Pith. sign in

REVIEW 3 major objections 6 minor 31 references

The paper argues that a weighted three-axis score — semantic coherence, diagnostic correctness, and clinical prioritization — can objectively grade LLM-written liver MRI reports, and that structured prompts with roughly 10–15 example report

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 ↗

A three-axis evaluation framework plus prompt-optimization study finds that instruction-rich prompts with 10–15 examples improve LLM-generated Chinese liver MRI report credibility, with Kimi-K2 and DeepSeek-V3 leading.

T0 review reviewed 2026-08-04 challenge →

load-bearing objection Large-scale prompt optimization study for Chinese liver MRI reports, but the Semantic Coherence metric is mislabeled and undermines some key interpretive claims. the 3 major comments →

arxiv 2510.23008 v3 pith:2IAFQWL7 submitted 2025-10-27 cs.AI

From Prompt Optimization to Multi-Dimensional Credibility Evaluation: Enhancing Trustworthiness of Chinese LLM-Generated Liver MRI Reports -- with Preliminary Extension to Lung Cancer

Qiuli Wang , Xinhuang Sun , Yonglin Chen , Jie Cheng , Yongxu Liu , Xingpeng Zhang , Xiaoming Li , Wei Chen This is my paper
classification cs.AI
keywords LLM report evaluationprompt optimizationliver MRI radiologycredibility assessmentsemantic coherencediagnostic correctnessclinical prioritizationChinese medical LLMs
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 reading

This paper tries to establish that the trustworthiness of Chinese LLM-written liver MRI reports can be scored objectively with a Multi-Dimensional Credibility Assessment (MDCA) framework that assigns a weighted score from three axes: semantic coherence (measured by BERT similarity to the ground-truth diagnosis sentence), diagnostic correctness (keyword matching of diagnosis targets), and clinical prioritization alignment (whether the most urgent findings lead the report). Using more than 15,000 real institutional reports, the authors also argue that prompt design matters more than model choice: structured instructions plus roughly 10–15 example reports from experienced radiologists give the best accuracy, fluency, and efficiency, while more than about 20 examples plateau or hurt performance. Across the models tested, Kimi-K2 and DeepSeek-V3 scored highest. A sympathetic reader would care because, if right, hospitals have a reproducible, model-agnostic way to assess and improve AI-generated reports without relying on LLM self-judgment or impractical manual review.

Core claim

The central claim is that a composite score MDCA = 0.2×SC + 0.4×DC + 0.4×CPA captures the clinical credibility of LLM-generated liver MRI reports in a model-agnostic and reproducible way, and that the same prompting recipe — role definition, task specification, tiered TOP diagnostic taxonomy, mandatory verification items, report structure standards, imaging principles, and about 10–15 example reports — produces consistent gains across different Chinese LLMs. The authors report that Kimi-K2-Instruct-0905 reached the highest composite score (76.149) and DeepSeek-V3 followed (75.410), and that example-based prompts primarily improved semantic coherence, while instruction components improved dia

What carries the argument

The MDCA framework: a composite score of 0.2×SC + 0.4×DC + 0.4×CPA, where SC = BERT(s_T, s_S) compares the generated diagnosis sentence against the ground-truth sentence using a pre-trained BERT language model, DC is keyword matching between the diagnosis targets in the generated and ground-truth reports, and CPA scores whether high-priority diagnoses appear in the top positions (Top-1, Top-3, Top-5). The prompt side of the machinery is an 11-configuration composition matrix that independently toggles six instruction components (role, core task, TOP taxonomy, verification items, structure standards, imaging principles) and varying numbers of expert example reports.

Load-bearing premise

The framework assumes that BERT similarity between the ground-truth diagnosis sentence and the generated sentence is a measure of textual coherence, when it is largely a semantic-similarity score against the gold diagnosis, entangling the 'coherence' axis with diagnostic accuracy.

What would settle it

Take a set of LLM-generated reports where the diagnosis sentence is fluent but the diagnosis is wrong (or the top finding is buried). Compute MDCA's SC for these reports and compare with radiologists' fluency ratings. If SC remains high while radiologists judge the prose fluent but the diagnosis wrong, the coherence dimension is doing its job; if SC drops sharply for wrong diagnoses, then SC is measuring diagnostic agreement, not coherence, and the framework's dimension separation is not as stated.

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

If this is right

  • If MDCA is valid, radiology departments can automate quality control of LLM reports by computing SC/DC/CPA in a deterministic, interpretable way instead of relying on LLM self-evaluation.
  • A prompt recipe of full instruction components plus 10–15 expert examples can serve as a baseline template for other institutions using Chinese liver MRI reporting.
  • Approximately 10–15 examples is a practical prompt-scaling guideline: performance rises to that point and then plateaus or mildly declines beyond 20.
  • Model ranking (Kimi-K2 and DeepSeek-V3 ahead of ByteDance-Seed and Qwen3) is stable across prompt complexities, suggesting the prompt design transfers across models.
  • The consistency check between MDCA scores and radiologist ratings suggests the framework can support trainee education and longitudinal monitoring of report quality.

Where Pith is reading between the lines

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

  • Because SC is computed as BERT similarity to the ground-truth diagnosis sentence, the reported 'semantic coherence' gains from examples likely reflect closer diagnostic mimicry, not simply smoother language; the 0.2 weight on SC may therefore be masking part of diagnostic agreement.
  • The roughly 10–15 example optimum suggests an economic corollary: when token budgets are tight, adding a few high-quality examples can substitute for verbose instructions.
  • The same protocol should extend naturally to other structured radiology reports (e.g., lung cancer CT), but the title's promised preliminary extension to lung cancer is not demonstrated in the body; testing it would require re-running the prompt matrix on a lung-cancer corpus.
  • The model-agnostic consistency indicates that investments in prompt design may dominate investments in model choice for this task — a hypothesis a hospital could test by A/B testing prompt designs on a single model before upgrading hardware.
Share X Bluesky LinkedIn Reddit HN

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 / 6 minor

Summary. The paper proposes a Multi-Dimensional Credibility Assessment (MDCA) framework for evaluating LLM-generated Chinese liver MRI diagnostic conclusions, and uses it to compare eleven prompt configurations and four Chinese LLMs on 15,127 institutional reports. MDCA scores reports on three dimensions — Semantic Coherence (SC), Diagnostic Correctness (DC), and Clinical Prioritization Alignment (CPA) — combined into a weighted composite. The authors report that Kimi-K2-Instruct-0905 achieves the highest composite score (76.149), closely followed by DeepSeek-V3 (75.410), that example-based prompts primarily enhance semantic coherence, and that approximately 10–15 examples strike the optimal performance/efficiency balance. The paper also includes a radiologist consistency check on 800 reports and an extension of the evaluation to the DeepSeek family (V3, V3.1, R1).

Significance. If the central measurement issues were resolved, this would be a valuable large-scale empirical study: it uses a real clinical corpus (15,127 reports), a broad prompt-composition matrix, multiple state-of-the-art Chinese LLMs under controlled API settings, and an independent radiologist rating check. The proposed prompt-design recipe and the model ranking are potentially actionable for radiology quality control and trainee education. The main contribution is the MDCA framework itself, but its construct validity is currently undermined by the definition of SC, which measures similarity to the ground-truth diagnosis rather than textual coherence. The authors also deserve credit for reporting standard deviations alongside means and for making the prompt components explicit in Tables I and II.

major comments (3)
  1. [Section II-B, Eq. (1)] The definition SC = BERT(s_T, s_S) computes similarity against the ground-truth diagnosis sentence s_T. This is not a measure of textual coherence or fluency; it is a content-agreement measure. The claim in the same paragraph that SC 'focused solely on textual coherence and did not assess diagnostic accuracy' is internally inconsistent. Because DC (Eq. 2) also compares the synthesized report to ground truth via keyword matching, SC and DC are not independent; the composite score in Eq. (4) double-counts diagnostic agreement. Consequently, the main mechanistic conclusion (Sections III-A, III-B, IV-A) that example-based prompts 'primarily enhanced semantic coherence' is not supported: the SC gains could simply reflect improved diagnostic content overlap with s_T. The Section III-D radiologist check does not rescue this dimension because it rates overall acceptability, not fluency. Please e
  2. [Section III-B, Table IV] The text lists Prompts 7, 4, 8, 6, 9, 10, and 11 as containing 0, 3, 5, 10, 15, and 20 examples, silently omitting the 25-example Prompt 11 (Table II). More substantively, the reported MDCA scores do not support the repeated claim that 10–15 examples is optimal (Abstract, Section III-B, Section VI). For Kimi-K2, MDCA rises monotonically from 0.7273 (P7) to 0.7696 (P11), with no clear plateau before 25; DeepSeek-V3 peaks at P10 (0.7613) and declines only slightly at P11 (0.7610). If the recommendation is based on efficiency, the paper needs an explicit score-per-token or cost-benefit analysis; as written, the claimed optimum is not demonstrated.
  3. [Section II-B, Eq. (4)] The MDCA weights 0.2/0.4/0.4 are introduced with the rationale that 'SC—calculated using BERT—focuses primarily on textual coherence... Therefore, the weight of SC may reasonably be set lower.' Since SC is not coherence (see Eq. (1) comment), this rationale collapses. Moreover, no sensitivity analysis is provided; the headline cross-model ranking (e.g., Kimi-K2 76.15 vs DeepSeek 75.41) and the prompt-comparison conclusions are all relative to this hand-set weighting. Please provide a robustness check over alternative weights (e.g., equal weights, weight extremes) or a principled derivation of the weights.
minor comments (6)
  1. [Section II-B, CPA derivation] In Step 2 of the CPA calculation, the example says t_RT2 would be matched with t_RS1, t_RS2, and t_RS2; the last should presumably be t_RS3. Also, the normalized score X_2 is described but the denominator is not fully specified for reports with missing or extra synthesized diagnoses.
  2. [Section III-B, Table IV] The list of example counts omits Prompt 11's 25 examples; the sentence should read 0, 3, 5, 10, 15, 20, and 25.
  3. [Section III-A, Table III] The text says Prompts 0 and 1 'achieved even lower scores,' but Table III shows that for both Kimi-K2 and DeepSeek-V3, Prompt 1 (role + task) is consistently lower than Prompt 0 (role only) on every metric. Grouping them together obscures this difference; please clarify whether adding a minimal task specification without structured support actively hurts performance.
  4. [Section III-C, Figure 4] The text states that 'all models demonstrated consistent performance gains with increasing prompt sophistication,' but then says ByteDance-Seed and Qwen3 showed limited improvements, particularly in SC and CPA. These statements need reconciliation, ideally with a table analogous to Table III for all four models.
  5. [Section II-D, DeepSeek family] The methods state that 'enable thinking' was uniformly set to False, yet DeepSeek-R1 is later evaluated 'with Chain-of-Thought.' Please clarify how Chain-of-Thought was enabled or disabled for each model in the DeepSeek family experiments.
  6. [Abstract] The abstract says 'more than 15,000 institutional liver cancer reports' were collected, but the study includes all eligible liver MRI reports (15,127), which include benign lesions and non-cancer diagnoses. Please align the wording with the actual inclusion criteria.

Circularity Check

0 steps flagged

No significant circularity: the evaluation is benchmarked against external ground truth and radiologist ratings; no self-citation chain or fitted prediction is present.

full rationale

The paper's central claims are empirical measurements rather than derivations that reduce to their own inputs. MDCA's submetrics are each defined against external ground-truth reports (Eqs. 1-3), and the composite weights in Eq. 4 (0.2/0.4/0.4) are set a priori with a stated rationale, not fitted to the reported MDCA outcomes. The 11 prompt configurations and multiple LLM comparisons are direct experiments on institutional reports; no parameter was fitted to a subset and then renamed a prediction. The radiologist consistency check in Section III-D provides an external benchmark against board-certified radiologist judgments, so the claim that MDCA 'effectively evaluates' report quality is not self-referential. There are no author self-citations serving as load-bearing evidence, no imported uniqueness theorems, and no ansatz smuggled in through prior work that itself rests on the same ansatz. The only substantive concern is construct validity of Semantic Coherence (Eq. 1), which measures BERT similarity to the target diagnosis rather than pure textual coherence; however, this is a metric-design/correctness issue, not circularity, because the observed direction and magnitude of SC changes are empirical and not forced by the equation. The limitations section appropriately acknowledges single-center, radiologist-only evaluation. Therefore no circularity is identified.

Axiom & Free-Parameter Ledger

2 free parameters · 3 axioms · 0 invented entities

The paper introduces no new physical or ontological entities. Its reliance on free parameters is limited to the hand-set MDCA weights and a data-driven example-count recommendation. The main epistemic burden is the domain assumptions about metric validity.

free parameters (2)
  • MDCA weights = 0.2 (SC), 0.4 (DC), 0.4 (CPA)
    Weights in Eq. (4) are chosen by hand based on the authors' judgment that SC is less important; no sensitivity analysis or justification from data. Model rankings may change under alternative weights.
  • Optimal example count = 10–15 examples
    Selected from the observed performance plateau around 20 examples; presented as a recommendation but is a data-derived threshold, not a predesignated parameter.
axioms (3)
  • domain assumption BERT similarity between target and synthesized diagnosis sentence measures semantic coherence/fluency
    Eq. (1) in Section II-B; no validation that BERT similarity corresponds to fluency rather than content overlap. This is the weakest link in the framework.
  • domain assumption Keyword matching (KMAT) between target and synthesized diagnosis terms accurately captures diagnostic correctness
    Section II-B; no details on synonym handling, terminology normalization, or partial matches, which is critical for Chinese medical text.
  • domain assumption Ground-truth diagnostic conclusions in the original clinical reports are the correct reference
    Section II-A; standard for retrospective evaluation but assumes original reports are error-free and that the reference standard is fixed.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of From Prompt Optimization to Multi-Dimensional Credibility Evaluation: Enhancing Trustworthiness of Chinese LLM-Generated Liver MRI Reports -- with Preliminary Extension to Lung Cancer." pith.science (2026). https://pith.science/paper/2IAFQWL7

@misc{pith2026251023008,
  author       = {Pith},
  title        = {Pith review of: From Prompt Optimization to Multi-Dimensional Credibility Evaluation: Enhancing Trustworthiness of Chinese LLM-Generated Liver MRI Reports -- with Preliminary Extension to Lung Cancer},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2IAFQWL7}},
  note         = {Machine review of arXiv:2510.23008}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Large language models (LLMs) have demonstrated promising performance in generating diagnostic conclusions from imaging findings, thereby supporting radiology reporting, trainee education, and quality control. However, systematic guidance on how to optimize prompt design across different clinical contexts remains underexplored. Moreover, a comprehensive and standardized framework for assessing the trustworthiness of LLM-generated radiology reports is yet to be established. This study aims to enhance the trustworthiness of LLM-generated liver MRI reports by introducing a Multi-Dimensional Credibility Assessment (MDCA) framework and providing guidance on institution-specific prompt optimization. The proposed framework is applied to evaluate and compare the performance of several advanced LLMs, including Kimi-K2-Instruct-0905, Qwen3-235B-A22B-Instruct-2507, DeepSeek-V3, and ByteDance-Seed-OSS-36B-Instruct, using the SiliconFlow platform.

Figures

Figures reproduced from arXiv: 2510.23008 by Jie Cheng, Qiuli Wang, Wei Chen, Xiaoming Li, Xingpeng Zhang, Xinhuang Sun, Yonglin Chen, Yongxu Liu.

Figure 1
Figure 1. Figure 1: The collection process of clinical reports. This [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Experimental results for Prompt Designs with KIMI-K2 and DeepSeek-V3. The figure illustrates the [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Experimental Results with Different Example Numbers. The figure illustrates how varying the number of [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Experimental Results with Four LLMs. The figure illustrates the performance of KIMI-K2, DeepSeek-V3, [PITH_FULL_IMAGE:figures/full_fig_p011_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Experimental Results with DeepSeek Family. The figure illustrates the performance of DeepSeek-V3, [PITH_FULL_IMAGE:figures/full_fig_p012_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Combined Scores and Proportions of Radiologist [PITH_FULL_IMAGE:figures/full_fig_p012_6.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

31 extracted references · 3 linked inside Pith

  1. [1]

    Immunotherapies for hepatocellular carci- noma,

    J. M. Llovet, F. Castet, M. Heikenwalder, M. K. Maini, V . Mazzaferro, D. J. Pinato, E. Pikarsky, A. X. Zhu, and R. S. Finn, “Immunotherapies for hepatocellular carci- noma,”Nature reviews Clinical oncology, vol. 19, no. 3, pp. 151–172, 2022

    2022

  2. [2]

    Management of hepatocellular carcinoma,

    J. Bruix and M. Sherman, “Management of hepatocellular carcinoma,”Hepatology, vol. 42, no. 5, pp. 1208–1236, 2005

    2005

  3. [3]

    Hepatocellular carcinoma,

    A. V ogel, T. Meyer, G. Sapisochin, R. Salem, and A. Saborowski, “Hepatocellular carcinoma,”The Lancet, vol. 400, no. 10360, pp. 1345–1362, 2022. [Online]. Available: https://www.sciencedirect.com/science/article/ pii/S0140673622012004

    2022

  4. [4]

    T. Aoki, N. Nishida, K. Ueshima, M. Morita, H. Chishina, M. Takita, S. Hagiwara, H. Ida, Y . Minami, A. Yamada et al., “Higher enhancement intrahepatic nodules on the hepatobiliary phase of gd-eob-dtpa-enhanced mri as a poor responsive marker of anti-pd-1/pd-l1 monotherapy for unresectable hepatocellular carcinoma,”Liver Cancer, vol. 10, no. 6, pp. 615–628, 2021

    2021

  5. [5]

    Gd-eob-dtpa-enhanced mri proves advantageous in selecting surgical candidates for patients with early-stage hepatocellular carcinoma: An analysis in terms of oncological outcomes,

    Z. Ye, J. Zhao, D. Hu, Z. Yang, J. Chen, L. Xu, Z. Zhou, M. Chen, and Y . Zhang, “Gd-eob-dtpa-enhanced mri proves advantageous in selecting surgical candidates for patients with early-stage hepatocellular carcinoma: An analysis in terms of oncological outcomes,”iLIVER, vol. 3, no. 4, p. 100117, 2024

    2024

  6. [6]

    A large language model improves clinicians’ diagnostic performance in complex critical illness cases,

    X. Wu, Y . Huang, and Q. He, “A large language model improves clinicians’ diagnostic performance in complex critical illness cases,”Critical Care, vol. 29, no. 1, p. 230, 2025

    2025

  7. [7]

    Kimi k2: Open agentic intelligence,

    K. Team, Y . Bai, Y . Bao, G. Chen, J. Chen, N. Chen, R. Chen, Y . Chen, Y . Chen, Y . Chenet al., “Kimi k2: Open agentic intelligence,”arXiv preprint arXiv:2507.20534, 2025

    Pith/arXiv arXiv 2025

  8. [8]

    Qwen3 technical report,

    A. Yang, A. Li, B. Yang, B. Zhang, B. Hui, B. Zheng, B. Yu, C. Gao, C. Huang, C. Lvet al., “Qwen3 technical report,”arXiv preprint arXiv:2505.09388, 2025

    Pith/arXiv arXiv 2025

  9. [9]

    Seed1. 5-thinking: Advancing superb reasoning models with reinforcement learning,

    B. Seed, J. Chen, T. Fan, X. Liu, L. Liu, Z. Lin, M. Wang, C. Wang, X. Wei, W. Xuet al., “Seed1. 5-thinking: Advancing superb reasoning models with reinforcement learning,”arXiv preprint arXiv:2504.13914, 2025

    arXiv 2025

  10. [10]

    Fuzzy reasoning chain: An innovative reasoning framework from fuzziness to clarity,

    P. Chen, X. Liu, Z. Liu, Z. Chen, X. Zhang, H. Hu, Z. Wang, K. Wang, S. Shi, and S. Lian, “Fuzzy reasoning chain: An innovative reasoning framework from fuzziness to clarity,”arXiv preprint arXiv:2509.22054, 2025

    arXiv 2025

  11. [11]

    Evaluating large language model workflows in clinical decision support for triage and referral and diagnosis,

    F. Gaber, M. Shaik, F. Allega, A. J. Bilecz, F. Busch, K. Goon, V . Franke, and A. Akalin, “Evaluating large language model workflows in clinical decision support for triage and referral and diagnosis,”npj Digital Medicine, vol. 8, no. 1, p. 263, 2025

    2025

  12. [12]

    Will generative large language models become radiologists’ invaluable allies?

    C. Marrocchio and N. Sverzellati, “Will generative large language models become radiologists’ invaluable allies?” p. e251259, 2025

    2025

  13. [13]

    Benchmark evaluation of deepseek large language models in clinical decision- making,

    S. Sandmann, S. Hegselmann, M. Fujarski, L. Bickmann, B. Wild, R. Eils, and J. Varghese, “Benchmark evaluation of deepseek large language models in clinical decision- making,”Nature Medicine, pp. 1–1, 2025

    2025

  14. [14]

    Comparative benchmarking of the deepseek large language model on medical tasks and clinical reasoning,

    M. Tordjman, Z. Liu, M. Yuce, V . Fauveau, Y . Mei, J. Hadjadj, I. Bolger, H. Almansour, C. Horst, A. S. Parihar et al., “Comparative benchmarking of the deepseek large language model on medical tasks and clinical reasoning,” Nature medicine, pp. 1–1, 2025

    2025

  15. [15]

    Multimedia- enhanced radiology reports: Concept, components, and challenges,

    L. Folio, L. Machado, and A. Dwyer, “Multimedia- enhanced radiology reports: Concept, components, and challenges,”RadioGraphics, vol. 38, pp. 462–482, 03 2018

    2018

  16. [16]

    Credibility assessment of a mech- anistic model of atherosclerosis to predict cardiovascu- lar outcomes under lipid-lowering therapy,

    Y . Wang, E. Courcelles, E. Peyronnet, S. Porte, A. Di- atchenko, E. Jacob, D. Angoulvant, P. Amarenco, F. Boc- cara, B. Cariouet al., “Credibility assessment of a mech- anistic model of atherosclerosis to predict cardiovascu- lar outcomes under lipid-lowering therapy,”npj Digital Medicine, vol. 8, no. 1, p. 171, 2025

    2025

  17. [17]

    Towards conversational diagnostic artificial intelligence,

    T. Tu, M. Schaekermann, A. Palepu, K. Saab, J. Freyberg, R. Tanno, A. Wang, B. Li, M. Amin, Y . Chenget al., “Towards conversational diagnostic artificial intelligence,” Nature, pp. 1–9, 2025

    2025

  18. [18]

    Opti- mizing large language models in radiology and mitigating pitfalls: prompt engineering and fine-tuning,

    T. T. Kim, M. Makutonin, R. Sirous, and R. Javan, “Opti- mizing large language models in radiology and mitigating pitfalls: prompt engineering and fine-tuning,”RadioGraph- ics, vol. 45, no. 4, p. e240073, 2025

    2025

  19. [19]

    Potential of chatgpt and gpt-4 for data mining of free-text ct reports on lung cancer,

    M. A. Fink, A. Bischoff, C. A. Fink, M. Moll, J. Kroschke, L. Dulz, C. P. Heußel, H.-U. Kauczor, and T. F. Weber, “Potential of chatgpt and gpt-4 for data mining of free-text ct reports on lung cancer,”Radiology, vol. 308, no. 3, p. e231362, 2023

    2023

  20. [20]

    Visual- textual attentive semantic consistency for medical report generation,

    Y . Zhou, L. Huang, T. Zhou, H. Fu, and L. Shao, “Visual- textual attentive semantic consistency for medical report generation,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 3985–3994

    2021

  21. [21]

    Visual-linguistic diagnostic semantic enhancement for medical report generation,

    J. Chen, G. Huang, X. Yuan, G. Zhong, Z. Tan, C.-M. Pun, and Q. Yang, “Visual-linguistic diagnostic semantic enhancement for medical report generation,”Journal of Biomedical Informatics, vol. 161, p. 104764, 2025

    2025

  22. [22]

    Autocodebench: Large language models are automatic code benchmark generators,

    J. Chou, A. Liu, Y . Deng, Z. Zeng, T. Zhang, H. Zhu, J. Cai, Y . Mao, C. Zhang, L. Tanet al., “Autocodebench: Large language models are automatic code benchmark generators,”arXiv preprint arXiv:2508.09101, 2025

    Pith/arXiv arXiv 2025

  23. [23]

    Deepseek’s “low-cost

    D. Zeng, Y . Qin, B. Sheng, and T. Y . Wong, “Deepseek’s “low-cost” adoption across china’s hospital systems: Too fast, too soon?”Jama, vol. 333, no. 21, pp. 1866–1869, 2025

    2025

  24. [24]

    Large language models for automated synoptic reports and resectability categoriza- tion in pancreatic cancer,

    R. Bhayana, B. Nanda, T. Dehkharghanian, Y . Deng, N. Bhambra, G. Elias, D. Datta, A. Kambadakone, C. G. Shwaartz, C.-A. Moultonet al., “Large language models for automated synoptic reports and resectability categoriza- tion in pancreatic cancer,”Radiology, vol. 311, no. 3, p. e233117, 2024

    2024

  25. [25]

    Radsearch, a semantic search model for accurate radiol- ogy report retrieval with large language model integration,

    C. H. Savage, G. Chaudhari, A. D. Smith, and J. H. Sohn, “Radsearch, a semantic search model for accurate radiol- ogy report retrieval with large language model integration,” Radiology, vol. 315, no. 1, p. e240686, 2025

    2025

  26. [26]

    A new step forward in the extraction of appropriate radiology reports,

    K. Yasaka and O. Abe, “A new step forward in the extraction of appropriate radiology reports,” p. e250867, 2025

    2025

  27. [27]

    Multisequence 3-t image synthesis from 64-mt low-field-strength mri using generative adversarial networks in multiple sclerosis,

    A. Lucas, T. C. Arnold, S. V . Okar, C. Vadali, K. D. Kawatra, Z. Ren, Q. Cao, R. T. Shinohara, M. K. Schindler, K. A. Daviset al., “Multisequence 3-t image synthesis from 64-mt low-field-strength mri using generative adversarial networks in multiple sclerosis,”Radiology, vol. 315, no. 1, p. e233529, 2025

    2025

  28. [28]

    Igniting language intelligence: The hitchhiker’s guide from chain- of-thought reasoning to language agents,

    Z. Zhang, Y . Yao, A. Zhang, X. Tang, X. Ma, Z. He, Y . Wang, M. Gerstein, R. Wang, G. Liuet al., “Igniting language intelligence: The hitchhiker’s guide from chain- of-thought reasoning to language agents,”ACM Computing Surveys, vol. 57, no. 8, pp. 1–39, 2025

    2025

  29. [29]

    Faithful chain- of-thought reasoning,

    Q. Lyu, S. Havaldar, A. Stein, L. Zhang, D. Rao, E. Wong, M. Apidianaki, and C. Callison-Burch, “Faithful chain- of-thought reasoning,” inThe 13th International Joint Conference on Natural Language Processing and the 3rd Conference of the Asia-Pacific Chapter of the Associa- tion for Computational Linguistics (IJCNLP-AACL 2023), 2023

    2023

  30. [30]

    Insights into deepseek- v3: Scaling challenges and reflections on hardware for ai architectures,

    C. Zhao, C. Deng, C. Ruan, D. Dai, H. Gao, J. Li, L. Zhang, P. Huang, S. Zhou, S. Maet al., “Insights into deepseek- v3: Scaling challenges and reflections on hardware for ai architectures,” inProceedings of the 52nd Annual Inter- national Symposium on Computer Architecture, 2025, pp. 1731–1745

    2025

  31. [31]

    Deepseek-r1 and gpt-4 are comparable in a complex diagnostic challenge: a historical control study,

    L. Chan, X. Xu, and K. Lv, “Deepseek-r1 and gpt-4 are comparable in a complex diagnostic challenge: a historical control study,”International Journal of Surgery, vol. 111, no. 6, pp. 4056–4059, 2025

    2025

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.