Pith. sign in

REVIEW 4 major objections 5 minor 69 references

PDD-RRG: Posterior Diagnostic Decision for Study-level Radiology Report Generation

T0 review · 4 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read The paper claims that adding a post-hoc Bayesian decision layer that fuses reports from varied input views improves radiology report clinical accuracy by up to 3.60 macro-F1 points on MIMIC-CXR without retraining the base generator.

desk verdict A useful post-hoc decision layer for RRG with solid label-level gains, but the report-level gain isn't cleanly attributed because the GPT-4 refinement step is uncontrolled. read the letter →

arxiv 2608.03055 v1 pith:55K3PVZF submitted 2026-08-04 cs.CV cs.CL

Yang Yu , Yiming Ji , Bin Dai , Dong Zhang , Zhiyong Zhou , Shoushan Li , Yakang Dai This is my paper · ORCID
classification cs.CVcs.CL
keywords radiologyreportgenerationposteriordiagnosticdecisionlikelihood-ratioaggregationBayesianmulti-viewinputsCheXbertlabelsMIMIC-CXRclinicalefficacy
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper argues that the usual one-pass, richest-input radiology report generation leaves diagnostic accuracy on the table, because adding more views or historical data can actually make a model miss findings it would have caught with less input. It proposes a post-hoc decision layer, PDD-RRG, that runs the existing frozen generator on several input subsets, turns each draft into a 14-class CheXbert label vector, and fuses them with a Bayesian likelihood-ratio posterior plus observation-specific thresholds. The fused diagnosis then drives an LLM to revise the most self-consistent draft report. On MIMIC-CXR, applying this layer to three strong baselines improves macro-F1 over the 14 observations by about 0.89 to 3.60 points without retraining any backbone. The point matters because it reframes RRG from pure generation to generation-plus-decision, and it offers a plug-in upgrade for existing models.

What carries the argument

The load-bearing object is the likelihood-ratio Bayesian posterior of Eq. (9): for each of the 14 clinical observations, the prior odds from training-set prevalence are multiplied by likelihood ratios $LR^{+}$ and $LR^{-}$ computed per input type, where a positive or uncertain prediction contributes $LR^{+}=R/FP$ and a negative prediction contributes $LR^{-}=(1-R)/(1-FP)$, with an amplification factor $\alpha$ for negative votes. The posterior log-odds is then thresholded by two observation-specific values, $th^{drop}_o$ and $th^{add}_o$, learned on the validation set to control false positives and false negatives asymmetrically. A consensus-based selection step picks the base report to edit. This machinery turns latent per-view diagnostic signals into a single revised diagnosis.

What would settle it

Measure the calibration of the posterior probabilities over the 14 observations on the test set, grouped by how many hypotheses share the same frontal image: if overconfidence grows as overlap increases, the independence assumption behind Eq. (9) is driving the decisions. Alternatively, restrict input configurations to disjoint image sets so the conditional-independence assumption holds by construction; if the macro-F1 gains vanish, the reported improvement depends on the very dependence the posterior ignores.

Watch

Extended reading notes

Core claim

The paper claims that a study-level radiology report can be made clinically more accurate by explicitly decoupling diagnostic decision-making from report generation: generate multiple draft reports from deliberately varied input configurations, aggregate their CheXbert labels through a Bayesian posterior that weights each input type by its training-set recall and false-positive rate for each observation, and apply per-observation add/drop thresholds tuned on validation data to decide the final status of each of 14 findings. The final 14-label diagnosis is then used to revise the consensus-selected base report via a frozen LLM. The authors show that this pipeline, which requires no retraining of the underlying RRG model, consistently outperforms both the richest-input single report and majority voting, with the largest gains coming from recovering false negatives that the richer input path missed.

Load-bearing premise

The method treats the reports generated from overlapping input subsets as independent pieces of evidence, even though they share the same frontal image, so the same model mistake can be counted several times and inflate the confidence of the final diagnosis.

Editorial extensions

If this is right

  • Applying PDD-RRG to MAIRA-2, LLM-RG4, and MLRG raises macro-F1 over 14 CheXbert observations by 3.60, 0.89, and 2.72 points on MIMIC-CXR test set.
  • The decision layer recovers far more false negatives than majority voting (61.35% vs 17.67% of recoverable FNs), while voting's false-positive suppression is largely a passive consensus-of-omission artifact.
  • The most input-complete report (Raw) is often clinically inferior to the consensus-selected report or the posterior-aggregated labels, so richer input does not guarantee better diagnosis.
  • No retraining or architectural change is needed; the framework edits the selected draft with a frozen LLM, so it can be layered onto any existing RRG model.

Reading between the lines

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

  • Because overlapping input views violate the conditional-independence assumption behind Eq. (9), the posterior probabilities are likely overconfident in proportion to how many hypotheses share the same frontal image; a reliability-diagram calibration check would separate the contribution of likelihood weighting from that of threshold tuning.
  • The method's decision vocabulary is fixed to CheXbert's 14 classes; extending the likelihood-ratio machinery to finer-grained or continuous severity labels is a natural next step that would require re-estimating the per-input-type statistics.
  • The LLM refinement step is a largely unexamined variable: substituting a cheaper editor or a rule-based sentence modifier would reveal whether the reported gains come from the posterior decision itself or from the language model's rewriting.
  • If the independence violation is addressed with a joint evidence model that treats the shared images once, the aggregation idea could transfer to other multi-source clinical report settings beyond chest X-rays.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes PDD-RRG, a post-hoc diagnostic decision layer for radiology report generation (RRG). Given a study with multiple current views, prior images, and prior reports, the method generates several draft reports by feeding different input subsets to a frozen RRG model, extracts 14 CheXbert observations from each draft, selects a consensus base report, computes per-observation Bayesian posterior probabilities using training-set likelihood ratios per input type, applies validation-tuned observation-specific drop/add thresholds to obtain an aggregated diagnosis, and finally uses GPT-4 to revise the base report to reflect this diagnosis. Experiments on MIMIC-CXR with MAIRA-2, LLM-RG4, and MLRG report improved macro/micro CheXbert F1 at both the label level and the report level, together with ablations, a majority-voting comparison, and a detailed error-repair analysis.

Significance. If the central claim holds, PDD-RRG is a useful model-agnostic, retraining-free plug-in that improves clinical efficacy of existing RRG systems, which is potentially valuable for medical AI deployment. The paper has real strengths: the experimental design covers three strong backbones with consistent reported gains; the code is promised; the ablation of likelihood-ratio modeling and observation-specific thresholds is sensible; and the comparison with majority voting, including the per-lesion repair analysis, is informative. However, two load-bearing issues prevent acceptance as-is: the Bayesian posterior in Eq. (9) relies on a conditional-independence assumption that is implausible for hypotheses sharing overlapping inputs, and the reported report-level gains are not cleanly attributed to the decision module because no control isolates the GPT-4 refinement stage. In addition, the significance of the modest report-level gains (e.g., LLM-RG4 +0.41 macro-F1 14) is not established without error bars or significance tests.

major comments (4)
  1. [Section 3.3, Eq. (9)] Eq. (9) computes posterior log-odds by summing alpha-weighted log likelihood ratios over all hypotheses as if the hypotheses were conditionally independent given the true label. This is not justified: the input configurations share overlapping images and priors (e.g., CF1-only vs. CF1+CL vs. CF1+prior+CL), so a model error on the shared CF1 image or prior report would appear in multiple hypotheses, inducing dependence. The resulting posterior probabilities are likely overconfident, and although the validation-tuned thresholds in Eq. (14) can partially absorb miscalibration, the independence assumption is load-bearing for the stated posterior interpretation and is never stated or tested. Please either justify the assumption, test posterior calibration on the validation set, or replace Eq. (9) with an aggregation that accounts for shared evidence.
  2. [Section 3.3, Eq. (15), and Table 1] The central reported artifact is the final report, but the contribution of the PDD label vector is not isolated from the GPT-4 refinement stage. Eq. (15) instructs GPT-4 to revise the base report using the PDD diagnosis, so the LLM is deliberately aligning the report with the PDD labels. Table 1 shows that report-level gains are smaller than label-level gains (e.g., LLM-RG4 macro-F1 14: +0.41 PDD-report vs. +0.89 PDD label-level), and Table 2's ablations are computed only on label-level PDD outputs. Without a control that applies the same GPT-4 revision protocol to the base report using the base report's own CheXbert labels (or majority-vote labels), the observed report-level improvement cannot be attributed to the posterior decision module; the LLM decoder itself could be responsible. This control is essential for the paper's central mechanistic claim.
  3. [Table 1] The reported gains lack measures of uncertainty. With a single test split and no error bars, bootstrap confidence intervals, or significance tests, small differences such as LLM-RG4 PDD-report macro-F1 14 (42.08 vs. 41.67, +0.41) may be within noise. Given that the headline claim is consistent improvement across backbones, please report variance estimates or at least bootstrap CIs for the main clinical metrics, especially the report-level numbers.
  4. [Section 3.3, Eq. (12)] The definition of alpha_{t,o} for uncertain predictions appears ill-posed as written: the denominator P(hat{h}=−1, type=t, hat{h}∈H_train) conditions on an event that is always true for training hypotheses, so it reduces to P(hat{h}=−1, type=t), not a conditional probability on negative labels. As written, Eq. (12) is not a likelihood ratio and its role in Eq. (9) is unclear. Please clarify the intended conditioning and correct the formula.
minor comments (5)
  1. [Section 3.2] The eight input types for MAIRA-2 are mentioned but not enumerated; the text says different parts of prior reports can be used, but the exact set of input configurations used in the experiments is not specified. Please list the eight types for reproducibility.
  2. [Section 3.3, Eq. (2)] The similarity function assigns 0.5 when one side is uncertain and the other is 0/1, but the rationale for this value, and for treating null as completely dissimilar, is not explained. A brief justification would improve interpretability.
  3. [Section 3.3, Eq. (13)] The decision rule contains a typo-prone condition: 'If both thresholds are met (th_add <= p < th_drop)' should presumably be 'th_add <= p < th_drop', but since thresholds are learned independently, this case may be impossible or ambiguous; please clarify what happens when th_drop <= p < th_add.
  4. [Section 4.1] The evaluation maps 'uncertain' labels to negative, but the decision module treats uncertain predictions as a separate category in Eq. (12). This asymmetry is not discussed; please state explicitly how uncertain labels from CheXbert are handled in training, validation, and evaluation.
  5. [General] The phrase 'first work to introduce a decision layer' in the contributions is strong; please soften it to 'first, to our knowledge' and consider citing any prior work on post-hoc label aggregation in medical report generation or related parse-and-refine pipelines.

Circularity Check

1 steps flagged · score 4.0 of 10

Report-level clinical-efficacy gains are partly constructed: the final report is generated from the same CheXbert label vector that is later used to score it, although the label-level PDD results remain an independent test-set prediction.

  1. fitted input called prediction [Section 3.3, Eq. (15) and Section 4.1, Evaluation Metrics]
    "Finally, we employ a LLM to revise ybase according to ˆh*, resulting in the final report ˆy*. ... Specifically, we map “uncertain” labels to negative and compute macro- and micro-average F1 scores for 14 observations extracted from reports."

    The final report is produced by prompting GPT-4 to revise ybase according to h* (Eq. 15), and h* is exactly the 14-category CheXbert label vector output by the posterior decision module. The clinical-efficacy metric in Section 4.1 is CheXbert F1 over the same 14 observations extracted from the final report. Thus the PDD-report row of Table 1 does not independently test the posterior decision: the report is instructed to contain h*, so its extracted labels are, up to LLM compliance and CheXbert noise, equal to h* by construction. The PDD label-level row is a genuine test-set prediction of the tuned decision module, so the circularity is partial and confined to the report-level claim.

full rationale

The central Bayesian aggregation (Eqs. 6-12) is a standard trained classifier: likelihood ratios are estimated on the training split and thresholds are tuned on the validation split, then applied to the test set. The label-level PDD results therefore have independent content and are not circular. The main circularity concern is the report-refinement step: Eq. (15) generates the final text from the same label vector that the evaluation metric extracts, so the PDD-report clinical-efficacy numbers are a textual realization of the tuned labels rather than an independent measurement. A control using the base report's own CheXbert labels in the same LLM revision would be needed to attribute the report-level gain to the posterior decision. No load-bearing self-citation or imported uniqueness theorem is present; the self-citations in the introduction are contextual. The conditional-independence assumption in Eq. (9) is a correctness risk, not a circularity.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The central claim rests on a small set of fitted thresholds and likelihood ratios, plus domain assumptions about label reliability and independence of evidence. No new medical entities are postulated.

free parameters (3)
  • alpha (certain predictions) = 1 or 1.2
    Hand-set amplification factor for positive/negative predictions in Eq. 9; chosen heuristically to prevent overconfidence.
  • alpha_t,o for uncertain predictions = Ratio of conditional probabilities estimated on training set
    Computed from training data via Eq. 12; used to weight uncertain evidence.
  • th_drop_o and th_add_o for 14 observations = Grid-searched on validation set (0 to 1 in steps of 0.05)
    Per-observation thresholds that maximize F1 on validation; these directly determine the final diagnosis.
assumptions (3)
  • domain assumption Conditional independence of hypotheses given the true observation label
    Eq. 9 multiplies likelihood ratios across hypotheses as if they are independent, but input types share overlapping images (e.g., CF1 alone vs CF1+CL), so errors are correlated.
  • domain assumption CheXbert labels of reference reports are reliable ground truth
    LRs and priors are estimated against CheXbert-extracted labels of reference reports; if CheXbert is biased, the whole calibration is biased.
  • ad hoc to paper 'Not mentioned' labels are treated as negative findings
    In LR estimation, 'null' is mapped to negative, which conflates silence with a genuine negative and influences the likelihood ratios.

how reviews work

0 comments
Cite this review

Pith. "Pith review of PDD-RRG: Posterior Diagnostic Decision for Study-level Radiology Report Generation." pith.science (2026). https://pith.science/paper/55K3PVZF

@misc{pith2026260803055,
  author       = {Pith},
  title        = {Pith review of: PDD-RRG: Posterior Diagnostic Decision for Study-level Radiology Report Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/55K3PVZF}},
  note         = {Machine review of arXiv:2608.03055}
}
read the original abstract

Automatic radiology report generation (RRG) aims to simulate the workflow of radiologists, assisting them in clinical diagnosis. However, existing methods often fall short in utilizing all information relevant to the examination, as is typically done in clinical practice. Although some works attempt to incorporate multi-view images and historical data, these additional inputs may sometimes lead to avoidable diagnostic errors on the contrary. To address these challenges, we introduce a decision-making stage after report generation for the first time and propose a Posterior Diagnostic Decision framework (PDD-RRG) to integrate potentially conflicting diagnoses. Specifically, we create various subsets of input data and utilize an existing RRG model to generate reports from different perspectives. Then the Bayesian posterior probability and the learned thresholds for each clinical observation are calculated to obtain an aggregated diagnostic conclusion, which is subsequently used to refine the generated report. Experiments on MIMIC-CXR demonstrate that our proposed PDD-RRG can effectively enhance the clinical efficacy of existing RRG models without any retraining.

Figures

Figures reproduced from arXiv: 2608.03055 by the authors.

Figure 1
Figure 1. Pipelines of RRG models and our proposed framework. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. A real case on MAIRA-2 where additional historical in [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 4
Figure 4. The architecture of our proposed pipeline. Note that [PITH_FULL_IMAGE:figures/full_fig_p003_4.png] view at source ↗
Figures from the paper (1 more)
Figure 6
Figure 6. Figure 6: Error repair statistics for FN (a) and FP (b). The numbers [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

69 extracted references · 40 canonical work pages

  1. [1]

    A Comprehensive Graph Framework for Question Answering with Mode-Seeking Preference Alignment , booktitle =

    Quanwei Tang and Sophia Yat Mei Lee and Junshuang Wu and Dong Zhang and Shoushan Li and Erik Cambria and Guodong Zhou , editor =. A Comprehensive Graph Framework for Question Answering with Mode-Seeking Preference Alignment , booktitle =. 2025 , url =

    work page 2025

  2. [2]

    2025 , url =

    Xincheng Ju and Dong Zhang and Junhui Li and Shoushan Li and Guodong Zhou , title =. 2025 , url =. doi:10.1109/TMM.2025.3543080 , timestamp =

    work page arXiv 2025

  3. [3]

    Emotion across Modalities and Cultures: Multilingual Multimodal Emotion-Cause Analysis with Memory-inspired Framework , booktitle =

    Dan Wu and Xincheng Ju and Dong Zhang and Shoushan Li and Erik Cambria and Guodong Zhou , editor =. Emotion across Modalities and Cultures: Multilingual Multimodal Emotion-Cause Analysis with Memory-inspired Framework , booktitle =. 2025 , url =. doi:10.1145/3746027.3755655 , timestamp =

    arXiv 2025

  4. [4]

    Two Heads are Better than One: Zero-shot Cognitive Reasoning via Multi-LLM Knowledge Fusion , booktitle =

    Liang Liu and Dong Zhang and Shoushan Li and Guodong Zhou and Erik Cambria , editor =. Two Heads are Better than One: Zero-shot Cognitive Reasoning via Multi-LLM Knowledge Fusion , booktitle =. 2024 , url =. doi:10.1145/3627673.3679744 , timestamp =

    arXiv 2024

  5. [5]

    Informatics in Medicine Unlocked , volume=

    Longitudinal data and a semantic similarity reward for chest X-ray report generation , author=. Informatics in Medicine Unlocked , volume=. 2024 , publisher=

    work page 2024

  6. [6]

    Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers),

    Wenjun Hou and Yi Cheng and Kaishuai Xu and Heng Li and Yan Hu and Wenjie Li and Jiang Liu , editor =. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers),. 2025 , url =

    work page 2025

  7. [7]

    The Impact of Auxiliary Patient Data on Automated Chest X-Ray Report Generation and How to Incorporate It , booktitle =

    Aaron Nicolson and Shengyao Zhuang and Jason Dowling and Bevan Koopman , editor =. The Impact of Auxiliary Patient Data on Automated Chest X-Ray Report Generation and How to Incorporate It , booktitle =. 2025 , url =

    work page 2025

  8. [8]

    2025 , url =

    Kang Liu and Zhuoqi Ma and Xiaolu Kang and Yunan Li and Kun Xie and Zhicheng Jiao and Qiguang Miao , title =. 2025 , url =. doi:10.1109/CVPR52734.2025.00968 , timestamp =

    arXiv 2025

Show all 69 references
  1. [9]

    AAAI-25, Sponsored by the Association for the Advancement of Artificial Intelligence, February 25 - March 4, 2025, Philadelphia, PA,

    Zhuhao Wang and Yihua Sun and Zihan Li and Xuan Yang and Fang Chen and Hongen Liao , editor =. AAAI-25, Sponsored by the Association for the Advancement of Artificial Intelligence, February 25 - March 4, 2025, Philadelphia, PA,. 2025 , url =. doi:10.1609/AAAI.V39I8.32890 , timestamp =

    2025 doi

  2. [10]

    Is classification all you need for radiology report generation? , author=

  3. [11]

    The Thirteenth International Conference on Learning Representations,

    Peng Xia and Kangyu Zhu and Haoran Li and Tianze Wang and Weijia Shi and Sheng Wang and Linjun Zhang and James Zou and Huaxiu Yao , title =. The Thirteenth International Conference on Learning Representations,. 2025 , url =

    2025

  4. [12]

    The Thirteenth International Conference on Learning Representations,

    Mohammad Shahab Sepehri and Zalan Fabian and Maryam Soltanolkotabi and Mahdi Soltanolkotabi , title =. The Thirteenth International Conference on Learning Representations,

  5. [13]

    Alistair E. W. Johnson and Tom J. Pollard and Seth J. Berkowitz and Nathaniel R. Greenbaum and Matthew P. Lungren and Chih. CoRR , volume =. 2019 , url =. 1901.07042 , timestamp =

    2019 arXiv

  6. [14]

    Ng and Matthew P

    Akshay Smit and Saahil Jain and Pranav Rajpurkar and Anuj Pareek and Andrew Y. Ng and Matthew P. Lungren , title =. Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing,. 2020 , url =. doi:10.18653/V1/2020.EMNLP-MAIN.117 , timestamp =

    2020 doi

  7. [15]

    Castro and Anton Schwaighofer and Sam Bond

    Shruthi Bannur and Kenza Bouzid and Daniel C. Castro and Anton Schwaighofer and Sam Bond. CoRR , volume =. 2024 , url =. doi:10.48550/ARXIV.2406.04449 , eprinttype =. 2406.04449 , timestamp =

    2024 doi arXiv

  8. [16]

    Xing , editor =

    Baoyu Jing and Pengtao Xie and Eric P. Xing , editor =. On the Automatic Generation of Medical Imaging Reports , booktitle =. 2018 , url =. doi:10.18653/V1/P18-1240 , timestamp =

    2018 doi

  9. [17]

    Xing , editor =

    Baoyu Jing and Zeya Wang and Eric P. Xing , editor =. Show, Describe and Conclude: On Exploiting the Structure Information of Chest X-ray Reports , booktitle =. 2019 , url =. doi:10.18653/V1/P19-1657 , timestamp =

    2019 doi

  10. [18]

    2021 , url =

    Fenglin Liu and Xian Wu and Shen Ge and Wei Fan and Yuexian Zou , title =. 2021 , url =. doi:10.1109/CVPR46437.2021.01354 , timestamp =

    2021

  11. [19]

    Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers),

    Wenjun Hou and Kaishuai Xu and Yi Cheng and Wenjie Li and Jiang Liu , title =. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers),. 2023 , url =. doi:10.18653/V1/2023.ACL-LONG.451 , timestamp =

    2023 doi

  12. [20]

    2024 , url =

    Shenshen Bu and Taiji Li and Yuedong Yang and Zhiming Dai , title =. 2024 , url =. doi:10.1109/CVPR52733.2024.01346 , timestamp =

    2024

  13. [21]

    AAAI-25, Sponsored by the Association for the Advancement of Artificial Intelligence, February 25 - March 4, 2025, Philadelphia, PA,

    Ting Xiao and Lei Shi and Peng Liu and Zhe Wang and Chenjia Bai , title =. AAAI-25, Sponsored by the Association for the Advancement of Artificial Intelligence, February 25 - March 4, 2025, Philadelphia, PA,. 2025 , url =. doi:10.1609/AAAI.V39I8.32936 , timestamp =

    2025 doi

  14. [22]

    Xing , title =

    Yuan Li and Xiaodan Liang and Zhiting Hu and Eric P. Xing , title =. Advances in Neural Information Processing Systems 31: Annual Conference on Neural Information Processing Systems 2018, NeurIPS 2018, December 3-8, 2018, Montr. 2018 , url =

    2018

  15. [23]

    Medical Image Computing and Computer Assisted Intervention -

    Jianbo Yuan and Haofu Liao and Rui Luo and Jiebo Luo , title =. Medical Image Computing and Computer Assisted Intervention -. 2019 , url =. doi:10.1007/978-3-030-32226-7\_80 , timestamp =

    2019 doi

  16. [24]

    Generating Radiology Reports via Memory-driven Transformer , booktitle =

    Zhihong Chen and Yan Song and Tsung. Generating Radiology Reports via Memory-driven Transformer , booktitle =. 2020 , url =. doi:10.18653/V1/2020.EMNLP-MAIN.112 , timestamp =

    2020 doi

  17. [25]

    Langlotz and Dan Jurafsky , title =

    Yasuhide Miura and Yuhao Zhang and Emily Bao Tsai and Curtis P. Langlotz and Dan Jurafsky , title =. Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies,. 2021 , url =. doi:10.18653/V1/2...

    2021 doi

  18. [26]

    Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing,

    Zhihong Chen and Yaling Shen and Yan Song and Xiang Wan , title =. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing,. 2021 , url =. doi:10.18653/V1/2021.ACL-LONG.4...

    2021 doi

  19. [27]

    Proceedings of the 29th International Conference on Computational Linguistics,

    Jingyi You and Dongyuan Li and Manabu Okumura and Kenji Suzuki , title =. Proceedings of the 29th International Conference on Computational Linguistics,. 2022 , url =

    2022

  20. [28]

    Findings of the Association for Computational Linguistics:

    Han Qin and Yan Song , title =. Findings of the Association for Computational Linguistics:. 2022 , url =. doi:10.18653/V1/2022.FINDINGS-ACL.38 , timestamp =

    2022 doi

  21. [29]

    2023 , url =

    Qingqiu Li and Jilan Xu and Runtian Yuan and Mohan Chen and Yuejie Zhang and Rui Feng and Xiaobo Zhang and Shang Gao , title =. 2023 , url =. doi:10.1109/BIBM58861.2023.10385817 , timestamp =

    2023

  22. [30]

    Machine Learning for Health, ML4H@NeurIPS 2023, 10 December 2023, New Orleans, Louisiana,

    Dang Nguyen and Chacha Chen and He He and Chenhao Tan , title =. Machine Learning for Health, ML4H@NeurIPS 2023, 10 December 2023, New Orleans, Louisiana,. 2023 , url =

    2023

  23. [31]

    Bai and Qiguang Miao , title =

    Kang Liu and Zhuoqi Ma and Xiaolu Kang and Zhusi Zhong and Zhicheng Jiao and Grayson Baird and Harrison X. Bai and Qiguang Miao , title =. Medical Image Computing and Computer Assisted Intervention -. 2024 , url =. doi:10.1007/978-3-031-72384-1\_41 , timestamp =

    2024 doi

  24. [32]

    CoRR , volume =

    Kang Liu and Zhuoqi Ma and Kun Xie and Zhicheng Jiao and Qiguang Miao , title =. CoRR , volume =. 2024 , url =. doi:10.48550/ARXIV.2411.10224 , eprinttype =. 2411.10224 , timestamp =

    2024 doi arXiv

  25. [33]

    Summers and Zhiyong Lu , title =

    Qingqing Zhu and Tejas Sudharshan Mathai and Pritam Mukherjee and Yifan Peng and Ronald M. Summers and Zhiyong Lu , title =. Medical Image Computing and Computer Assisted Intervention -. 2023 , url =. doi:10.1007/978-3-031-43904-9\_19 , timestamp =

    2023 doi

  26. [34]

    Findings of the Association for Computational Linguistics:

    Francesco Dalla Serra and Chaoyang Wang and Fani Deligianni and Jeff Dalton and Alison O'Neil , title =. Findings of the Association for Computational Linguistics:. 2023 , url =. doi:10.18653/V1/2023.FINDINGS-EMNLP.325 , timestamp =

    2023 doi

  27. [35]

    Findings of the Association for Computational Linguistics:

    Wenjun Hou and Yi Cheng and Kaishuai Xu and Wenjie Li and Jiang Liu , title =. Findings of the Association for Computational Linguistics:. 2023 , url =. doi:10.18653/V1/2023.FINDINGS-EMNLP.140 , timestamp =

    2023 doi

  28. [36]

    Computer Vision -

    Fuying Wang and Shenghui Du and Lequan Yu , title =. Computer Vision -. 2024 , url =. doi:10.1007/978-3-031-73001-6\_11 , timestamp =

    2024 doi

  29. [37]

    Proceedings of the 32nd

    Xin Mei and Rui Mao and Xiaoyan Cai and Libin Yang and Erik Cambria , title =. Proceedings of the 32nd. 2024 , url =. doi:10.1145/3664647.3681377 , timestamp =

    2024

  30. [38]

    AAAI-25, Sponsored by the Association for the Advancement of Artificial Intelligence, February 25 - March 4, 2025, Philadelphia, PA,

    Tengfei Liu and Jiapu Wang and Yongli Hu and Mingjie Li and Junfei Yi and Xiaojun Chang and Junbin Gao and Baocai Yin , title =. AAAI-25, Sponsored by the Association for the Advancement of Artificial Intelligence, February 25 - March 4, 2025, Philadelphia, PA,. 2025 , url =. ...

    2025 doi

  31. [39]

    Applications of Medical Artificial Intelligence - Third International Workshop,

    Yanfeng Li and Jinghan Sun and Liansheng Wang , title =. Applications of Medical Artificial Intelligence - Third International Workshop,. 2024 , url =. doi:10.1007/978-3-031-82007-6\_24 , timestamp =

    2024 doi

  32. [40]

    2024 , url =

    Ke Zhang and Hanliang Jiang and Jian Zhang and Qingming Huang and Jianping Fan and Jun Yu and Weidong Han , title =. 2024 , url =. doi:10.1109/TMM.2023.3273390 , timestamp =

    2024

  33. [41]

    Aman Madaan and Niket Tandon and Prakhar Gupta and Skyler Hallinan and Luyu Gao and Sarah Wiegreffe and Uri Alon and Nouha Dziri and Shrimai Prabhumoye and Yiming Yang and Shashank Gupta and Bodhisattwa Prasad Majumder and Katherine Hermann and Sean Welleck and Amir Yazdanbakh...

    2023

  34. [42]

    Noah Shinn and Federico Cassano and Ashwin Gopinath and Karthik Narasimhan and Shunyu Yao , title =. Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 202...

    2023

  35. [43]

    The Twelfth International Conference on Learning Representations,

    Zhibin Gou and Zhihong Shao and Yeyun Gong and Yelong Shen and Yujiu Yang and Nan Duan and Weizhu Chen , title =. The Twelfth International Conference on Learning Representations,. 2024 , url =

    2024

  36. [44]

    Goodman , title =

    Eric Zelikman and Yuhuai Wu and Jesse Mu and Noah D. Goodman , title =. Advances in Neural Information Processing Systems 35: Annual Conference on Neural Information Processing Systems 2022, NeurIPS 2022, New Orleans, LA, USA, November 28 - December 9, 2022 , year =

    2022

  37. [45]

    Learning From Mistakes Makes

    Shengnan An and Zexiong Ma and Zeqi Lin and Nanning Zheng and Jian. Learning From Mistakes Makes. CoRR , volume =. 2023 , url =. doi:10.48550/ARXIV.2310.20689 , eprinttype =. 2310.20689 , timestamp =

    2023 doi arXiv

  38. [46]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers),

    Yongqi Tong and Dawei Li and Sizhe Wang and Yujia Wang and Fei Teng and Jingbo Shang , title =. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers),. 2024 , url =. doi:10.18653/V1/2024.ACL-LONG.169 , timestamp =

    2024 doi

  39. [47]

    Vicky Zhao and Conghui He and Lijun Wu , title =

    Zhuoshi Pan and Yu Li and Honglin Lin and Qizhi Pei and Zinan Tang and Wei Wu and Chenlin Ming and H. Vicky Zhao and Conghui He and Lijun Wu , title =. Findings of the Association for Computational Linguistics,. 2025 , url =

    2025

  40. [48]

    Proceedings of the 32nd

    Xiao Liang and Yanlei Zhang and Di Wang and Haodi Zhong and Ronghan Li and Quan Wang , title =. Proceedings of the 32nd. 2024 , url =. doi:10.1145/3664647.3681201 , timestamp =

    2024

  41. [49]

    Findings of the Association for Computational Linguistics:

    Wenjun Hou and Yi Cheng and Kaishuai Xu and Yan Hu and Wenjie Li and Jiang Liu , title =. Findings of the Association for Computational Linguistics:. 2024 , url =. doi:10.18653/V1/2024.FINDINGS-EMNLP.528 , timestamp =

    2024 doi

  42. [50]

    Proceedings of the 31st International Conference on Computational Linguistics,

    Heng Yin and Shanlin Zhou and Pandong Wang and Zirui Wu and Yongtao Hao , title =. Proceedings of the 31st International Conference on Computational Linguistics,. 2025 , url =

    2025

  43. [51]

    AAAI-25, Sponsored by the Association for the Advancement of Artificial Intelligence, February 25 - March 4, 2025, Philadelphia, PA,

    Xiaofei Huang and Wenting Chen and Jie Liu and Qisheng Lu and Xiaoling Luo and Linlin Shen , title =. AAAI-25, Sponsored by the Association for the Advancement of Artificial Intelligence, February 25 - March 4, 2025, Philadelphia, PA,. 2025 , url =. doi:10.1609/AAAI.V39I4.3239...

    2025 doi

  44. [52]

    Findings of the Association for Computational Linguistics:

    Chentao Huang and Guangli Li and Xinjiong Zhou and Yafeng Ren and Hongbin Zhang , title =. Findings of the Association for Computational Linguistics:. 2025 , url =. doi:10.18653/V1/2025.FINDINGS-NAACL.215 , timestamp =

    2025 doi

  45. [53]

    arXiv preprint arXiv:1308.0850 , year=

    Generating sequences with recurrent neural networks , author=. arXiv preprint arXiv:1308.0850 , year=

    arXiv

  46. [54]

    arXiv preprint arXiv:2403.08002 , year=

    Towards a clinically accessible radiology foundation model: open-access and lightweight, with automated evaluation , author=. arXiv preprint arXiv:2403.08002 , year=

    arXiv

  47. [55]

    LLaVA-Rad MIMIC-CXR Annotations , author=

  48. [56]

    arXiv preprint arXiv:2401.10815 , year=

    Rad-dino: Exploring scalable medical image encoders beyond text supervision , author=. arXiv preprint arXiv:2401.10815 , year=

    arXiv

  49. [57]

    See https://vicuna

    Vicuna: An open-source chatbot impressing gpt-4 with 90\ author=. See https://vicuna. lmsys. org (accessed 14 April 2023) , volume=

    2023

  50. [58]

    arXiv preprint arXiv:2109.11635 , year=

    Revisiting the uniform information density hypothesis , author=. arXiv preprint arXiv:2109.11635 , year=

    arXiv

  51. [59]

    Cognitive psychology , volume=

    Redundancy and reduction: Speakers manage syntactic information density , author=. Cognitive psychology , volume=. 2010 , publisher=

    2010

  52. [60]

    Transactions of the Association for Computational Linguistics , volume=

    A cross-linguistic pressure for Uniform Information Density in word order , author=. Transactions of the Association for Computational Linguistics , volume=. 2023 , publisher=

    2023

  53. [61]

    Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing , year=

    Is Information Density Uniform in Task-Oriented Dialogues? , author=. Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing , year=

    2021

  54. [62]

    arXiv preprint arXiv:2501.03895 , year=

    Llava-mini: Efficient image and video large multimodal models with one vision token , author=. arXiv preprint arXiv:2501.03895 , year=

    arXiv

  55. [63]

    Preparing a collection of radiology examinations for distribution and retrieval , journal =

    Dina Demner. Preparing a collection of radiology examinations for distribution and retrieval , journal =. 2016 , url =. doi:10.1093/JAMIA/OCV080 , timestamp =

    2016 doi

  56. [64]

    Aaron Nicolson and Jason Dowling and Bevan Koopman , title =. Artif. Intell. Medicine , volume =. 2023 , url =. doi:10.1016/J.ARTMED.2023.102633 , timestamp =

    2023

  57. [65]

    Kevin Zhou and Li Xiao , title =

    Shuxin Yang and Xian Wu and Shen Ge and Zhuozhao Zheng and S. Kevin Zhou and Li Xiao , title =. Medical Image Anal. , volume =. 2023 , url =. doi:10.1016/J.MEDIA.2023.102798 , timestamp =

    2023

  58. [67]

    Interactive and Explainable Region-guided Radiology Report Generation , booktitle =

    Tim Tanida and Philip M. Interactive and Explainable Region-guided Radiology Report Generation , booktitle =. 2023 , url =. doi:10.1109/CVPR52729.2023.00718 , timestamp =

    2023

  59. [68]

    PromptMRG: Diagnosis-Driven Prompts for Medical Report Generation , booktitle =

    Haibo Jin and Haoxuan Che and Yi Lin and Hao Chen , editor =. PromptMRG: Diagnosis-Driven Prompts for Medical Report Generation , booktitle =. 2024 , url =. doi:10.1609/AAAI.V38I3.28038 , timestamp =

    2024 doi

  60. [69]

    Behl and Alon Benhaim and Misha Bilenko and Johan Bjorck and S

    Marah I Abdin and Sam Ade Jacobs and Ammar Ahmad Awan and Jyoti Aneja and Ahmed Awadallah and Hany Awadalla and Nguyen Bach and Amit Bahree and Arash Bakhtiari and Harkirat S. Behl and Alon Benhaim and Misha Bilenko and Johan Bjorck and S. Phi-3 Technical Report:. CoRR , volum...

    2024 doi arXiv

  61. [70]

    Phi-4-Mini Technical Report: Compact yet Powerful Multimodal Language Models via Mixture-of-LoRAs , journal =

    Abdelrahman Abouelenin and Atabak Ashfaq and Adam Atkinson and Hany Awadalla and Nguyen Bach and Jianmin Bao and Alon Benhaim and Martin Cai and Vishrav Chaudhary and Congcong Chen and Dong Chen and Dongdong Chen and Jun. Phi-4-Mini Technical Report: Compact yet Powerful Multi...

    2025 doi arXiv

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.