: the 'reasoning leap' recipe is not medical-specific; any domain with a concept graph and a sparse-to-dense generation gap (legal, maintenance, engineering) could use the same two-level policy to decide when to jump to an underrepresented concept group.
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

5 major / 7 minor

Summary. The paper proposes ReinRAG, a retrieval-augmented generation framework for generating long-form discharge instructions from pre-admission patient information (allergies, chief complaint, and history of present illness). ReinRAG trains a two-level retrieval policy over the UMLS knowledge graph: a semantic-level policy decides whether to stay in or leap across UMLS semantic groups, and a concept-level policy selects neighboring concepts to extend reasoning paths. The retriever is trained with GRO, a GRPO-inspired group-normalized REINFORCE objective, using rewards based on overlap with ground-truth discharge concepts. The retrieved reasoning paths are then prompted into an LLM (Mistral-7B-Instruct-v0.3 in the main results). Experiments on 4,000 MIMIC-IV-note summaries compare against vanilla LLMs, medical LLMs, and retrieval baselines using clinical efficacy (precision, recall, F1, Jaccard, Hamming loss) and NLG metrics, along with ablations, parameter sensitivity analyses, and a small human evaluation.

Significance. The task of generating discharge instructions from pre-admission data alone is practically relevant and underexplored, and the idea of RL-controlled reasoning granularity over a medical knowledge graph with group-normalized rewards is interesting. The paper includes useful ablations, parameter sensitivity experiments, and a human evaluation, even if the latter is small. However, the main quantitative evidence is compromised by internally inconsistent clinical-efficacy metrics in Table 2, missing statistical uncertainty throughout, and a reward/evaluation alignment that makes the reported CE gains unsurprising. If the authors correct the metrics, clarify the data split, and provide uncertainty quantification, the contributions could be of value to clinical NLP; as reported, the central empirical claims are not verifiable.

major comments (5)
  1. [Table 2 and §4.2] The CE metrics in Table 2 are internally inconsistent. For any two concept sets, Jaccard J = P·R/(P+R−P·R) and F1 = 2J/(1+J). In the ReinRAG concept-level row, P=99.20%, R=40.73% imply J≈34.4%, not 6.42%, and J=6.42% implies F1≈12.1%, not 56.01%. The same order-of-magnitude mismatch appears in every row (e.g., N-gram LLaMA: J=5.77 vs implied ≈23.5%). Since Section 4.2's headline claims (≥12% and ≥6% improvements in recall and F1) rest on these numbers, the empirical core of the paper is not verifiable as reported. Please recompute from raw concept sets, define whether J/HL are macro- or micro-averaged, and ensure the identity F1=2J/(1+J) holds per row.
  2. [§4.1–§4.6] No error bars, confidence intervals, or significance tests are reported for any automatic metric or human evaluation. With a test set of 1,000 and differences of 6–12 percentage points, paired tests (e.g., bootstrap or Wilcoxon) are needed to establish that the gains are not noise. The human evaluation uses only 20 cases and two professionals with no inter-annotator agreement; the distribution plots in Figure 7 are descriptive. Please add uncertainty quantification and, for the human study, agreement statistics.
  3. [§3.3 and §4.1] The retrieval reward in Eq. (7) directly maximizes overlap with ground-truth discharge concepts (the indicator sum and λ·sim(P, Ĉ)), while the CE evaluation metrics measure overlap between generated and ground-truth discharge concepts. The retriever is thus optimized on the same gold concept sets used for evaluation. Although this is not full circularity because evaluations are on held-out patients, it means the CE gains are an expected consequence of supervised reward design rather than evidence of 'reasoning.' Please report additional analyses that separate the contribution of direct concept matching from the leap mechanism (e.g., retrieval-only recall of held-out GT concepts, or an ablation that removes the GT-concept term from the reward).
  4. [§3.2 and §4.2] The claim that cross-cluster 'reasoning leaps' in UMLS semantic groups correspond to clinically meaningful inference (Sec. 4.2, 'Effectiveness of Reasoning Leaps') is not validated. The paper introduces a 'cluster leap' pseudo-relation and allows transitions between UMLS semantic groups, but provides no clinician assessment of whether the retrieved paths are sound diagnostic or therapeutic reasoning. The human evaluation in Sec. 4.6 is qualitative and limited to 20 cases. Please include a path-level validation (e.g., clinician ratings of retrieved paths or an error analysis of leaps) to support the mechanism's interpretation.
  5. [§4.1] The train/test split is described as selecting 4,000 summaries with 3,000 for training and 1,000 for testing, but it is not stated whether the split is at the patient level. MIMIC-IV-note contains multiple notes per patient; if summaries from the same patient appear in both training and test, the retriever could memorize patient-specific concepts and inflate CE. Please clarify the split and, if necessary, re-run at the patient level.
minor comments (7)
  1. [Section 1] In the paragraph beginning 'Patient discharge instructionsummarizes', there is a missing space ('instructionsummarizes') and a duplicated citation to Kononenko (2001) in the same sentence.
  2. [Table 2/3 captions] The ∆ columns are ambiguous in sign; for example, '97.20(-1.6)' appears to mean that ReinRAG is 1.6 points higher than the baseline, but the negative sign is confusing. Please define in the caption whether ∆ = baseline − ReinRAG or ReinRAG − baseline.
  3. [Section 4.2] In the baselines list, 'Retrieved-Based Methods' should read 'Retrieval-Based Methods'.
  4. [Section 4.6] Typographical errors: 'medical processionals' should be 'medical professionals', and 'or event incorrect instructions' should be 'or even incorrect instructions'.
  5. [Figure 3] The axes are not clearly labeled: the CE panels show only F1 and HL without indicating which bars correspond to which G or NP values, and the N-gram/Concept distinction is not marked on the x-axis. Please add a legend and axis titles.
  6. [Eq. (7)] The sum over concepts in P counts repeated concepts multiple times if duplicates appear in a path; please clarify whether concepts are deduplicated before computing the reward.
  7. [Appendix A.1] The discount factor γ=0.1 is unusually low for a 5-step episodic task; please justify this choice or report its sensitivity.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the RL retriever is trained on training-set ground-truth concepts and evaluated on held-out test-set generated text, so the central comparison is a genuine generalization test; the reported Table 2 metric inconsistency is a correctness concern, not a circularity.

full rationale

The reward in Eq. 7 (binary reward for path concepts appearing in the ground-truth discharge instruction, plus a soft embedding-similarity term to the ground-truth concept set) is aligned with the clinical-efficacy evaluation, which matches generated concepts against ground-truth concepts. However, this does not make the evaluation circular: the policy is trained on 3,000 patients and evaluated on 1,000 held-out patients, and the reported CE metrics are computed on the LLM's generated discharge instructions, not on the retrieved paths themselves. There is no equation in the paper showing that the reward equals or determines the evaluation score by construction; the LLM must still convert the retrieved paths into free text, and the held-out split requires the retriever to generalize to new ground-truth concept sets. Self-citations in the paper (Ting et al., 2025a,b; Zhao et al., 2025) are contextual references and are not load-bearing for the central claim. The Limitations section explicitly acknowledges that more comprehensive physician evaluation is needed, which is an honest limitation rather than a circular step. The apparent mathematical inconsistency among P/R/F1 and Jaccard values in Table 2, if real, is a numerical reporting or metric-computation error and should be investigated under correctness review, not treated as circularity. Overall, the derivation chain does not reduce to its own inputs.

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

The central claim rests on six explicit or implicit premises plus six hand-set hyperparameters. The most consequential is that UMLS cluster connectivity supports clinically valid reasoning leaps. The paper provides no independent validation of this premise, and the reward aligns with the evaluation metric, which inflates the apparent soundness of the clinical efficacy numbers.

free parameters (6)
  • lambda (reward weight) = 10
    Eq. 7 weight balancing binary ground-truth concept hits against embedding similarity to ground-truth concepts; set by hand in Appendix A.1.
  • gamma (discount factor) = 0.1
    Eq. 9 discount factor; unusually low, giving early retrieval steps more credit for the episodic reward; chosen by hand.
  • T (maximum retrieval steps) = 5
    Fixed path length in Appendix A.1; controls how far reasoning can leap and directly affects path content.
  • G (retrieval processes per patient) = 10
    Group size in Eq. 8; sensitivity analysis in Fig. 3a shows performance varies with G, and G=10 appears to be chosen from test curves.
  • NP (number of prompted paths) = 10
    Number of reasoning paths fed to the LLM; sensitivity analysis in Fig. 3b shows a strong effect on metrics.
  • beta (entropy coefficient) = decayed, start value not reported
    Exploration term in Eq. 10; decayed during training but the initial value and schedule are unspecified, making replication harder.
assumptions (6)
  • domain assumption UMLS semantic groups and SNOMED CT relations encode clinically meaningful reasoning steps.
    Section 3.2 uses cross-cluster jumps as reasoning leaps; no independent clinical validation that such graph transitions correspond to valid diagnostic or therapeutic inference.
  • domain assumption Ground-truth discharge instructions are complete and correct targets.
    Reward Eq. 7 and CE metrics use the ground-truth concept set; if ground truth is noisy or incomplete, both training signal and evaluation are distorted.
  • domain assumption Pre-admission fields (allergies, chief complaint, HPI) contain enough signal to generate discharge instructions.
    Task setup in Sec. 3.1 and Sec. 4.1; no clinical evidence that these fields suffice for the full discharge instruction.
  • domain assumption SapBERT cosine similarity measures useful semantic relatedness between clinical concepts and paths.
    Eqs. 2, 6, and 7 rely on embedding similarity for concept selection and soft reward; no calibration against clinician judgments.
  • standard math REINFORCE with group-normalized rewards is a valid optimization for this sparse-reward retrieval MDP.
    Eqs. 9 and 10 follow standard policy gradient (Williams 1992; Shao et al. 2024); convergence and stability are not proven for this setting.
  • domain assumption The prompted LLM follows the retrieved reasoning paths faithfully.
    Eq. 11 assumes the LLM conditions on the paths as intended; no faithfulness analysis is provided.
invented entities (1)
  • cluster leap pseudo-relation
    purpose: Inserted into reasoning paths to mark a semantic-level jump between UMLS clusters during retrieval (Sec. 3.2 Connect step and Figure 2).
    A synthetic graph relation not present in UMLS. It allows the retriever to express cross-cluster transitions in the path string, but no independent evidence shows it is clinically meaningful.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Leaps Beyond the Seen: Reinforced Reasoning Augmented Generation for Clinical Notes." pith.science (2026). https://pith.science/paper/TDXMSYZZ

@misc{pith2026250605386,
  author       = {Pith},
  title        = {Pith review of: Leaps Beyond the Seen: Reinforced Reasoning Augmented Generation for Clinical Notes},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TDXMSYZZ}},
  note         = {Machine review of arXiv:2506.05386}
}
read the original abstract

Clinical note generation aims to produce free-text summaries of a patient's condition and diagnostic process, with discharge instructions being a representative long-form example. While recent LLM-based methods pre-trained on general clinical corpora show promise in clinical text generation, they fall short in producing long-form notes from limited patient information. In this paper, we propose ReinRAG, a reinforced reasoning augmented generation (RAG) for long-form discharge instructions based on pre-admission information. ReinRAG retrieves reasoning paths from a medical knowledge graph to provide explicit semantic guidance to the LLM. To bridge the information gap, we propose group-based retriever optimization (GRO) which improves retrieval quality with group-normalized rewards, encouraging reasoning leaps for deeper inference by the LLM. Comprehensive experiments on the real-world dataset show that ReinRAG outperforms baselines in both clinical efficacy and natural language generation metrics. Further analysis reveals that ReinRAG fills semantic gaps in sparse input scenarios, and retrieved reasoning paths help LLMs avoid clinical misinterpretation by focusing on key evidence and following coherent reasoning.

Figures

Figures reproduced from arXiv: 2506.05386 by the authors.

Figure 1
Figure 1. Keyword distribution across UMLS seman￾tic clusters in patients’ pre-admission information and discharge instructions. Keywords from pre-admission information are concentrated in the Disorders cluster, whereas those in discharge instructions span a broader range of semantic clusters, revealing a substantial infor￾mation gap. Patient discharge instruction summarizes a wide range of information, including diagnoses, m… view at source ↗
Figure 2
Figure 2. The overview of ReinRAG. After extracting the patient’s pre-admission information and matching keywords with the UMLS KG, the retrieval network performs two-level retrieval based on RL to form reasoning paths. Then, the group-based retriever optimization leverages group-relative rewards to optimize the policy network. Finally, the well-trained ReinRAG generates reasoning paths to guide the LLM in discharge instructi… view at source ↗
Figure 3
Figure 3. Parameter sensitivity analysis of ReinRAG with Mistral-7B-Instruct-v0.3. 45 55 65 75 F1 HL(↓) -1.6% -8.9% +1.2% +6.1% (%) Full Model w/o exploration w/o leaps CE Metrics (a) CE Evaluation. 10 15 20 25 RG-L MTR -2.8% -14.41% -0.8% -6.3% (%) NLG Metrics Full Model w/o exploration w/o leaps (b) NLG Evaluation [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Ablation study of ReinRAG with Mistral. meaning of ground-truth instructions with less ir￾relevant descriptions. Although it obtains a lower BLEU-2 score, the highest ROUGE-L, BERTScore (F1BERT), METEOR and Sentence-BERT similar￾ity scores confirm that ReinRAG produces…
Figure 5
Figure 5. Figure 5: Recall of vanilla Mistral-7B-Instruct-v0.3 [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 6
Figure 6. Figure 6: Overall human evaluation performance of Vanilla Mistral-7B-Instruct-v0.3 and ReinRAG. Scores range from 1 to 5, with higher scores indicating better performance. 0 0.2 0.4 0.6 0.8 1 1 2 3 4 5 Vanilla Mistral ReinRAG + Mistral Strongly Disagree Strongly Agree Reading Co…
Figure 7
Figure 7. Figure 7: Evaluation scores over four aspects [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

18 extracted references · 6 canonical work pages

  1. [4]

    arXiv preprint arXiv:2310.06825

    Mistral 7b. arXiv preprint arXiv:2310.06825. Haibo Jin, Haoxuan Che, Yi Lin, and Hao Chen

    arXiv

  2. [5]

    arXiv preprint arXiv:2411.15124

    T\" ulu 3: Pushing frontiers in open language model post-training. arXiv preprint arXiv:2411.15124. Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Hein- rich Küttler, Mike Lewis, Wen-tau Yih, Tim Rock- täschel, et al

    arXiv

  3. [6]

    In Proceedings of the 2024 Joint In- ternational Conference on Computational Linguis- tics, Language Resources and Evaluation (LREC- COLING 2024), pages 10632–10641

    Llamacare: An instruction fine-tuned large language model for clinical nlp. In Proceedings of the 2024 Joint In- ternational Conference on Computational Linguis- tics, Language Resources and Evaluation (LREC- COLING 2024), pages 10632–10641. Yunxiang Li, Zihan Li, Kai Zhang, Ruilong Dan, Steve Jiang, and You Zhang

    work page 2024

  4. [8]

    ai: An open-source retrieval-augmented large language model system for answering medical questions using scientific litera- ture

    Clinfo. ai: An open-source retrieval-augmented large language model system for answering medical questions using scientific litera- ture. In PACIFIC SYMPOSIUM ON BIOCOMPUT- ING 2024, pages 8–23. World Scientific. National Library of Medicine (US)

    work page 2024

  5. [9]

    https://huggingface.co/prithivMLmods/ Qwen-UMLS-7B-Instruct

    Qwen-umls-7b-instruct. https://huggingface.co/prithivMLmods/ Qwen-UMLS-7B-Instruct . Hugging Face model card. License: CreativeML OpenRAIL-M. Accessed: 2025-07-26. Pengcheng Qiu, Chaoyi Wu, Xiaoman Zhang, Weixiong Lin, Haicheng Wang, Ya Zhang, Yanfeng Wang, and Weidi Xie

    work page 2025

  6. [11]

    arXiv preprint arXiv:2402.03300

    Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300. Chang Shu, Baian Chen, Fangyu Liu, Zihao Fu, Ehsan Shareghi, and Nigel Collier

    arXiv

  7. [13]

    arXiv preprint arXiv:2306.09968

    Clinicalgpt: large language models finetuned with diverse medical data and comprehensive evaluation. arXiv preprint arXiv:2306.09968. Yilin Wen, Zifeng Wang, and Jimeng Sun

    arXiv

  8. [14]

    arXiv preprint arXiv:2308.09729

    Mindmap: Knowledge graph prompting sparks graph of thoughts in large language models. arXiv preprint arXiv:2308.09729. Christopher YK Williams, Jaskaran Bains, Tianyu Tang, Kishan Patel, Alexa N Lucas, Fiona Chen, Brenda Y Miao, Atul J Butte, and Aaron E Kornblith

    arXiv

Show all 18 references
  1. [16]

    In 2019 IEEE international conference on data mining (ICDM), pages 728–737

    Automatic generation of medical imaging diagnostic report with hierarchical recurrent neural network. In 2019 IEEE international conference on data mining (ICDM), pages 728–737. IEEE. Cyril Zakka, Rohan Shad, Akash Chaurasia, Alex R Dalal, Jennifer L Kim, Michael Moor, Robyn F...

    2019

  2. [17]

    arXiv preprint arXiv:2305.15075

    Hu- atuogpt, towards taming language model to be a doc- tor. arXiv preprint arXiv:2305.15075. Jinghan Zhang, Xiting Wang, Weijieying Ren, Lu Jiang, Dongjie Wang, and Kunpeng Liu

    arXiv

  3. [18]

    arXiv preprint arXiv:2508.01191

    Is chain-of-thought reasoning of llms a mirage? a data distribution lens. arXiv preprint arXiv:2508.01191. A Implementation Details A.1 Hyperparameter Settings For model training, the maximum number of re- trieval steps is set to 5, and the embedding dimen- sion is

    arXiv

  4. [1991]

    dis- charge me!

    Function opti- mization using connectionist reinforcement learning algorithms. Connection Science, 3(3):241–268. Chaoyi Wu, Weixiong Lin, Xiaoman Zhang, Ya Zhang, Weidi Xie, and Yanfeng Wang. 2024a. Pmc-llama: toward building open-source language models for medicine. Journal o...

    2024 arXiv

  5. [2016]

    In MedIR workshop, sigir, pages 1–4

    Quickumls: a fast, unsupervised approach for medical concept extraction. In MedIR workshop, sigir, pages 1–4. Lo Pang-Yun Ting, Hong-Pei Chen, An-Shan Liu, Chun- Yin Yeh, Po-Lin Chen, and Kun-Ta Chuang. 2025a. Early detection of patient deterioration from real- time wearable m...

    arXiv

  6. [2019]

    arXiv preprint arXiv:1908.10084

    Sentence-bert: Sentence embeddings using siamese bert-networks. arXiv preprint arXiv:1908.10084. Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Y Wu, et al

    1908 arXiv

  7. [2021]

    In NAACL-HLT, pages 4228–4238

    Self-alignment pretraining for biomedical entity representations. In NAACL-HLT, pages 4228–4238. Association for Computational Linguistics. Jingyu Liu, Jiaen Lin, and Yong Liu. 2024b. How much can rag help the reasoning of llm? arXiv preprint arXiv:2410.02338. Alejandro Lozano...

    arXiv

  8. [2023]

    arXiv preprint arXiv:2311.16079

    Meditron-70b: Scaling medical pretraining for large language mod- els. arXiv preprint arXiv:2311.16079. Michael Denkowski and Alon Lavie

    arXiv

  9. [2024]

    In AAAI 2024 Spring Symposium on Clinical Foundation Models

    Automated generation of hospital dis- charge summaries using clinical guidelines and large language models. In AAAI 2024 Spring Symposium on Clinical Foundation Models. Yanjun Gao, Ruizhe Li, Emma Croxford, John Caskey, Brian W Patterson, Matthew Churpek, Timothy Miller, Dmitr...

    2024

  10. [2025]

    arXiv preprint arXiv:2501.12948

    Deepseek-r1: In- centivizing reasoning capability in llms via reinforce- ment learning. arXiv preprint arXiv:2501.12948. Albert Q Jiang, Alexandre Sablayrolles, Arthur Men- sch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guil- ...

    arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.