Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

Grounding chain-of-thought in a biomedical knowledge graph lets small language models match classical clinical predictors on very little data.

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

2026-08-04 06:40 UTC pith:CWFKHX5M

load-bearing objection Useful low-data pipeline with real KG-grounded CoT gains on MIMIC-III, but the cross-cohort generalization claim is overstated and the CoT supervision is answer-conditioned. the 3 major comments →

arxiv 2512.01210 v4 pith:CWFKHX5M submitted 2025-12-01 cs.AI

Knowledge Graph Augmented Large Language Models for Disease Prediction

Ruiyu Wang , Tuan Vinh , Ran Xu , Yuyin Zhou , Jiaying Lu , Francisco Pasquel , Mohammed K Ali , Carl Yang This is my paper
classification cs.AI
keywords knowledge-graph-guided reasoningchain-of-thoughtdisease predictionelectronic health recordsMIMIC-IIIzero-shot transferclinical explainabilitysmall language model fine-tuning
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 claims that the reason small language models struggle at clinical prediction from electronic health records is not lack of scale but lack of grounded supervision. It builds a training corpus by mapping ICD-9 codes to nodes in a biomedical knowledge graph, extracting disease-relevant paths, prompting a language model to write step-by-step rationales under strict temporal rules, and keeping only rationales whose final Yes/No matches the true next-visit outcome. Fine-tuning 7–8B parameter models on 400–1,000 labeled visits with this corpus yields AUROC of 0.66–0.70 and macro-AUPR of 0.40–0.47 on MIMIC-III, competitive with logistic regression and XGBoost while producing explanations clinicians prefer in a blinded study. The paper also reports zero-shot transfer to a separate type-2 diabetes cohort, where accuracy rises from 0.40–0.51 to 0.72–0.77. A sympathetic reader would care because this is a concrete recipe for data-efficient, interpretable clinical decision support without proprietary-scale models.

Core claim

The central claim is that KG-anchored CoT supervision—not model size—is what lets a small LLM learn prognostic reasoning. Concretely, the paper establishes that a pipeline of (1) mapping each patient's ICD-9 codes to entities in a biomedical knowledge graph, (2) mining short paths from present features to the target disease, and (3) filtering generated explanations so that only those whose stated conclusion matches the observed outcome survive, produces a compact supervision set. Fine-tuning on that set, an 8B and a 7B instruction-tuned model outperform strong classical baselines on MIMIC-III under strict data scarcity, and their explanations are preferred over an untuned model's by clinicia

What carries the argument

The load-bearing object is the KG-guided CoT supervision corpus: a set of (visit, disease, explanation, outcome) tuples in which each explanation is anchored to nodes and paths from a biomedical knowledge graph (PrimeKG) and is temporally consistent by construction—features are treated as observations at the index visit, the target label refers to the next visit, and chronic/persistent evidence is weighted over transient findings. The filtering rule—keep a trace only if its concluding Yes/No equals the ground-truth label—does the work of aligning the explanation with the observed outcome, while the KG scaffold does the work of keeping the explanation clinically grounded. The paper also treat

Load-bearing premise

The zero-shot transfer claim depends on an unspecified alignment between MIMIC-III's 7,423 ICD-9 features and the target cohort's 12,724 binary features; without a well-defined mapping, the reported accuracy jump cannot be interpreted as generalization of the learned reasoning.

What would settle it

Recover the exact feature-space mapping between MIMIC-III and the type-2 diabetes cohort from the released code, then recalibrate the transferred model's decision threshold to the target cohort's prevalence; if the accuracy gain disappears while AUROC stays near 0.52, the cross-cohort result is an operating-point shift rather than generalization of learned reasoning.

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

If this is right

  • If the central claim holds, data-efficient clinical prediction does not require massive model scale: a 7–8B open-weight model plus a KG-guided, label-consistent supervision set suffices to reach classical-baseline-level AUROC/AUPR.
  • Clinician-facing explanations from such models can be aligned with biomedical structure and temporal logic at generation time, not as post hoc attribution, which is what the blinded preference results support.
  • The cross-cohort transfer result, taken at face value, says the learned decision rules are not bound to a specific EHR vocabulary or coding practice, provided a feature-space mapping exists.
  • The ablations imply that filtering and KG grounding play complementary roles: filtering drives accuracy/F1, KG grounding drives ranking (AUROC/AUPR); a practitioner would therefore tune them separately.

Where Pith is reading between the lines

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

  • The reported cross-cohort accuracy gain (0.40–0.51 to 0.72–0.77) with only modest AUROC/AUPR changes suggests the fine-tuned models mainly shift their decision threshold; a recalibrated comparison would determine whether any genuine ranking improvement transfers.
  • Because mapping, relevance selection, and path pruning rely on a proprietary foundation model, the pipeline as described is not fully reproducible with open-weight tools; substituting an open model for the curator is a direct test of the method's independence from any single LLM.
  • A natural extension the paper leaves implicit: the label-consistency filter makes the training set reflect the training cohort's base rate, so deploying on a cohort with different prevalence would require recalibration—this can be tested by evaluating on a second external cohort with known prevalence.
  • The single-visit formulation discards longitudinal history; extending to multi-visit inputs could remove the need for explicit temporal heuristics by letting the model see persistence directly.

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

3 major / 5 minor

Summary. The paper proposes a knowledge-graph-guided chain-of-thought (KG-CoT) framework for next-visit disease prediction on MIMIC-III. ICD-9 codes are mapped to PrimeKG; disease-relevant nodes and shortest paths are mined; GPT-4o generates CoT rationales conditioned on the index-visit features, KG paths, and the ground-truth next-visit label; traces whose final verdict matches the label are retained; LLaMA-3.1-8B and Gemma-7B are fine-tuned on the resulting corpus. The authors report improved AUROC (0.66–0.70) and macro-AUPR (0.40–0.47) over classical baselines at 400–1,000 training visits, zero-shot transfer to the CRADLE cohort (accuracy 0.72–0.77), and clinician preference for KG-guided rationales.

Significance. If the results hold, the paper would demonstrate a practical recipe for combining biomedical knowledge graphs with LLM reasoning to produce data-efficient, interpretable EHR models. The release of code and the explicit data-efficiency setup are commendable. However, two load-bearing issues prevent accepting the central claims as stated: the zero-shot CRADLE transfer is not interpretable because the feature-space alignment is undefined, and the answer-conditioned CoT supervision creates rationales that are guaranteed to match labels, which weakens the 'clinically grounded reasoning' claim. These are fixable in a revision, but they are not merely presentation concerns.

major comments (3)
  1. [§Results, Table 4; §Discussion, Finding 3; Conclusion] The zero-shot transfer claim is not supported because the paper never describes how a MIMIC-trained model with input dimension N=7,423 consumes CRADLE's 12,724 binary features. The sentence 'MIMIC-trained models transfer to the CRADLE cohort ... across cohorts and spaces' is insufficient. Without the exact feature alignment (e.g., intersection of ICD-9/ICD-10 codes, mapping to PrimeKG nodes, or a shared vocabulary), the accuracy jump from 0.40–0.51 to 0.72–0.77 cannot be attributed to learned generalization. Table 4 shows that AUROC changes only from ≈0.52 to ≈0.53 and AUPR stays ≈0.25, so the gains are threshold/operating-point shifts; the abstract's 'generalizes across cohorts' overstates this. The authors must either specify the alignment and show that it preserves clinical meaning, or remove/recalibrate the cross-cohort generalization claim.
  2. [§Method, CoT Generation with KG-Guided Reasoning and Filtering; Eq. (Ct,d)] The CoT generator is explicitly conditioned on the ground-truth label y^(d)_{t+1} as an input, and the filtering rule keeps a trace only if its concluding verdict matches that label. Consequently, every training rationale is constructed to agree with the known outcome; the prompt's instruction 'do not assume the given answer is correct' does not change the fact that the answer is provided. This risks distilling post-hoc rationalizations rather than independent prognostic reasoning, and it directly affects the explanation-quality claim: the clinician preference study (Figure 3) compares KG-guided traces against an untuned baseline, not whether the traces are causally faithful. The authors should either (a) evaluate an answer-unconditioned variant (or a counterfactual analysis), or (b) explicitly frame the supervision as label-guided rationalization and temper the 'clinically grounded reas
  3. [§Results, Table 2] LLM rows (LLaMA3-8B + KG-CoT, Gemma-7B + KG-CoT) are reported as single values while classical baselines show means ± standard deviations. Since fine-tuning exhibits stochastic variation and the paper selects the checkpoint with best validation macro-AUPR, the claim that KG-guided LLMs 'consistently outperform' baselines is not statistically supported. Provide multiple seeds/runs or confidence intervals for the LLM rows, or explicitly state the number of runs used; otherwise the comparison at equal compute is unverifiable.
minor comments (5)
  1. [§Results, Dataset paragraph] Typo: 'e split 12,353 index visits' should be 'We split'.
  2. [§Results, Human Evaluation] Cohen's κ values κ_clar, κ_cov, κ_corr are named but never reported; only preference percentages are given. Report the κ values so inter-rater reliability is transparent.
  3. [§Results, Table 3] The labels 'unfiltered' and 'no KG' are not precisely defined in the text. Clarify what each ablation removes (KG paths vs. filtering only vs. both) and how the input differs.
  4. [§Method, KG Entity Mapping] The mapping yields 1,513 mapped feature nodes out of 7,423 ICD-9 codes. The paper should state how unmapped features are handled in both the CoT prompt and the fine-tuned model input (are they included as raw codes, ignored, or projected?).
  5. [§Discussion, Finding 3] The discussion appropriately notes that the CRADLE results 'mainly shift the operating point,' but the abstract and conclusion do not carry this caveat. Please align the abstract language with Finding 3.

Circularity Check

1 steps flagged

Held-out disease predictions are independent, but the explanation-quality claim is partly circular because the CoT supervision is generated from and filtered to the ground-truth label.

specific steps
  1. self definitional [Methods: 'CoT Generation with KG-Guided Reasoning' and 'Filtering'; prompt template in Figure 4]
    "CoT Generation with KG-Guided Reasoning. ... Ct,d = LLM(d, x+t, R+d(t), R−d(t), Pd, y(d)t+1 | Igen) ... Filtering. ... Ct,d is kept if and only if the CoT-implied label matches the ground-truth outcome y(d)t+1 ... If your final conclusion contradicts the given answer, acknowledge potential discrepancies without mentioning provided answer ... and refine your response again accordingly."

    The CoT generator receives the ground-truth label y(d)t+1 as an explicit input, the prompt instructs it to 'refine your response' whenever the conclusion contradicts that answer, and the filter retains only traces with ŷ=y. So every training rationale is a post-hoc rationalization of the outcome it is supposed to explain. The paper then uses these rationales to fine-tune models and claims 'clinically grounded prognostic reasoning,' with clinicians preferring them on 'correctness & soundness.' That explanation claim is partly circular: the rationales were manufactured to match the label shown to clinicians. Held-out disease predictions are not circular (test labels are absent at inference), so the circularity is confined to the reasoning/explanations claim.

full rationale

The central predictive claims on MIMIC-III (AUROC 0.66–0.70, macro-AUPR 0.40–0.47) are evaluated on a held-out 10% of visits and are not forced by the training construction; the fine-tuned models do not receive y(d)t+1 at test time. The CRADLE zero-shot numbers are also evaluated against CRADLE ground truth, not the MIMIC labels used to construct the CoT corpus, so they are not circular in the 'prediction equals input' sense. The main circularity concern is real but narrower: the CoT supervision corpus is generated by an LLM that is handed the answer and told to refine any conclusion that disagrees with it, then filtered on agreement. Hence the paper's explanation-quality claims—'clinically grounded prognostic reasoning' and clinician preference on 'correctness & soundness'—are partly self-fulfilling; the rationales are rationalizations of labels, not independent derivations. This does not invalidate the AUROC/AUPR evidence, but it does mean the 'reasoning' component should be treated as label-conditioned distillation rather than independent reasoning. Separately, the cross-cohort generalization claim is under-specified: the paper says MIMIC-trained models transfer 'across cohorts and spaces' without defining how the 7,423-feature MIMIC space relates to CRADLE's 12,724 features, and Finding 3 concedes the accuracy gain is mainly an operating-point shift. These are correctness/reproducibility risks, not circularity. No load-bearing self-citation chain is present; the CRADLE cohort citation ([25]) is to the authors' prior work but only supplies the cohort definition, and the central derivation does not reduce to it. Score 5 reflects partial circularity in the explanation claim while the predictive derivation remains independent.

Axiom & Free-Parameter Ledger

4 free parameters · 6 axioms · 0 invented entities

The paper relies on PrimeKG and GPT-4o as external oracles, plus hand-set hyperparameters (τ, L, K_node, K_path) and a label-conditioned CoT generation/filtering step. No new physical or clinical entities are invented. The label-conditioned filtering is the main circularity burden.

free parameters (4)
  • similarity threshold τ = 0.85
    Stage 2 of ICD-to-KG mapping keeps candidates only above cosine 0.85; no sensitivity analysis is given, and it affects the set of mapped nodes.
  • max path length L = 5
    Shortest-path enumeration is capped at L=5 to avoid overlong chains; the choice is heuristic and not varied in experiments.
  • KG relevance/path counts K_node, K_path = 8 and 5
    GPT-4o selects 8 relevant nodes and 5 paths per disease; these hand-set caps control the KG evidence used in CoT generation.
  • decision threshold = 0.5
    Fixed threshold used to convert probabilities to labels for accuracy/F1; on CRADLE the transfer gains are attributed largely to threshold shift in the Discussion.
axioms (6)
  • domain assumption PrimeKG is a reliable and sufficiently complete substrate for disease-relevant biomedical relations.
    All entity mapping, relevance mining, and path extraction use PrimeKG; the Limitations section states the approach depends on PrimeKG coverage and quality.
  • domain assumption GPT-4o can correctly validate ICD-9-to-KG mappings and select clinically relevant nodes/paths.
    Stages 2/3 of mapping and the node/path selection steps delegate correctness judgments to GPT-4o without gold-standard evaluation; acknowledged in Limitations.
  • ad hoc to paper Shortest paths up to length L=5 between mapped feature nodes and disease nodes capture the clinically meaningful mechanisms.
    The hop bound and the path count K_path are set heuristically; the paper does not validate that pruned paths include all relevant chains.
  • ad hoc to paper Answer-conditioned CoT traces, after filtering on label match, are valid supervision for learning reasoning rather than memorizing label statistics.
    Ct,d is generated with y(d)_{t+1} in the prompt and kept only when the CoT conclusion equals y; the paper does not test the independence of the reasoning from the provided answer.
  • domain assumption A single index visit's ICD-9 features are sufficient to predict next-visit disease in this setup.
    The task uses adjacent visit pairs following Harutyunyan et al.; Limitations acknowledge that longitudinal history is not fully exploited.
  • domain assumption Clinician preference on clarity/relevance/correctness is a valid proxy for explanation quality.
    The human evaluation uses two clinicians' preferences as the main evidence for explanation quality; no objective correctness metric is used.

pith-pipeline@v1.3.0-alltime-deepseek · 11145 in / 18636 out tokens · 165864 ms · 2026-08-04T06:40:48.896944+00:00 · methodology

0 comments
read the original abstract

Electronic health records (EHRs) enable strong clinical prediction, but explanations are often coarse and hard to use for patient-level decisions. We propose a knowledge graph (KG)-guided chain-of-thought (CoT) framework for visit-level disease prediction on MIMIC-III. We map ICD-9 codes to PrimeKG, mine disease-relevant nodes and paths, and use these paths to scaffold temporally consistent CoT rationales, retaining only samples whose conclusions match observed outcomes. We fine-tune lightweight instruction-tuned LLMs (LLaMA-3.1-Instruct-8B and Gemma-7B) on two small cohorts (400 and 1,000 index visits) across ten PrimeKG-mapped diseases. Our models outperform strong classical baselines, reaching AUROC 0.66-0.70 and macro-AUPR 0.40-0.47. Without additional training, the models transfer zero-shot to the CRADLE cohort, improving accuracy from 0.40-0.51 to 0.72-0.77. In a blinded clinician study, KG-guided CoT rationales are consistently preferred for clarity, relevance, and correctness. Code is available at: https://github.com/JonathanWry/KG-guided-LLM-pipeline

Figures

Figures reproduced from arXiv: 2512.01210 by Carl Yang, Francisco Pasquel, Jiaying Lu, Mohammed K Ali, Ran Xu, Ruiyu Wang, Tuan Vinh, Yuyin Zhou.

Figure 1
Figure 1. Figure 1: General pipeline for KG-guided CoT data generation. Observed clinical features are denoted in Blue; Intermediate KG nodes are denoted in Green, and the Target disease is denoted in Red. Preliminaries. A knowledge graph (KG) is a structured representation of entities and their relationships, typically modeled as a graph G = (V, E) where V denotes the set of entities (nodes) and E the set of typed edges enco… view at source ↗
Figure 2
Figure 2. Figure 2: Schematic and example of KG-guided CoT generation and filtering. Observed clinical features are denoted in Blue; Intermediate KG nodes are denoted in Green, and the Target disease is denoted in Red [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Clinician preference for KG-guided explanations vs. the untuned baseline across three dimensions (donuts show % preferred; remainder is baseline). On the 115 annotated cases, both clinicians preferred the KG-guided model (Model A) over the untuned base￾line (Model B): 96.5% for Clarity & Coherence, 94.8% for Coverage & Relevance, and 98.3% for Correctness & Soundness. Qualitatively, clinicians noted that K… view at source ↗
Figure 4
Figure 4. Figure 4: Prompt templates used in our KG-guided CoT pipeline: (top-left) disease-relevant node selection, (top￾right) KG path pruning and selection, (bottom) CoT generation conditioned on KG evidence and visit features [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗

discussion (0)

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

Forward citations

Cited by 1 Pith paper

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

  1. Lung-R1: A Knowledge Graph-Guided LLM for Pulmonary Diagnostic Reasoning

    cs.AI 2026-06 unverdicted novelty 5.0

    Introduces the first structured pulmonary knowledge graph LungKG and uses it to train Lung-R1, which reaches SOTA on EMR-based pulmonary diagnosis tasks.

Reference graph

Works this paper leans on

33 extracted references · 4 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Jamoom, and Michael F

    Jennifer King, Vaishali Patel, Eric W. Jamoom, and Michael F. Furukawa. Clinical benefits of electronic health record use: national findings.Health Services Research, 49(1 Pt 2):392–404, 2014

    2014

  2. [2]

    Representation learning for clinical time series prediction tasks in electronic health records.BMC Medical Informatics and Decision Making, 19(Suppl 8):259, 2019

    Tong Ruan, Liqi Lei, Yangming Zhou, Jie Zhai, Le Zhang, Ping He, and Ju Gao. Representation learning for clinical time series prediction tasks in electronic health records.BMC Medical Informatics and Decision Making, 19(Suppl 8):259, 2019

    2019

  3. [3]

    Automatically detecting acute myocardial infarction events from ehr text: a preliminary study

    Jiaping Zheng, Jorge Yarzebski, Balaji Polepalli Ramesh, Robert J Goldberg, and Hong Yu. Automatically detecting acute myocardial infarction events from ehr text: a preliminary study. InAMIA Annual Symposium Proceedings, volume 2014, page 1286, 2014

    2014

  4. [4]

    Fogel and Joseph C

    Alexander L. Fogel and Joseph C. Kvedar. Artificial intelligence powers digital medicine.NPJ Digital Medicine, 1(5):1–4, 2018

    2018

  5. [5]

    Glicksberg, Hyejin C

    Irene Landi, Benjamin S. Glicksberg, Hyejin C. Lee, Sarah Cherng, Giovanni Landi, Mattia Danieletto, Christina Lu, Chih-lin Hsu, Ray Chen, and Joel T. Dudley. Deep representation learning of electronic health records to unlock patient stratification at scale.NPJ Digital Medicine, 3(1):96, 2020

    2020

  6. [6]

    Alistair E. W. Johnson, Tom J. Pollard, Lu Shen, Li-Wei H. Lehman, Mengling Feng, Mohammad Ghassemi, Benjamin Moody, Peter Szolovits, Leo A. Celi, and Roger G. Mark. Mimic-iii, a freely accessible critical care database.Scientific Data, 3:160035, 2016

    2016

  7. [7]

    Goldstein, Ann Marie Navar, Michael J

    Benjamin A. Goldstein, Ann Marie Navar, Michael J. Pencina, and John P. A. Ioannidis. Opportunities and challenges in developing risk prediction models with electronic health record data: a systematic review.Journal of the American Medical Informatics Association, 24(1):198–208, 2017

    2017

  8. [8]

    Retain: An interpretable predictive model for healthcare using reverse time attention mechanism.Advances in neural information processing systems, 29, 2016

    Edward Choi, Mohammad Taha Bahadori, Jimeng Sun, Joshua Kulas, Andy Schuetz, and Walter Stewart. Retain: An interpretable predictive model for healthcare using reverse time attention mechanism.Advances in neural information processing systems, 29, 2016

    2016

  9. [9]

    Intelligible models for healthcare: Predicting pneumonia risk and hospital 30-day readmission

    Rich Caruana, Yin Lou, Johannes Gehrke, Paul Koch, Marc Sturm, and Noemie Elhadad. Intelligible models for healthcare: Predicting pneumonia risk and hospital 30-day readmission. InProceedings of the 21st ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD ’15), pages 1721–1730, 2015

    2015

  10. [10]

    Stop explaining black box machine learning models for high stakes decisions and use inter- pretable models instead.Nature Machine Intelligence, 1(5):206–215, 2019

    Cynthia Rudin. Stop explaining black box machine learning models for high stakes decisions and use inter- pretable models instead.Nature Machine Intelligence, 1(5):206–215, 2019

    2019

  11. [11]

    Chain-of-thought prompting elicits reasoning in large language models.Advances in neural information pro- cessing systems, 35:24824–24837, 2022

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models.Advances in neural information pro- cessing systems, 35:24824–24837, 2022

    2022

  12. [12]

    Ni, et al

    Jun Wang, Meng Fang, Ziyu Wan, Muning Wen, Jiachen Zhu, Anjie Liu, Ziqin Gong, Yan Song, Lei Chen, Lionel M. Ni, et al. Openr: An open source framework for advanced reasoning with large language models. arXiv preprint arXiv:2410.09671, 2024

    Pith/arXiv arXiv 2024

  13. [13]

    Towards reasoning in large language models: A survey.Findings of the Association for Computational Linguistics: ACL 2023, 2023

    Jie Huang and Kevin Chen-Chuan Chang. Towards reasoning in large language models: A survey.Findings of the Association for Computational Linguistics: ACL 2023, 2023

    2023

  14. [14]

    A preliminary study of o1 in medicine: Are we closer to an ai doctor?, 2024

    Yunfei Xie, Juncheng Wu, Haoqin Tu, Siwei Yang, Bingchen Zhao, Yongshuo Zong, Qiao Jin, Cihang Xie, and Yuyin Zhou. A preliminary study of o1 in medicine: Are we closer to an ai doctor?, 2024

    2024

  15. [15]

    s1: Simple test-time scaling.Conference on Empirical Methods in Natural Language Processing (EMNLP), 2025

    Niklas Muennighoff, Zitong Yang, Weijia Shi, Xiang Lisa Li, Li Fei-Fei, Hannaneh Hajishirzi, Luke Zettle- moyer, Percy Liang, Emmanuel Cand`es, and Tatsunori Hashimoto. s1: Simple test-time scaling.Conference on Empirical Methods in Natural Language Processing (EMNLP), 2025

    2025

  16. [16]

    Limo: Less is more for reasoning

    Yixin Ye, Zhen Huang, Yang Xiao, Ethan Chern, Shijie Xia, and Pengfei Liu. Limo: Less is more for reasoning. Conference on Language Modeling (COLM), 2025

    2025

  17. [17]

    Ram- ehr: Retrieval augmentation meets clinical predictions on electronic health records

    Ran Xu, Wenqi Shi, Yue Yu, Yuchen Zhuang, Bowen Jin, May Dongmei Wang, Joyce Ho, and Carl Yang. Ram- ehr: Retrieval augmentation meets clinical predictions on electronic health records. InProceedings of the Annual Meeting of the Association for Computational Linguistics (ACL), pages 754–765, 2024

    2024

  18. [18]

    Liu et al. Improving large language model applications in biomedicine with retrieval-augmented generation: A systematic review, meta-analysis, and clinical development guidelines.Journal of the American Medical Informatics Association, 2025

    2025

  19. [19]

    Building a knowledge graph to enable precision medicine

    Payal Chandak, Kexin Huang, and Marinka Zitnik. Building a knowledge graph to enable precision medicine. Scientific Data, 10(1):155, 2023

    2023

  20. [20]

    O1-pruner: Length-harmonizing fine-tuning for o1-like reasoning pruning.arXiv preprint arXiv:2501.12570, 2025

    Haotian Luo, Li Shen, Haiying He, Yibo Wang, Shiwei Liu, Wei Li, Naiqiang Tan, Xiaochun Cao, and Dacheng Tao. O1-pruner: Length-harmonizing fine-tuning for o1-like reasoning pruning.arXiv preprint arXiv:2501.12570, 2025

    Pith/arXiv arXiv 2025

  21. [21]

    Do not think that much for 2+3=? on the overthinking of o1-like llms.Proceed- ings of the 42nd International Conference on Machine Learning (ICML)), 2024

    Xingyu Chen, Jiahao Xu, Tian Liang, Zhiwei He, Jianhui Pang, Dian Yu, Linfeng Song, Qiuzhi Liu, Mengfei Zhou, Zhuosheng Zhang, et al. Do not think that much for 2+3=? on the overthinking of o1-like llms.Proceed- ings of the 42nd International Conference on Machine Learning (ICML)), 2024

    2024

  22. [22]

    The llama 3 herd of models.arXiv preprint arXiv:2407.21783, 2024

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, et al. The llama 3 herd of models.arXiv preprint arXiv:2407.21783, 2024

    Pith/arXiv arXiv 2024

  23. [23]

    Gemma: Open models based on gemini research and technol- ogy.arXiv preprint arXiv:2403.08295, 2024

    Gemma Team, Thomas Mesnard, Cassidy Hardin, Robert Dadashi, Surya Bhupatiraju, Shreya Pathak, Laurent Sifre, Morgane Rivi`ere, Mihir Sanjay Kale, et al. Gemma: Open models based on gemini research and technol- ogy.arXiv preprint arXiv:2403.08295, 2024

    Pith/arXiv arXiv 2024

  24. [24]

    Multitask learning and benchmarking with clinical time series data.Scientific Data, 6(1):96, 2019

    Hrayr Harutyunyan, Hrant Khachatrian, David C Kale, Greg Ver Steeg, and Aram Galstyan. Multitask learning and benchmarking with clinical time series data.Scientific Data, 6(1):96, 2019

    2019

  25. [25]

    role": "system

    Ran Xu, Mohammed K. Ali, Joyce C. Ho, and Carl Yang. Hypergraph transformers for ehr-based clinical predictions.AMIA Joint Summits on Translational Science Proceedings, 2023:582–591, 2023. Implementation Details.We fine-tuned LLaMA-3.1-8B-Instruct and Gemma-7B-Instruct on a single NVIDIA H200 (512 GB host memory; Emory HPC); full hyperparameters and train...

    2023

  26. [26]

    Explore the question and first filter out irrelevant partial reasoning paths provided for disease verification, leaving only useful paths based on your expertise

  27. [27]

    Treat potentially_relevant_present as confirmed present (1) and potentially_relevant_absent as confirmed absent (0) at time t

    Incorporate the remaining paths and potentially_relevant_features naturally as if you discovered them yourself, then supplement with your new reasoning path from provided features to form complete reasoning paths based on your expertise. Treat potentially_relevant_present as confirmed present (1) and potentially_relevant_absent as confirmed absent (0) at time t

  28. [28]

    Diagnostic evidence at time t only supports t+1 if it is chronic/persistent

    Analyze and evaluate the reasoning paths with explicit attention to TEMPORAL V ALIDITY , expanding on the most relevant ones together with explicit diagnostic evidence to construct a logical, well-supported explanation. Diagnostic evidence at time t only supports t+1 if it is chronic/persistent. Prefer objective findings (imaging, labs, pathognomonic sign...

  29. [29]

    Avoid using assumptions, indirect risk factors, or therapy context as critical evidence

    Base your decision ONL Y on confirmed, definitive evidence (1/0). Avoid using assumptions, indirect risk factors, or therapy context as critical evidence. Do NOT equate risk factor with a confirmed diagnosis

  30. [30]

    unspecified hemorrhage

    If there are resolution cues (e.g., “unspecified hemorrhage” without source, single-episode events, acute procedures that typically resolve), down-weight them for predicting t+1. Always check for **resolution indicators** (e.g., ""treated,"" ""status post,"" ""resolved,"" ""acute episode"") and treat such findings as low-weight for t+1 prediction

  31. [31]

    Do not mention the existence of predefined reasoning paths or the provided answer in your response

  32. [32]

    Instead, determine the answer solely based on your reasoning

    Do not assume the given answer is correct. Instead, determine the answer solely based on your reasoning

  33. [33]

    "Wait, there might be something wrong

    If your final conclusion contradicts the given answer, acknowledge potential discrepancies without mentioning provided answer (e.g., ""Wait, there might be something wrong"") and refine your response again accordingly. ### Output: Finding reasoning paths: (you ""discover"" potential reasoning paths yourself by using the given paths if useful or generating...