Pith. sign in

REVIEW 4 major objections 5 minor 47 references

A new framework, Condition-Gated Reasoning, embeds patient-specific conditions into knowledge-graph edges and gates traversal so that contraindicated answers are pruned, yielding large gains on conditional biomedical QA while matching state

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

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

T0 review · deepseek-v4-flash

2026-08-02 22:04 UTC pith:2AAO76KF

load-bearing objection A genuinely useful conditional-reasoning mechanism and benchmark, but the evaluation needs a label-quality audit and prompt-matched baselines before the headline gains can be trusted. the 4 major comments →

arxiv 2602.17911 v3 pith:2AAO76KF submitted 2026-02-20 cs.CL cs.AI

Condition-Gated Reasoning for Context-Dependent Biomedical Question Answering

Jash Rajesh Parekh , Wonbin Kweon , Joey Chan , Rezarta Islamaj , Robert Leaman , Pengcheng Jiang , Chih-Hsuan Wei , Zhizheng Wang
show 2 more authors
Zhiyong Lu Jiawei Han
This is my paper
classification cs.CL cs.AI
keywords conditional reasoningbiomedical question answeringknowledge graphsretrieval-augmented generationcontraindicationsmulti-hop reasoningbenchmarkgated traversal
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 show that biomedical question answering needs explicit conditionality: the same fact can have different correct answers depending on patient specifics like comorbidities or pregnancy. It introduces CondMedQA, a 100-question diagnostic benchmark where the correct answer changes under a stated condition, and Condition-Gated Reasoning (CGR), a framework that attaches conditions to knowledge-graph edges and blocks traversal of edges whose conditions the query violates. On CondMedQA, CGR reaches 82% exact match with a frontier LLM, 20 points above the strongest baseline, and it also beats baselines on two established multi-hop benchmarks. If right, the work argues that encoding constraints as first-class graph structure is a practical way to make medical QA safer and more interpretable, rather than relying on the LLM to infer them implicitly.

Core claim

The paper's central claim is that explicitly modeling the conditions under which medical relationships hold—rather than treating them as implicit LLM knowledge—makes retrieval-based biomedical QA more reliable on questions where a contraindication or special population changes the answer. CGR extracts n-tuples of the form (subject, relation, object, conditions) from documents, builds a directed graph, and during traversal gates each edge by evaluating whether the query context satisfies, violates, or is silent about each condition. Only edges whose conditions are not violated are traversed. With this mechanism, CGR achieves 82.00 EM on the new CondMedQA benchmark versus 62.00 for the stronge

What carries the argument

The central object is a condition-aware knowledge graph: edges are 4-tuples (u, r, v, C) where C is a list of conditions under which the relationship holds or is blocked, e.g., ⟨hypertension, treated_by, lisinopril, [¬bilateral renal artery stenosis, ¬pregnancy]⟩. Traversal is controlled by a gating function that blocks an edge if any of its conditions evaluates to false for the query, where the evaluation is done in one bulk LLM call returning true/false/null for every unique condition. This turns conditional reasoning into a graph reachability problem with an explicit 'don't know' policy (null conditions do not block), rather than a hidden inference inside the answer generator.

Load-bearing premise

The empirical claims rest on CondMedQA's gold labels being correct and unambiguous; the authors' own 30-question medical audit reached only 66.7% exact all-annotator agreement on answer accuracy, so if a meaningful fraction of the 100 labels are wrong or ambiguous, the reported gains cannot be attributed solely to conditional reasoning.

What would settle it

Take CondMedQA, have a panel of clinicians re-annotate all 100 answers with adjudication, and re-run the exact same experiments against the revised labels. If CGR's 20-point exact-match lead over the strongest baseline shrinks substantially or disappears, the claim that CGR reliably selects condition-appropriate answers is not supported.

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

If this is right

  • CGR can be applied to any domain where the correct answer depends on contextual constraints, not just biomedicine.
  • Because gating happens before answer generation, the evidence paths are auditable: a clinician or user can see which conditions blocked which candidates.
  • The method keeps pace with or exceeds graph-based and RAG baselines on non-conditional multi-hop benchmarks, so adding conditionality does not come at the cost of general QA performance.
  • Potential deployment direction: in clinical decision support, CGR could be paired with a final human review; the paper explicitly positions it as a research tool, not a standalone medical advisor.

Where Pith is reading between the lines

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

  • The reported margin on CondMedQA may be inflated by label ambiguity: the authors' own audit found only 66.7% exact three-annotator agreement on answer accuracy. Re-annotating with adjudicated disagreements could shrink the 20-point gap.
  • A much simpler baseline—standard retrieval followed by a post-hoc filter that drops any candidate matching a stated contraindication—might capture part of CGR's advantage; comparing against such a controller would isolate what the graph gating contributes over a rule check.
  • CGR's bottleneck per its own error analysis is retrieval and path ranking, not gating. Improving condition-aware ranking (e.g., boosting paths that match the query's conditions) is an obvious next step the paper identifies.
  • The gating policy treats 'null' (unknown) conditions as pass-through, which is conservative but could admit unsafe answers when the graph simply lacks a contraindication. A harm-averse variant might block unknown-condition edges in high-stakes settings.

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

4 major / 5 minor

Summary. The paper introduces CondMedQA, a 100-question benchmark for conditional biomedical question answering, and Condition-Gated Reasoning (CGR), a framework that extracts condition-aware n-tuples, builds a knowledge graph with edge-level conditions, gates traversal by an LLM-based condition lookup, ranks reasoning paths, and generates a final answer. On CondMedQA, CGR with GPT-5.2 reaches 82.00 EM versus 62.00 for the best baseline (StructRAG), with large gains also reported on MedHopQA (86.75 vs 75.75) and MedHopQA-Cond (80.00 vs 74.29). The paper includes an ablation showing that removing gating drops CGR from 82.0 to 57.0 EM on CondMedQA, hyperparameter sensitivity analyses, extraction-model ablations, and an error analysis. The central claim is that explicit condition gating yields more reliable condition-appropriate answers while matching or exceeding state-of-the-art performance on established biomedical QA benchmarks.

Significance. If the empirical claims hold, the contribution is valuable: a reusable gating mechanism for condition-aware reasoning, a diagnostic benchmark that targets an under-tested capability, and a complexity analysis showing that condition evaluation requires only O(1) LLM calls. The paper's strengths include the explicit encoding of conditions as first-class graph-edge attributes, the interpretable traversal, the multi-model evaluation, and the commitment to releasing code and data. The gating ablation (Table 4) provides internal evidence that the mechanism matters. However, the headline result depends on the correctness and unambiguousness of CondMedQA's gold labels, which are only weakly validated (66.7% exact all-annotator agreement on Answer Accuracy, AC1=0.67; §3.3, Table 2). Because roughly one third of audited labels are contested, and because the answer-generation prompt used for CGR differs from what is reported for baselines (Appendix G.4 vs §5.1.4), the 20-point EM gap on CondMedQA does not yet conclusively isolate conditional reasoning. These issues are fixable and do not negate the framework's promise, but they must be addressed before the central claim is established.

major comments (4)
  1. [§3.3 / Table 2 / Appendix F] The empirical core of the claim rests on CondMedQA gold labels being correct and unambiguous, but the reported audit found only 66.7% exact three-annotator agreement on Answer Accuracy (Gwet's AC1=0.67) and 56.7% on Question Quality (AC2=0.60). No item-level adjudication or labels for the remaining 70 questions are provided, and §7.1 does not list this as a limitation. Appendix F concretely shows ambiguity: Q99 asks which drug should be replaced in RIPE for an HIV patient, yet the gold answer is rifabutin (the replacement), and the model's literal reading (rifampin) is marked wrong; Q28's gold 'V/Q scan' is contested by 'MRI without contrast' which avoids radiation entirely. With about a third of audited labels contested on a 100-question benchmark, the 20-point EM gain over baselines may partly reflect label disputes rather than conditional reasoning. Please release adjudicated item-lev
  2. [§5.2 / §G.4 / §5.1.4] The comparison between CGR and baselines is conflated with answer-generation prompting. CGR's answer-generation prompt (Appendix G.4) explicitly instructs the model to provide the 'BEST AVAILABLE ANSWER', to treat 'Insufficient evidence' as a 'LAST RESORT', and to interpret warnings as evidence of use. The paper states that 'All baseline methods use GPT-5.2 for answer generation' but does not report the baseline prompt. If baselines were given a more conservative 'answer only if confident' instruction, the EM gap would be inflated by prompt engineering rather than by gating. Please use an identical generation protocol for all methods (or report the exact baseline prompts), and re-run the comparison under that protocol. The gating ablation in Table 4 is also more convincing if the no-gating variant uses the same prompt.
  3. [§5.1.1 (MedHopQA-Cond)] MedHopQA-Cond is described as 'a disjoint, clinically reviewed set of 35 conditional questions drawn from the same MedHopQA source but excluded from the 400-question split above.' This is a new evaluation set, not an established benchmark, yet its construction and review process are not specified. Who performed the clinical review, with what instructions and agreement? How were the 35 questions selected, and are they a random sample or an adversarial filter? Without this information, the 80.00 vs 74.29 EM result on MedHopQA-Cond is hard to interpret, and the claim that CGR 'matches or exceeds' SOTA on conditional benchmarks is only as strong as this subset's validity.
  4. [§5.3 / Table 4] The gating ablation, which shows a 25-point EM drop on CondMedQA when gating is removed, is informative but inherits the same gold-label problem: if a meaningful fraction of CondMedQA answers are disputed, the ablation's magnitude is also affected. Moreover, the ablation paragraph in §5.3 attributes the drop to gating, but the no-gating variant also changes the evidence assembly: ungated traversal may retrieve more paths, potentially overwhelming the answer-generation model. Please report the number of paths and evidence size in the gated vs ungated conditions, and verify that the only difference is edge filtering.
minor comments (5)
  1. [§3.3] The text says Krippendorff's alpha and Cohen's kappa 'are unreliable under highly skewed marginal distributions' and therefore Gwet's AC1/AC2 are used. This is a reasonable choice, but the reported AC1=0.67 for Answer Accuracy is moderate, not high; the paper should state that this indicates only fair-to-moderate reliability and discuss the implications. Also, 'AC2' for ordinal quality is defined, but no reference is given for the ordinal extension.
  2. [§5.3 / Table 5] In the extraction-model ablation, all three extractors achieve exactly 65.00 EM on CondMedQA, so the conclusion that 'higher quality n-tuple extraction leads to better CGR performance' is not supported by the CondMedQA column. The claim is supported by MedHopQA/MedHopQA-Cond F1 only. Also, the ablation uses a random 20-question subset (seed 42); please report confidence intervals or bootstrap estimates.
  3. [§4.3.1 / Eq. (1)] Equation (1) uses the notation '⊮[·]' for an indicator function; this is nonstandard and likely a typographical artifact. Use '𝟙[·]' or define the notation explicitly. Also, the example in §4.3.1 writes 'L(¬BRAS)=false', but the condition lookup table L is defined over conditions rather than negated literals; clarify how negated conditions are represented and evaluated.
  4. [§4.3.2 / Eq. (2)] The path ranking score in Eq. (2) is a sum over query keywords without normalization by the number of keywords, so queries with more keywords automatically have larger scores. Consider averaging or max-pooling over keywords, and state why the unnormalized sum was chosen.
  5. [§7 (Limitations)] The Limitations section mentions benchmark scope and knowledge-source dependence but omits the major limitation of gold-label reliability identified in §3.3 and Appendix F. This should be disclosed here, not only in the error analysis. Also, the reproducibility statement says 'We will release' — for a journal submission, please provide a URL or supplement at review time.

Circularity Check

0 steps flagged

No significant circularity: CGR's claims rest on external benchmarks and no prediction reduces to a fitted parameter or self-citation.

full rationale

CGR is an externally evaluated method: its output (Eq. 4) is a function of the query, gated traversal paths, and evidence snippets, all produced by pretrained models (Qwen2.5-14B for extraction, MedEmbed for ranking, GPT-5.2 for answer generation). No parameter is fitted to CondMedQA gold labels, and no result is defined in terms of the benchmark answers. The gating rule (Eq. 1) is an explicit mechanism, not a fitted prediction; the reported 82.00 EM on CondMedQA is a measurement against a held-out label set, and the MedHopQA, MedHopQA-Cond, and BioASQ-B results provide independent external support. Self-citations (SARG [33], MedHopQA [17]) appear only as related-work context and are not load-bearing for the framework's design or for the experimental conclusions. The substantive concerns in the paper are benchmark-validity issues—66.7% all-annotator agreement on Answer Accuracy, and Q99's wording conflicting with its gold label—but these are correctness risks, not circularity: they do not make CGR's derivation equivalent to its inputs. The central contribution is a reusable gating mechanism plus a diagnostic benchmark, both of which can be falsified on data not generated by the method itself.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 0 invented entities

Central claims draw on the benchmark labels (LLM-generated, author-verified, 30-item peer audit), on the reliability of the LLM condition evaluator, and on comparable prompt/evidence conditions for baselines. The method itself introduces no equations whose predictions reduce to fitted values; the main free knobs are retrieval/ranking hyperparameters and the unreported threshold/depth/seed.

free parameters (6)
  • k_paths (top-k reasoning paths) = 3
    Selected by hyperparameter sweep in §5.3/Fig 3 on the same benchmarks; performance degrades at k=10.
  • k_nodes (top-k entry nodes) = 5
    Selected by the same sweep; k=5 judged sufficient.
  • entry node selection threshold τ = not reported
    Mentioned in §4.3.1 without a stated value.
  • max traversal depth d_max = not reported
    Termination criterion in §4.3.1; no value is given.
  • BioASQ-B random 500-question sample seed = not stated
    Appendix B says the 500 questions were randomly sampled but no seed is reported, so the BioASQ-B column is not exactly reproducible.
  • extraction model choice = Qwen2.5-14B-Instruct-GPTQ-Int4
    Selected from the extraction-model ablation (§5.3); results differ across extractors.
axioms (5)
  • domain assumption Wikipedia articles used to generate CondMedQA answers are sufficiently accurate medical references; author review corrects any errors.
    Benchmark construction (§3.2) grounds Q/A in Wikipedia; no external clinical guideline audit is provided.
  • domain assumption LLM-based condition evaluation returns correct True/False/Null labels for unseen paraphrases, numeric age, and negations.
    §4.2.2 relies on Qwen2.5 to build the lookup table; no standalone accuracy measurement for the evaluator is reported.
  • domain assumption Unmentioned conditions do not invalidate an edge (null conditions do not block traversal).
    Eq. (1) only blocks on explicit false; in medicine, absent information is not evidence, but this assumption is unstated and could admit contraindicated paths.
  • ad hoc to paper MedEmbed-based top-k path ranking surfaces the condition-appropriate path when gating succeeds.
    Appendix F attributes 11/18 errors to gold paths existing in the KG but not in the top-k; ranking quality is assumed, not established.
  • ad hoc to paper All baselines received equivalent answer-generation instructions to CGR's.
    Only CGR's answer prompt is given (§G.4); baseline prompts are not shown, so instruction-level comparability is an assumption.

pith-pipeline@v1.3.0-alltime-deepseek · 17461 in / 16374 out tokens · 138104 ms · 2026-08-02T22:04:05.148528+00:00 · methodology

0 comments
read the original abstract

Current biomedical question answering (QA) systems often assume that medical knowledge applies uniformly, yet real-world clinical reasoning is inherently conditional: nearly every decision depends on patient-specific factors such as comorbidities and contraindications. Existing benchmarks do not evaluate such conditional reasoning, and retrieval-augmented or graph-based methods lack explicit mechanisms to ensure that retrieved knowledge is applicable to given context. To address this gap, we propose CondMedQA, the first benchmark for conditional biomedical QA, consisting of multi-hop questions whose answers vary with patient conditions. Furthermore, we propose Condition-Gated Reasoning (CGR), a novel framework that constructs condition-aware knowledge graphs and selectively activates or prunes reasoning paths based on query conditions. Our findings show that CGR more reliably selects condition-appropriate answers while matching or exceeding state-of-the-art performance on biomedical QA benchmarks, highlighting the importance of explicitly modeling conditionality for robust medical reasoning.

Figures

Figures reproduced from arXiv: 2602.17911 by Chih-Hsuan Wei, Jash Rajesh Parekh, Jiawei Han, Joey Chan, Pengcheng Jiang, Rezarta Islamaj, Robert Leaman, Wonbin Kweon, Zhiyong Lu, Zhizheng Wang.

Figure 1
Figure 1. Figure 1: Example of condition-gated reasoning in the biomedical domain. Existing KG-RAG extracts triples and retrieves contraindi￾cated treatments. CGR extracts n-tuples with patient-specific con￾ditions, gating unsafe paths and retrieving only safe alternatives. preferences shift under specific conditions. Biomedical question answering is inherently conditional, as clinical decision-making requires reasoning over … view at source ↗
Figure 2
Figure 2. Figure 2: Overview of Condition-Gated Reasoning (CGR) compared to traditional KG-RAG. Given a clinical query about a patient with hypertension and bilateral renal artery stenosis (BRAS), traditional KG-RAG extracts standard relation triples and traverses all paths indiscriminately, retrieving evidence for contraindicated treatments (e.g., Lisinopril, an ACE inhibitor). CGR extends triples to n-tuples that include pa… view at source ↗
Figure 3
Figure 3. Figure 3: Hyperparameter sensitivity analysis. Top row: varying 𝑘paths ∈ {1, 3, 5, 10}. Bottom row: varying 𝑘nodes ∈ {1, 3, 5, 10}. Based on these results, we set 𝑘paths = 3 and 𝑘nodes = 5 for all experiments. 6 Interdisciplinary Contributions Contributions to Biomedical Research. The CGR algorithm enables traceable conditional reasoning by constructing evidence trails grounded in published medical research. Additio… 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

47 extracted references · 9 linked inside Pith

  1. [1]

    2024.MedEmbed: Medical-Focused Embedding Models

    Abhinand Balachandran. 2024.MedEmbed: Medical-Focused Embedding Models. https://github.com/abhinand5/MedEmbed

    2024

  2. [2]

    Boyu Chen, Zirui Guo, Zidan Yang, Yuluo Chen, Junze Chen, Zhenghao Liu, Chuan Shi, and Cheng Yang. 2025. Pathrag: Pruning graph-based retrieval augmented generation with relational paths.arXiv preprint arXiv:2502.14902 (2025)

    arXiv 2025

  3. [3]

    Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. 2024. 8 Condition-Gated Reasoning for Context-Dependent Biomedical Question Answering Scaling instruction-finetuned language models.Journal of Machine Learning Research25, 70 (2024), 1–53

    2024

  4. [4]

    Jacob Cohen. 1960. A coefficient of agreement for nominal scales.Educational and psychological measurement20, 1 (1960), 37–46

    1960

  5. [5]

    Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. 2024. The llama 3 herd of models.arXiv e-prints(2024), arXiv–2407

    2024

  6. [6]

    Phan Minh Dung. 1995. On the acceptability of arguments and its fundamen- tal role in nonmonotonic reasoning, logic programming and n-person games. Artificial intelligence77, 2 (1995), 321–357

    1995

  7. [7]

    Darren Edge, Ha Trinh, Newman Cheng, Joshua Bradley, Alex Chao, Apurva Mody, Steven Truitt, Dasha Metropolitansky, Robert Osazuwa Ness, and Jonathan Larson. 2024. From Local to Global: A GraphRAG Approach to Query-Focused Summarization.arXiv preprint arXiv:2404.16130(2024)

    Pith/arXiv arXiv 2024

  8. [8]

    Yichun Feng, Jiawei Wang, Ruikun He, Lu Zhou, and Yixue Li. 2025. A retrieval- augmented knowledge mining method with deep thinking LLMs for biomedical research and clinical support.GigaScience14 (2025), giaf109

    2025

  9. [9]

    Melvin Fitting. 1992. Michael Gelfond and Vladimir Lifschitz. The stable model semantics for logic programming. Logic programming, Proceedings of the fifth international conference and symposium, Volume 2, edited by Robert A. Kowalski and Kenneth A. Bowen, Series in logic programming, The MIT Press, Cambridge, Mass., and London, 1988, pp. 1070–1080.-Kit F...

    1992

  10. [10]

    2000.Safe and Sound: Artificial Intelligence in Haz- ardous Applications

    John Fox and Subrata Das. 2000.Safe and Sound: Artificial Intelligence in Haz- ardous Applications. MIT Press, Cambridge, Mass

    2000

  11. [11]

    Gemini Team, Google. 2025. Gemini 3: Introducing the latest Gemini AI Model from Google. https://blog.google/products/gemini/gemini-3/ Accessed: 2025- 11-18

    2025

  12. [12]

    Zirui Guo, Lianghao Xia, Yanhua Yu, Tu Ao, and Chao Huang. 2025. LightRAG: Simple and Fast Retrieval-Augmented Generation. InFindings of the Association for Computational Linguistics: EMNLP 2025

    2025

  13. [13]

    Bernal Jiménez Gutiérrez, Yiheng Shu, Weijian Qi, Sizhe Zhou, and Yu Su. 2025. From rag to memory: Non-parametric continual learning for large language models.arXiv preprint arXiv:2502.14802(2025)

    Pith/arXiv arXiv 2025

  14. [14]

    Kilem Gwet. 2001. Handbook of inter-rater reliability.Gaithersburg, MD: STATAXIS Publishing Company(2001), 223–246

    2001

  15. [15]

    Kilem Li Gwet. 2008. Computing inter-rater reliability and its variance in the presence of high agreement.Brit. J. Math. Statist. Psych.61, 1 (2008), 29–48

    2008

  16. [16]

    Tzu-Sheng Hsu, Hung-Yi Chen, et al . 2024. KRAGEN: Knowledge-Graph- Augmented Generation for Biomedical Question Answering.arXiv preprint arXiv:2403.01340(2024)

    Pith/arXiv arXiv 2024

  17. [17]

    Rezarta Islamaj et al. 2025. Overview of BioCreative IX Track 1: MedHopQA – Multi-hop Biomedical Question Answering. InBioCreative IX Challenge and Workshop, IJCAI 2025

    2025

  18. [18]

    Pengcheng Jiang, Lang Cao, Ruike Zhu, Minhao Jiang, Yunyi Zhang, Jimeng Sun, and Jiawei Han. 2025. RAS: Retrieval-And-Structuring for Knowledge-Intensive LLM Generation.arXiv preprint arXiv:2502.10996(2025)

    arXiv 2025

  19. [19]

    Pengcheng Jiang, Siru Ouyang, Yizhu Jiao, Ming Zhong, Runchu Tian, and Jiawei Han. 2025. Retrieval and structuring augmented generation with large lan- guage models. InProceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2. 6032–6042

    2025

  20. [20]

    Xu, Luyu Gao, Pengfei Liu, and Graham Neubig

    Ziyan Jiang, Frank F. Xu, Luyu Gao, Pengfei Liu, and Graham Neubig. 2024. LongRAG: Enhancing Retrieval-Augmented Generation with Long-context LLMs. arXiv preprint arXiv:2406.15319(2024)

    Pith/arXiv arXiv 2024

  21. [21]

    Bernal Jimenez Gutierrez, Yi Shu, Yu Gu, Michihiro Yasunaga, and Yu Su. 2024. HippoRAG: Neurobiologically Inspired Long-Term Memory for Large Language Models. InAdvances in Neural Information Processing Systems (NeurIPS), Vol. 37

    2024

  22. [22]

    Qiao Jin, Bhuwan Dhingra, Zhengping Liu, William Cohen, and Xinghua Lu

  23. [23]

    Qiao Jin, Won Kim, Qingyu Chen, Donald C Comeau, Lana Yeganova, W John Wilbur, and Zhiyong Lu. 2023. Medcpt: Contrastive pre-trained transformers with large-scale pubmed search logs for zero-shot biomedical information retrieval. Bioinformatics39, 11 (2023), btad651

    2023

  24. [24]

    Yunsoo Kim, Yusuf Abdulle, and Honghan Wu. 2025. Biohopr: A benchmark for multi-hop, multi-answer reasoning in biomedical domain. InFindings of the Association for Computational Linguistics: ACL 2025. 12894–12908

    2025

  25. [25]

    Klaus Krippendorff. 2011. Computing Krippendorff’s alpha-reliability. (2011)

    2011

  26. [26]

    2025.vLLM: An Efficient Inference Engine for Large Language Models

    Woosuk Kwon. 2025.vLLM: An Efficient Inference Engine for Large Language Models. Ph. D. Dissertation. UC Berkeley

    2025

  27. [27]

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rock- täschel, et al. 2020. Retrieval-augmented generation for knowledge-intensive nlp tasks.Advances in neural information processing systems33 (2020), 9459–9474

    2020

  28. [28]

    Zhuoqun Li, Xuanang Chen, Haiyang Yu, Hongyu Lin, Yaojie Lu, Qiaoyu Tang, Fei Huang, Xianpei Han, Le Sun, and Yongbin Li. 2024. Structrag: Boosting knowledge intensive reasoning of llms via inference-time hybrid information structurization.arXiv preprint arXiv:2410.08815(2024)

    Pith/arXiv arXiv 2024

  29. [29]

    Liu, Kevin Lin, Danqi Chen, Christopher D

    Nelson F. Liu, Kevin Lin, Danqi Chen, Christopher D. Manning, Ravi Pandey, and Omer Levy. 2024. Lost in the middle: How language models use long contexts. Transactions of the Association for Computational Linguistics (TACL)12 (2024), 157–173

    2024

  30. [30]

    Haoran Luo, Haihong E, Guanting Chen, Yandan Zheng, Xiaobao Wu, Yikai Guo, Qika Lin, Yu Feng, Zemin Kuang, Meina Song, Yifan Zhu, and Anh Tuan Luu. 2025. HyperGraphRAG: Retrieval-Augmented Generation via Hypergraph-Structured Knowledge Representation. InAdvances in Neural Information Processing Systems (NeurIPS)

    2025

  31. [31]

    Yuanjie Lyu, Zihan Niu, Zheyong Xie, Chao Zhang, Tong Xu, Yang Wang, and Enhong Chen. 2024. Retrieve-plan-generation: An iterative planning and an- swering framework for knowledge-intensive llm generation.arXiv preprint arXiv:2406.14979(2024)

    Pith/arXiv arXiv 2024

  32. [32]

    John McCarthy. 1980. Circumscription—a form of non-monotonic reasoning. Artificial intelligence13, 1-2 (1980), 27–39

    1980

  33. [33]

    Jash Rajesh Parekh, Pengcheng Jiang, and Jiawei Han. 2025. Structure-Augmented Reasoning Generation. https://api.semanticscholar.org/CorpusID:279261260

    2025

  34. [34]

    Hongjin Qian, Li Zhu, Hongying Zhang, Yan Chen, Zhaocheng Wang, and Chao Chen. 2024. Slimmer: Real-time Memory-based Context Compression. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing (EMNLP)

    2024

  35. [35]

    Raymond Reiter. 1980. A logic for default reasoning.Artificial intelligence13, 1-2 (1980), 81–132

    1980

  36. [36]

    Peri L Schuyler, William T Hole, Mark S Tuttle, and David D Sherertz. 1993. The UMLS Metathesaurus: representing different views of biomedical concepts. Bulletin of the Medical Library Association81, 2 (1993), 217

    1993

  37. [37]

    Yucheng Shi, Shaochen Xu, Tianze Yang, Zhengliang Liu, Tianming Liu, Xiang Li, and Ninghao Liu. 2025. Mkrag: Medical knowledge retrieval augmented gen- eration for medical question answering. InAMIA Annual Symposium Proceedings, Vol. 2024. 1011

    2025

  38. [38]

    Aaditya Singh, Adam Fry, Adam Perelman, Adam Tart, Adi Ganesh, Ahmed El-Kishky, Aidan McLaughlin, Aiden Low, AJ Ostrow, Akhila Ananthram, et al

  39. [39]

    Qwen Team et al. 2024. Qwen2 technical report.arXiv preprint arXiv:2407.10671 2, 3 (2024)

    Pith/arXiv arXiv 2024

  40. [40]

    George Tsatsaronis et al. 2015. An overview of the BIOASQ large-scale biomedical semantic indexing and question answering competition.BMC Bioinformatics16, 1 (2015), 1–28

    2015

  41. [41]

    Peiyi Wang, Zelun Li, Hongyi Zhang, Yan Chen, Zhaocheng Wang, and Chao Chen. 2024. Long-Context Fine-Tuning of Large Language Models. InProceed- ings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics (NAACL)

    2024

  42. [42]

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

    2022

  43. [43]

    Johannes Welbl, Pontus Stenetorp, and Sebastian Riedel. 2018. Constructing Datasets for Multi-hop Reading Comprehension Across Documents.Transactions of the Association for Computational Linguistics (TACL)6 (2018), 287–302

    2018

  44. [44]

    Jing Zhang, Yifan Wang, et al. 2024. MedGraphRAG: Bridging Large Language Models and Domain-Specific Knowledge Graphs for Medical Question Answer- ing.arXiv preprint arXiv:2408.03988(2024)

    Pith/arXiv arXiv 2024

  45. [45]

    in children

    Xuejiao Zhao, Siyan Liu, Su-Yin Yang, and Chunyan Miao. 2025. Medrag: Enhanc- ing retrieval-augmented generation with knowledge graph-elicited reasoning for healthcare copilot. InProceedings of the ACM on Web Conference 2025. 4442–4457. 9 Parekh et al. A Use of Large Language Models In this work, large language models (LLMs) were used to assist with editi...

    2025

  46. [2019]

    In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing (EMNLP)

    PubMedQA: A Dataset for Biomedical Research Question Answering. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing (EMNLP)

    2019

  47. [2025]

    Openai gpt-5 system card.arXiv preprint arXiv:2601.03267(2025)

    Pith/arXiv arXiv 2025