Pith. sign in

REVIEW 3 major objections 5 minor 52 references

Compiling clinical guidelines into executable, disease-specific scoring functions and fusing them with an LLM's ranked differential diagnosis improves diagnostic accuracy across four benchmarks and four backbones — without updating the mode

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

2026-08-01 00:36 UTC pith:SFYRJPU6

load-bearing objection Genuinely useful idea and a strong multi-backbone headline, but the empirical gains rest on a feature-grounding step the paper never validates; treat the accuracy numbers as provisional. the 3 major comments →

arxiv 2607.26160 v1 pith:SFYRJPU6 submitted 2026-07-28 cs.AI

GuideSkill: Evolving Executable LLM Agent Skills for Guideline-Grounded Clinical Reasoning

Lang Cao , Yuhao Shen , Tianyang Luo , Simo Du , Hao Peng , Yue Guo This is my paper
classification cs.AI
keywords clinical practice guidelinesexecutable skillsLLM agentsdifferential diagnosisICD-10diagnostic reasoningskill libraryguideline grounding
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 tries to establish that clinical practice guidelines can be turned from passive reference text into an external library of executable diagnostic rules — Python functions that return a 0–3 support tier for a candidate disease — and that fusing those deterministic scores with an LLM's ranked differential diagnosis improves accuracy. GuideSkill-Zero, built from guidelines alone, already beats guideline retrieval on every backbone. GuideSkill-Evo, which refines existing skills and adds diagnoses from labeled cases, raises skill coverage from 56.5% to 99.5% and improves over direct inference by 18.49% relative, with the best macro-average accuracy for all four tested backbones. If correct, guideline knowledge can be maintained, inspected, and updated outside the model, and reused across different LLMs without retraining.

Core claim

The central claim is that diagnostic criteria from clinical practice guidelines can be compiled into executable, disease-indexed Python skills that assign one of four ordinal support levels (Confirmed, Strongly Suggestive, Compatible, Not Supported) to the evidence in a patient case. At inference, an LLM proposes a ranked differential of five ICD-10 categories; for each candidate, the framework grounds only the features that candidate's skill requires, executes the skill to get a normalized evidence score, and fuses it with the LLM ranking score (default α=0.5) to pick the final diagnosis. The paper shows this two-signal fusion outperforms direct prompting, chain-of-thought, retrieval-augmen

What carries the argument

The executable diagnostic skill: a Python function generated from guideline recommendations (or case-distilled rules) that takes a case dictionary of boolean/numeric clinical features and returns a 0–3 tier plus a normalized score. Each skill encodes disease-specific decision logic, including contradicting-evidence downgrades and missing-value handling, and is indexed by three-character ICD-10 category. The work it does is to translate vague guideline prose into fixed, inspectable, deterministic scoring rules that are comparable across diseases on the same tier scale, so that the LLM's ranking can be corrected by explicit evidence-based support scores. The feature-grounding step supplies the

Load-bearing premise

The entire empirical gain rests on the feature-grounding step faithfully converting the free-text patient case into the exact boolean/numeric keys each compiled skill expects — including treating unmentioned features as absent — because a single wrong key can change the support tier and flip the fused diagnosis.

What would settle it

Replace the LLM feature-grounding step with oracle features drawn from the gold-structured record for a sample of cases and compare accuracy; if the gap between oracle-grounded and LLM-grounded GuideSkill-Evo is large, the reported gains depend on grounding fidelity rather than skill logic. Alternatively, perturb a single grounded feature per case (e.g., toggle biopsy_noncaseating_granuloma) and measure how often the final diagnosis changes.

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

If this is right

  • GuideSkill-Zero, using only guideline-derived skills, outperforms guideline RAG on every backbone, so executable compilation alone adds diagnostic value before any case data is used.
  • Case-derived evolution adds 267 new ICD-10 skills and lifts gold-label coverage from 56.5% to 99.5%, improving accuracy on newly covered diagnoses in all 12 available backbone–benchmark comparisons.
  • On the held-out MedThink-Bench, excluded from evolution, GuideSkill-Evo improves over GuideSkill-Zero on all four backbones, indicating case-derived rules transfer to an unseen benchmark.
  • On Qwen3.5-9B, GuideSkill-Evo (50.98 macro-average) beats the strongest parameter-update baseline (45.86) and the guideline-decision-tree baseline without updating the backbone.
  • Executable skills show zero run-to-run accuracy variation, whereas textual rubrics vary, indicating a more reproducible scoring interface.

Where Pith is reading between the lines

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

  • The feature-grounding step is a single point of failure: a controlled study that perturbs grounded features (e.g., toggling one boolean per case) would reveal how much of the gain is due to skill logic versus grounding fidelity.
  • Because skills are external and inspectable, the same library could double as a clinician-auditable artifact for guideline validation, separate from the LLM that invokes it.
  • The compilation pipeline is not diagnosis-specific; evolving executable skills from labeled cases could apply to treatment, screening, or triage guidelines, wherever ordinal evidence rules can be stated.
  • The tier contract (0–3) creates a common currency for evidence strength across diseases, which could be used to build interpretable differential-scoring dashboards for real clinical workflows.

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 introduces GuideSkill, an external reasoning layer for clinical diagnosis that compiles disease-specific criteria from clinical practice guidelines into executable Python functions returning ordinal support tiers (GuideSkill-Zero), then refines and extends this library using labeled patient cases (GuideSkill-Evo). At inference, an LLM proposes a top-K differential diagnosis, the relevant skills are retrieved, the LLM grounds each skill's required features from the free-text case (Eq. 12), the skills are executed deterministically, and the LLM ranking is fused with the skill scores using a weight α=0.5. The authors report experiments across four benchmarks (MedCaseReasoning, ER-Reason, MIMIC-CDM-FI, MedThink-Bench) and four backbones (GPT-5.4, Claude-Sonnet-4.6, MedGemma-27B, Qwen3.5-9B), claiming that GuideSkill-Evo achieves the highest macro-average accuracy for every backbone, improves over direct inference by 18.49% relatively, raises gold-label skill coverage from 56.5% to 99.5%, and outperforms parameter-update baselines on Qwen3.5-9B without updating the backbone. A clinician assessment of ten skills is also reported.

Significance. If the empirical claims hold, the contribution is genuinely useful: it provides a model-agnostic, inspectable, and extensible way to operationalize guideline knowledge for LLM-based diagnosis, and it shows transfer to a benchmark (MedThink-Bench) excluded from skill evolution. The design is clean and the evaluation is broad in terms of backbones and benchmarks. The paper also includes several strengths that should be acknowledged: MedThink-Bench is held out from evolution, the clinician assessment uses a blinded tier-assignment protocol with look-alike cases, the error analysis identifies candidate recall as the dominant remaining failure mode, and the efficient variant reduces cost while retaining most of the accuracy gain. The main reservations concern the correctness criterion, the unvalidated feature-grounding step, and the absence of statistical reliability measures; these are fixable but currently block full confidence in the reported magnitudes and in the attribution of the gains to guideline-derived skill logic.

major comments (3)
  1. [§4 'Evaluation' vs Appendix H 'Evaluation Protocol'] The manuscript gives two incompatible correctness criteria. §4 states that predictions are evaluated by exact equality of ICD-10 codes, with malformed or code-free outputs counted as incorrect. Appendix H states that semantic matching between the predicted diagnosis and the gold label is performed by an LLM-as-a-judge (claude-haiku-4-5), and flags judge-based matching as a limitation. This is not a presentation detail: every accuracy number in Tables 2–4 depends on which criterion was actually used. If semantic matching was used, the reported gains could partly reflect judge bias toward one method's output format; if exact equality was used, Appendix H and its stated limitation are misleading. Please state the primary protocol, run both variants, and report judge agreement or an error analysis for the semantic matcher.
  2. [§3.4, Eq. (12); Appendix K; Appendix L] Skill Feature Grounding is the only link between free-text patient cases and deterministic skill execution, but no direct grounding accuracy is reported anywhere in the manuscript. The Appendix K trace shows that the final decision can flip on boolean keys such as `biopsy_noncaseating_granuloma` and `mimics_excluded`; with α=0.5, a one-tier grounding error changes the fused score by 1/6 and a two-tier error by 1/3, enough to overturn the LLM ranking. Since the skill docstring names the disease, the grounding LLM may systematically mark candidate-supporting features as true, in which case the skill score partially recapitulates the same LLM's ranking rather than external guideline logic. The clinician assessment (Table 5) covers 10 of 473 skills and 51 held-out cases, with κ=0.60, which is moderate support but not validation of the full library. Please report per-feature grounding precisi
  3. [§5, Table 2; Appendix H (inference settings)] No variance, confidence intervals, or significance tests are reported. MedThink-Bench has only 55 test cases, so a difference of two to five correct cases changes accuracy by roughly 4–9 percentage points; the macro-average weights this small benchmark equally with the 894-case MedCaseReasoning. GPT-5.4 is queried at the API default temperature of 1, so even the proprietary-model results are not deterministic. Before claiming 'highest macro-average for every backbone,' please provide bootstrap confidence intervals, per-benchmark case-level improvement counts, or repeated-run statistics for at least the small benchmarks and for the GPT-5.4 runs.
minor comments (5)
  1. [Table 4 / Appendix C] The executable-vs-textual comparison reports accuracy values (e.g., ER-Reason 63.00) that do not match the corresponding Claude-Sonnet-4.6 GuideSkill-Evo values in Table 2 (50.56). Appendix C says 'We evaluate 349 cases' while the table lists four datasets; the composition and selection of the 349 cases should be clarified.
  2. [Appendix G] The cross-reference 'As shown in Table 6' appears to point to the efficiency table; the error-analysis table is Table 7. Please fix the reference.
  3. [Appendix D and E] The text refers to the α-sensitivity figure and the top-k figure as Figures 4 and 5 in the main text, but Appendix D says 'Figure 2' and Appendix E says 'Figure 3'. The figure numbering is inconsistent.
  4. [Appendix H (evolution rule)] The statement that a matching guideline recommendation 'is never weakened' deserves a short discussion. The clinician assessment found one case where evolution overgeneralized and changed an originally correct rule into an inappropriate one; if guideline-derived rules cannot be weakened, that failure mode may be systematically unfixable by the proposed evolution mechanism.
  5. [Throughout] Minor typos and formatting issues: 'The analysis is judge and cateogirze also by Claude-Sonnet-4.6' in Appendix G; the phrase 'Table3' appears without a space in §5; Figure 2's legend uses '34 7 6...' obscurely and should be relabeled for readability.

Circularity Check

0 steps flagged

No significant circularity: central claims are checked on held-out benchmarks and supervised evolution is ordinary fitting, not disguised prediction.

full rationale

The paper's derivation chain is self-contained rather than circular. GuideSkill-Zero is compiled from CPG documents before benchmark labels are used ('The result is an ICD-10-indexed library of 349 executable diagnostic skills'), and GuideSkill-Evo refines skills using training splits, which is standard supervised fitting. The central external check is MedThink-Bench: 'MedThink-Bench is reserved for external evaluation and excluded from evolution,' and Table 2 shows GuideSkill-Evo improves over direct inference on that benchmark for every backbone (e.g., 34.55 vs 21.82 for Qwen3.5-9B). The coverage increase from 56.5% to 99.5% is largely a by-construction consequence of adding skills for training diagnoses, but the paper reports it as a mechanism rather than as a prediction, and accuracy is measured on held-out test cases. The only fragile component is Eq. 12 (Skill Feature Grounding), where an LLM converts free text into the exact feature dict expected by each skill; an error can change support tiers and flip the fused diagnosis (Appendix K). This is a validation/robustness threat, not a circular reduction: no equation makes s_skill equal to s_LLM or to a fitted parameter. Self-citations (e.g., MedGuideX, Shen et al. 2026b) appear only in related-work positioning and are not load-bearing; no uniqueness theorem or ansatz is imported from them. Therefore I can exhibit no specific circular step, and the low score reflects only minor self-reference and the unvalidated grounding interface rather than circularity.

Axiom & Free-Parameter Ledger

2 free parameters · 4 axioms · 1 invented entities

The central method consists of two learned components: the compiled skills (from guidelines via an LLM) and the evolution rules (from cases via an LLM). The main hand-set hyperparameters are α and K. The method relies on the reliability of the source guidelines, the ICD-10 normalization by an LLM, and the LLM judge used for evaluation.

free parameters (2)
  • fusion weight α = 0.5 (default)
    Controls the blend of LLM ranking score and skill evidence score (Eq. 15). Set to 0.5 by hand; sensitivity analysis in Appendix D shows the optimal value ranges from 0.0 (MedThink) to 0.5 (MedCase, ER-Reason), so the default is not individually optimal for every benchmark.
  • candidate set size K = 5 (default)
    Number of top diagnoses proposed by the LLM (Eq. 1). Chosen as a trade-off in Appendix E; accuracy saturates near K=5, but the value is a hand-set hyperparameter.
axioms (4)
  • domain assumption Clinical practice guidelines in the curated corpus (epfl-llm/guidelines) provide correct and current diagnostic criteria for the covered diseases.
    The entire GuideSkill-Zero library is compiled from these documents (§B.1). If a guideline is outdated or wrong, the skill inherits the error.
  • domain assumption Three-character ICD-10 categories are a meaningful granularity for evaluating diagnosis predictions.
    All gold labels and predictions are mapped to ICD-10 categories (§3.1, §4). The paper's own error analysis shows coding granularity causes 74.8% of recall-miss errors, indicating this assumption is imperfect.
  • domain assumption The LLM-as-a-judge (claude-haiku-4-5) correctly decides whether a predicted diagnosis matches the gold ICD-10 category.
    Accuracy is computed via this judge (§H Evaluation Protocol). The authors themselves note LLM judges can be biased; if the judge is inconsistent, all reported numbers shift.
  • ad hoc to paper The evolution rule that an existing guideline-derived recommendation is 'never weakened' preserves correctness.
    In case-to-recommendation refinement, when a matching guideline recommendation exists, it is used as the backbone and cannot be weakened (§H Skill Evolution). If a guideline is wrong for a particular population, the evolved skill cannot correct it.
invented entities (1)
  • Executable disease skill library (GuideSkill) independent evidence
    purpose: External, disease-indexed repository of compiled guideline criteria and case-derived rules that output ordinal support scores (tier 0–3).
    The library is the method's central artifact, and it is falsifiable through held-out predictions; the MedThink-Bench transfer (excluded from evolution) provides an external handle on whether the skills generalize.

pith-pipeline@v1.3.0-alltime-deepseek · 25341 in / 12099 out tokens · 118604 ms · 2026-08-01T00:36:01.608333+00:00 · methodology

0 comments
read the original abstract

Clinical practice guidelines (CPGs) encode diagnostic criteria, but LLM systems typically retrieve guideline text or absorb it through training rather than execute its rules. We introduce GuideSkill, an external reasoning layer that compiles disease-specific criteria into executable functions returning ordinal diagnostic-support scores. GuideSkill-Zero is initialized from guidelines, while GuideSkill-Evo uses case--diagnosis pairs to refine covered skills and add missing diagnoses. At inference, an LLM proposes a differential diagnosis, grounds the features required by each matched skill, and fuses its ranking with the executed skill scores. Across four benchmarks and four backbones, GuideSkill-Zero improves macro-average accuracy over guideline RAG by 13.45% on average. GuideSkill-Evo achieves the highest macro-average for every backbone, improves over direct inference by 18.49% relatively, and increases gold-label skill coverage from 56.5% to 99.5%. On Qwen3.5-9B, it also exceeds the strongest parameter-update baseline by 11.16% without updating the backbone. Expert evaluation further indicates that GuideSkill produces clinically sound and broadly acceptable skills, suggesting that its initialized and evolved rules are reliable and practically meaningful. These results support executable skills as a model-agnostic mechanism for combining guideline-derived procedures with case-derived diagnostic patterns.

Figures

Figures reproduced from arXiv: 2607.26160 by Hao Peng, Lang Cao, Simo Du, Tianyang Luo, Yue Guo, Yuhao Shen.

Figure 1
Figure 1. Figure 1: Overview of GuideSkill. Guideline recommendations are compiled into the initial executable library (GuideSkill-Zero); labeled cases refine covered skills and add missing diagnoses (GuideSkill-Evo); and, during inference, executed skill scores are fused with an LLM-generated differential ranking. The final diagnosis score combines the LLM ranking score and the skill-based evidence score: S(x, d) = αsLLM(x, … view at source ↗
Figure 2
Figure 2. Figure 2: Test-split diagnosis distribution by ICD-10 chapter [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Accuracy of GuideSkill-Zero and GuideSkill-Evo on initially covered and newly added ICD-10 categories. Evolution improves all new-skill settings while preserving or improving existing-skill performance in 11 of 16 settings. Method MedCaseReasoning ER-Reason MIMIC-CDM-FI MedThink-Bench Average Qwen3.59B Direct Inference 19.69 39.44 92.55 21.82 43.38 Fine-tuning w/ Guidelines (Staniek, Sokolov, and Riezler 2… view at source ↗
Figure 4
Figure 4. Figure 4: Sensitivity analysis of the fusion weight [PITH_FULL_IMAGE:figures/full_fig_p011_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Sensitivity analysis of candidate set size. We vary [PITH_FULL_IMAGE:figures/full_fig_p011_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

52 extracted references · 5 canonical work pages · 3 internal anchors

  1. [1]

    arXiv preprint arXiv:2505.11733 , year=

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

    Pith/arXiv arXiv

  2. [2]

    Nature , volume=

    Towards accurate differential diagnosis with large language models , author=. Nature , volume=. 2025 , doi=

    2025

  3. [3]

    Applied Sciences , volume=

    What disease does this patient have? a large-scale open domain question answering dataset from medical exams , author=. Applied Sciences , volume=. 2021 , publisher=

    2021

  4. [4]

    Nature medicine , volume=

    Evaluation and mitigation of the limitations of large language models in clinical decision-making , author=. Nature medicine , volume=. 2024 , publisher=. doi:10.1038/s41591-024-03097-1 , url=

    doi:10.1038/s41591-024-03097-1 2024

  5. [5]

    arXiv preprint arXiv:2505.22919 , year=

    Er-reason: A benchmark dataset for llm-based clinical reasoning in the emergency room , author=. arXiv preprint arXiv:2505.22919 , year=. doi:10.48550/arXiv.2505.22919 , url=

    Pith/arXiv arXiv doi:10.48550/arxiv.2505.22919

  6. [6]

    arXiv preprint arXiv:2601.03267 , year=

    Openai gpt-5 system card , author=. arXiv preprint arXiv:2601.03267 , year=

    Pith/arXiv arXiv

  7. [7]

    arXiv preprint arXiv:2505.09388 , year=

    Qwen3 technical report , author=. arXiv preprint arXiv:2505.09388 , year=

    Pith/arXiv arXiv

  8. [8]

    arXiv preprint arXiv:2311.16079 , year=

    Meditron-70b: Scaling medical pretraining for large language models , author=. arXiv preprint arXiv:2311.16079 , year=

    Pith/arXiv arXiv

  9. [9]

    arXiv preprint arXiv:2512.00601 , year=

    Clinical-R1: Empowering Large Language Models for Faithful and Comprehensive Reasoning with Clinical Objective Relative Policy Optimization , author=. arXiv preprint arXiv:2512.00601 , year=

    arXiv

  10. [10]

    arXiv preprint arXiv:2412.18925 , year=

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

    Pith/arXiv arXiv

  11. [11]

    arXiv preprint arXiv:2510.04284 , year=

    Doctor-r1: Mastering clinical inquiry with experiential agentic reinforcement learning , author=. arXiv preprint arXiv:2510.04284 , year=

    arXiv

  12. [12]

    International Conference on Human-Computer Interaction , pages=

    An Efficient Interaction Human-AI Synergy System Bridging Visual Awareness and Large Language Model for Intensive Care Units , author=. International Conference on Human-Computer Interaction , pages=. 2025 , organization=

    2025

  13. [13]

    arXiv preprint arXiv:2601.21340 , year=

    EHR-RAG: Bridging Long-Horizon Structured Electronic Health Records and Large Language Models via Enhanced Retrieval-Augmented Generation , author=. arXiv preprint arXiv:2601.21340 , year=

    arXiv

  14. [14]

    arXiv preprint arXiv:2509.20067 , year=

    Macd: Multi-agent clinical diagnosis with self-learned knowledge for llm , author=. arXiv preprint arXiv:2509.20067 , year=

    arXiv

  15. [15]

    arXiv preprint arXiv:2512.03838 , year=

    Training and Evaluation of Guideline-Based Medical Reasoning in LLMs , author=. arXiv preprint arXiv:2512.03838 , year=. doi:10.48550/arXiv.2512.03838 , url=

    doi:10.48550/arxiv.2512.03838

  16. [16]

    JMIR Formative Research , volume=

    Guideline-incorporated large language model-driven evaluation of medical records using MedCheckLLM , author=. JMIR Formative Research , volume=. 2025 , publisher=. doi:10.2196/53335 , url=

    work page doi:10.2196/53335 2025

  17. [17]

    arXiv preprint arXiv:2601.03475 , year=

    CPGPrompt: Translating Clinical Guidelines into LLM-Executable Decision Support , author=. arXiv preprint arXiv:2601.03475 , year=. doi:10.48550/arXiv.2601.03475 , url=

    doi:10.48550/arxiv.2601.03475

  18. [18]

    Enhancing Large Language Models for Clinical Decision Support by Incorporating Clinical Practice Guidelines

    Enhancing large language models for clinical decision support by incorporating clinical practice guidelines , author=. 2024 IEEE 12th International Conference on Healthcare Informatics (ICHI) , pages=. 2024 , organization=. doi:10.48550/arXiv.2401.11120 , url=

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2401.11120 2024

  19. [19]

    arXiv preprint arXiv:2510.05194 , year=

    Reinforcement Learning for Clinical Reasoning: Aligning LLMs with ACR Imaging Appropriateness Criteria , author=. arXiv preprint arXiv:2510.05194 , year=. doi:10.48550/arXiv.2510.05194 , url=

    doi:10.48550/arxiv.2510.05194

  20. [20]

    arXiv preprint arXiv:2312.02441 , year=

    Meddm: Llm-executable clinical guidance tree for clinical decision-making , author=. arXiv preprint arXiv:2312.02441 , year=. doi:10.48550/arXiv.2312.02441 , url=

    Pith/arXiv arXiv doi:10.48550/arxiv.2312.02441

  21. [21]

    MedGuideX: Internalizing Decision Logic from Executable Guidelines into Large Language Models for Clinical Reasoning

    MedGuideX: Internalizing Decision Logic from Executable Guidelines into Large Language Models for Clinical Reasoning , author=. arXiv preprint arXiv:2605.26567 , year=. doi:10.48550/arXiv.2605.26567 , url=

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2605.26567

  22. [22]

    arXiv preprint arXiv:2603.25158 , year=

    Trace2Skill: Distill Trajectory-Local Lessons into Transferable Agent Skills , author=. arXiv preprint arXiv:2603.25158 , year=. doi:10.48550/arXiv.2603.25158 , url=

    Pith/arXiv arXiv doi:10.48550/arxiv.2603.25158

  23. [23]

    arXiv preprint arXiv:2604.17870 , year=

    GraSP: Graph-Structured Skill Compositions for LLM Agents , author=. arXiv preprint arXiv:2604.17870 , year=

    Pith/arXiv arXiv

  24. [24]

    arXiv preprint arXiv:2604.05333 , year=

    Graph of Skills: Dependency-Aware Structural Retrieval for Massive Agent Skills , author=. arXiv preprint arXiv:2604.05333 , year=

    Pith/arXiv arXiv

  25. [25]

    arXiv preprint arXiv:2604.08377 , year=

    SkillClaw: Let Skills Evolve Collectively with Agentic Evolver , author=. arXiv preprint arXiv:2604.08377 , year=. doi:10.48550/arXiv.2604.08377 , url=

    Pith/arXiv arXiv doi:10.48550/arxiv.2604.08377

  26. [26]

    arXiv e-prints , pages=

    SkillRouter: Skill Routing for LLM Agents at Scale , author=. arXiv e-prints , pages=

  27. [27]

    An Empirical Study of Agent Skills for Healthcare: Practice, Gaps, and Governance

    An Empirical Study of Agent Skills for Healthcare: Practice, Gaps, and Governance , author=. arXiv preprint arXiv:2605.02709 , year=. doi:10.48550/arXiv.2605.02709 , url=

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2605.02709

  28. [28]

    bioRxiv , year=

    STELLA: Towards a Biomedical World Model with Self-Evolving Multimodal Agents , author=. bioRxiv , year=

  29. [29]

    arXiv preprint arXiv:2602.12670 , year=

    SkillsBench: Benchmarking how well agent skills work across diverse tasks , author=. arXiv preprint arXiv:2602.12670 , year=

    Pith/arXiv arXiv

  30. [30]

    arXiv preprint arXiv:2602.08234 , year=

    Skillrl: Evolving agents via recursive skill-augmented reinforcement learning , author=. arXiv preprint arXiv:2602.08234 , year=

    Pith/arXiv arXiv

  31. [31]

    arXiv preprint arXiv:2603.02766 , year=

    Evoskill: Automated skill discovery for multi-agent systems , author=. arXiv preprint arXiv:2603.02766 , year=

    Pith/arXiv arXiv

  32. [32]

    New England Journal of Medicine , volume=

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

    2006

  33. [33]

    Swiss medical weekly , volume=

    Diagnostic errors and flaws in clinical reasoning: mechanisms and prevention in practice , author=. Swiss medical weekly , volume=

  34. [34]

    2024 , publisher=

    Medical decision making , author=. 2024 , publisher=

    2024

  35. [35]

    npj Digital Medicine , year=

    Automating expert-level medical reasoning evaluation of large language models , author=. npj Digital Medicine , year=. doi:10.1038/s41746-025-02208-7 , url=

    doi:10.1038/s41746-025-02208-7

  36. [36]

    arXiv preprint arXiv:2604.07269 , year=

    Joint Optimization of Reasoning and Dual-Memory for Self-Learning Diagnostic Agent , author=. arXiv preprint arXiv:2604.07269 , year=

    Pith/arXiv arXiv

  37. [37]

    arXiv preprint arXiv:2603.27820 , year=

    Improving clinical diagnosis with counterfactual multi-agent reasoning , author=. arXiv preprint arXiv:2603.27820 , year=

    Pith/arXiv arXiv

  38. [38]

    2011 , address=

    Clinical Practice Guidelines We Can Trust , author=. 2011 , address=. doi:10.17226/13058 , url=

    work page doi:10.17226/13058 2011

  39. [39]

    2019 , note=

    International Statistical Classification of Diseases and Related Health Problems, 10th Revision , author=. 2019 , note=

    2019

  40. [40]

    Advances in Neural Information Processing Systems , volume=

    Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks , author=. Advances in Neural Information Processing Systems , volume=. 2020 , doi=

    2020

  41. [41]

    Advances in Neural Information Processing Systems , volume=

    Chain-of-Thought Prompting Elicits Reasoning in Large Language Models , author=. Advances in Neural Information Processing Systems , volume=. 2022 , doi=

    2022

  42. [42]

    Advances in Neural Information Processing Systems , volume=

    Language Models are Few-Shot Learners , author=. Advances in Neural Information Processing Systems , volume=. 2020 , doi=

    2020

  43. [43]

    2022 , doi=

    Hu, Edward J and Shen, Yelong and Wallis, Phillip and Allen-Zhu, Zeyuan and Li, Yuanzhi and Wang, Shean and Wang, Lu and Chen, Weizhu , booktitle=. 2022 , doi=

    2022

  44. [44]

    International Conference on Learning Representations , year=

    Decoupled Weight Decay Regularization , author=. International Conference on Learning Representations , year=. doi:10.48550/arXiv.1711.05101 , url=

    Pith/arXiv arXiv doi:10.48550/arxiv.1711.05101

  45. [45]

    arXiv preprint arXiv:2402.03300 , year=

    DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models , author=. arXiv preprint arXiv:2402.03300 , year=. doi:10.48550/arXiv.2402.03300 , url=

    Pith/arXiv arXiv doi:10.48550/arxiv.2402.03300

  46. [46]

    Proceedings of the 29th Symposium on Operating Systems Principles , pages=

    Efficient Memory Management for Large Language Model Serving with PagedAttention , author=. Proceedings of the 29th Symposium on Operating Systems Principles , pages=. 2023 , doi=

    2023

  47. [47]

    arXiv preprint arXiv:2306.05685 , year=

    Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena , author=. arXiv preprint arXiv:2306.05685 , year=. doi:10.48550/arXiv.2306.05685 , url=

    Pith/arXiv arXiv doi:10.48550/arxiv.2306.05685

  48. [48]

    2026 , howpublished=

    2026

  49. [49]

    2026 , url=

    Claude Sonnet 4.6 , author=. 2026 , url=

    2026

  50. [50]

    2026 , url=

    MedGemma 27B Instruction-Tuned Model Card , author=. 2026 , url=

    2026

  51. [51]

    2026 , howpublished=

    text-embedding-3-small Model Documentation , author=. 2026 , howpublished=

    2026

  52. [52]

    2026 , doi=

    Shen, Shuaike and Cheng, Wenduo and Ma, Mingqian and Turcan, Alistair and Zhang, Martin Jinye and Ma, Jian , journal=. 2026 , doi=

    2026