The reviewed record of science sign in
Pith

arxiv: 2607.05880 · v1 · pith:KLNVJOK2 · submitted 2026-07-07 · cs.CV · cs.AI

Harrison.Rad 1.5 Technical Report: A radiology foundation model that can draft reports from images, priors and clinical context

Suneeta Mall , Vladimir Nekrasov , Ashnil Kumar , Sajith Karunasena , Aiden Nibali , Alix Bird , Mateo Diaz Shine , Jarrel Seah This is my paper

Reviewed by Pith T0 review T1 audit T2 compute T3 formal T4 kernel 2026-07-08 21:43 UTCglm-5.2pith:KLNVJOK2record.jsonopen to challenge →

classification cs.CV cs.AI
keywords clinicalmodelradiologyreportingreportsacrosschestframework
0
0 comments X

The pith

Radiology AI passes mock FRCR exam where all rivals fail

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

This paper presents Harrison.Rad 1.5 (HR1.5), a radiology-specific multimodal large language model that interprets X-ray images, integrates clinical history and prior studies, and drafts structured radiology reports across the full breadth of plain-film radiology. The authors train the model through a three-stage pipeline — domain adaptation of a base language model on radiology reports, contrastive vision-encoder training with curriculum-based hard negatives on approximately six million image-report instances, and visual-question-answering fine-tuning on multi-turn clinical conversations. The central claim is that HR1.5 is the only system evaluated — among general-purpose frontier models and medical-domain models alike — to meet the passing standard on a simulated Fellowship of the Royal College of Radiologists (FRCR) 2B Short Case examination, achieving a 62.5% per-sheet pass rate while every other model scored 0%. The authors also introduce a Findings-Diagnosis scoring framework that extends RadGraph-XL entity extraction with ontology-based synonym matching and polarity-contradiction detection, arguing that conventional text-overlap metrics (BLEU, ROUGE, BERTScore) fail to distinguish clinically correct reports from clinically wrong ones. The paper further presents an explainability engine that probes whether the model genuinely encodes and uses radiological concepts (via activation steering), localises answers to image regions (via Grad-CAM and attention methods), and estimates confidence (via attention entropy, head disagreement, and representation-trajectory instability, corrected for attention-sink artefacts).

Core claim

The core discovery is that a radiology-specific foundation model, trained through staged domain adaptation, contrastive alignment, and instruction tuning on approximately 6.5 million studies, can pass a simulated version of a standardised radiologist certification examination (FRCR 2B Short Case) while all general-purpose and medical-domain models evaluated fail every exam sheet. The authors attribute this to radiology-specific training that spends model capacity on the distinctions that matter for image interpretation and reporting rather than on general-purpose breadth. A secondary contribution is the Findings-Diagnosis scoring methodology, which demonstrates that conventional report-level

What carries the argument

HR1.5 is a multimodal large language model built through a three-stage pipeline: (1) domain adaptation of a base LLM on radiology reports, (2) contrastive vision-encoder training with curriculum-based hard negatives on ~6 million image-report instances, and (3) visual-question-answering fine-tuning on ~18 million crafted multi-turn clinical conversations. The HR1.5+ configuration adds an agentic framework that consults task-appropriate models for higher-level tasks like report generation. The Findings-Diagnosis scoring framework extends RadGraph-XL entity extraction with ontology-based synonym matching and polarity-contradiction detection. The explainability engine uses activation steering (

If this is right

  • If a radiology-specific model can pass a simulated certification exam, draft reports across multiple body parts, and integrate clinical priors, it could substantially reduce the time radiologists spend producing reports — the bottleneck driving reporting backlogs.
  • The Findings-Diagnosis scoring methodology, if adopted by the community, could shift how radiology report generation is evaluated away from lexical-overlap metrics toward clinical-correctness metrics that penalise polarity contradictions.
  • The activation-steering analysis showing that concepts like pulmonary oedema, cardiomegaly, and fracture are bidirectionally controllable suggests that internal model representations could be used for monitoring and controlled intervention, not just passive observation.
  • The confidence estimation approach — combining attention entropy, head disagreement, and representation-trajectory instability, corrected for attention sinks — offers a path toward giving clinicians a trust signal alongside each AI-generated answer, though the authors note it is not yet calibrated.
  • The proposed C6 evaluation framework (Coverage, Correctness, Comprehensiveness, Care, Coherence, Conformity) would decompose report quality into interpretable axes, separating failure modes like over-calling from under-describing.

Where Pith is reading between the lines

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

  • The fact that HR1.5+ regresses on the legacy FRCR Rapids format (15.7% vs HR1.5's 24.3%) because it tends to produce detailed responses penalised on normal cases suggests a fundamental tension: models optimised for thorough reporting may underperform on triage-style tasks that reward brevity, implying different clinical workflows may require different model configurations.
  • The observation that the model maps an out-of-distribution input (a synthetic X-ray of an apple in a chest) onto the nearest learnt concept (breast prosthesis) with high confidence reveals that the confidence signal cannot detect conceptual novelty — only distributional unfamiliarity — which is a safety-critical distinction for clinical deployment.
  • If the FRCR simulation's case selection, difficulty distribution, or scoring rubric diverges from the actual RCR exam, the headline passing claim may not generalise to the real certification setting — a question that could only be resolved by evaluation on official exam materials.
  • The concept geometry analysis showing that acute fracture directions are anti-aligned with chronic degenerative directions (cosine similarity −0.76) raises the question of whether the model's internal representation space could be used to detect novel findings: a finding that does not align with any existing concept direction might flag a genuinely novel pathology.

Load-bearing premise

The simulated FRCR examination, the Angoff cut-off thresholds, and the internal evaluation datasets (RadBench, RadCoverage-VQA) were all constructed by the model's developers and have not been independently validated against the actual RCR examination or externally reviewed. If the case selection, difficulty distribution, or scoring rubric of this simulation diverges from the real FRCR exam, the headline passing claim may not generalise.

What would settle it

If HR1.5 were evaluated on official RCR FRCR 2B Short Case examination materials (or an independently constructed simulation with verified difficulty distribution and scoring), and its per-sheet pass rate fell below the Angoff cut-off, the central claim would be undermined. Similarly, if an independent multi-reader study found that radiologists rejected HR1.5's draft reports at rates comparable to the general-purpose models it outperforms on internal benchmarks, the claim of clinically superior reporting would weaken.

Figures

Figures reproduced from arXiv: 2607.05880 by Aiden Nibali, Alix Bird, Ashnil Kumar, Jarrel Seah, Mateo Diaz Shine, Sajith Karunasena, Suneeta Mall, Vladimir Nekrasov.

Figure 1.1
Figure 1.1. Figure 1.1: The radiology foundation-model landscape, 2020–2026. Models are positioned by year of release, with the bars showing the total released per year; marker colour denotes access level (open source, open weights, research/closed, API/preview, FDA-cleared/commercial) and marker shape denotes capability tier, from precursor segmentation and backbone encoders (1), through contrastive or single-modality vision-l… view at source ↗
Figure 1
Figure 1. Figure 1: summarises this pipeline at a conceptual level: the progression from a domain-adapted [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 1.2
Figure 1.2. Figure 1.2: Conceptual overview of the HR1.5 training pipeline. From radiology data (images, reports, and clinical context), training proceeds through multi-stage pre-training (with and without contrastive objectives) followed by instruction fine-tuning for image-grounded visual question answering, yielding Harrison.Rad 1.5. Objectives are kept predominantly radiology-relevant; internal architecture, stage count, an… view at source ↗
Figure 1.3
Figure 1.3. Figure 1.3: Data engineering for HR1.5 training, shown conceptually. From a de-identified clinical source, two engineered learning signals are produced: (top) hard negatives that vary along clinically meaningful axes (breadth) under a difficulty curriculum (depth), for contrastive alignment; (bottom) context-grounded crafted dialogues derived from real cases, with deliberate oversampling of under-represented finding… view at source ↗
Figure 1
Figure 1. Figure 1: illustrates this philosophy schematically: a single de-identified clinical source is [PITH_FULL_IMAGE:figures/full_fig_p008_1.png] view at source ↗
Figure 3.1
Figure 3.1. Figure 3.1: Scaphoid fracture. Case courtesy of Subash Thapa, Radiopaedia.org (case 40623), used under licence. 22 [PITH_FULL_IMAGE:figures/full_fig_p022_3_1.png] view at source ↗
Figure 3.2
Figure 3.2. Figure 3.2: 1. Metastatic lung disease, portable chest radiograph. Case courtesy of Daniel Hyeong Seok Kim, Radiopaedia.org (case 153711), used under licence. Image only (no clinical history). 23 [PITH_FULL_IMAGE:figures/full_fig_p023_3_2.png] view at source ↗
Figure 3.2
Figure 3.2. Figure 3.2: 2. Metastatic lung disease, portable chest radiograph. Case courtesy of Daniel Hyeong Seok Kim, Radiopaedia.org (case 153711), used under licence. Image with clinical history. 24 [PITH_FULL_IMAGE:figures/full_fig_p024_3_2.png] view at source ↗
Figure 3.3
Figure 3.3. Figure 3.3: Pneumococcal pneumonia, current (left) and prior (right) chest radiographs. Case courtesy of Jack Ren, Radiopaedia.org (case 29090), used under licence. 26 [PITH_FULL_IMAGE:figures/full_fig_p026_3_3.png] view at source ↗
Figure 3.4
Figure 3.4. Figure 3.4: Left upper lobe mass. Case courtesy of Dai Roberts, Radiopaedia.org (case 80445), used under licence. 27 [PITH_FULL_IMAGE:figures/full_fig_p027_3_4.png] view at source ↗
Figure 3.5
Figure 3.5. Figure 3.5: 1. Congestive cardiac failure, PA and lateral chest radiographs. Case courtesy of Frank Gaillard, Radiopaedia.org (case 29092), used under licence. 29 [PITH_FULL_IMAGE:figures/full_fig_p029_3_5.png] view at source ↗
Figure 3.5
Figure 3.5. Figure 3.5: 2. Patellar dislocation with medial avulsion fracture and joint effusion. Case courtesy of Rodney Strahan, Radiopaedia.org (case 51097), used under licence. 30 [PITH_FULL_IMAGE:figures/full_fig_p030_3_5.png] view at source ↗
Figure 3.6
Figure 3.6. Figure 3.6: Interstitial thickening. Case courtesy of Bruno Di Muzio, Radiopaedia.org (case 45314), used under licence. 32 [PITH_FULL_IMAGE:figures/full_fig_p032_3_6.png] view at source ↗
Figure 4.1
Figure 4.1. Figure 4.1: CXR concept cosine similarity. Where directions do align, the alignment tracks clinical relationships rather than arbitrary correlation. In CXR, the overlapping airspace and parenchymal processes group together: atelectasis and consolidation are the most aligned pair (0.77), with lung collapse (0.69) and ground glass opacity (0.67) nearby, while the support devices form their own cluster, the central ven… view at source ↗
Figure 4.2
Figure 4.2. Figure 4.2: MSK concept cosine similarity. • Subtractive: subtracting reliably suppresses the finding, but adding does not sustain its induction; dependable for detection and suppression, not for induction. • Additive: adding induces the finding, but subtracting does not suppress it; usable for eliciting or amplifying the finding, not for detection or suppression. • None: neither direction produces the expected effe… view at source ↗
Figure 4.3
Figure 4.3. Figure 4.3: FracAtlas localisation: ground-truth fracture regions (left of each pair) versus the model’s attended regions (right) across a sample of cases. MS-CXR examples cannot be shown here owing to dataset licensing restrictions, but the localisation behaviour is analogous and is deducible from the hit-rate results above. 40 [PITH_FULL_IMAGE:figures/full_fig_p040_4_3.png] view at source ↗
Figure 4.4
Figure 4.4. Figure 4.4: Left apical lung mass case. Case courtesy of Dai Roberts, Radiopaedia.org (case 80445), used under licence. The engine reports a raw confidence of 0.81. Detecting and correcting for attention sinks lowers this to 0.79, an inflation of 0.03. A small portion of the apparent certainty was a structural artefact rather than genuine engagement with the image and question. The modest gap leaves the overall trus… view at source ↗
Figure 4.5
Figure 4.5. Figure 4.5: Ankle Weber B fracture. Case courtesy of Frank Gaillard, Radiopaedia.org (case 7480), used under licence. Across the conversation the corrected confidence stays in a narrow band (0.80–0.82) with a consistent, small sink inflation (∼0.02), showing that the model’s certainty is stable as the dialogue moves from description to classification to a full report, and does not spuriously spike. The activation hi… view at source ↗
Figure 4.6
Figure 4.6. Figure 4.6: Querying with the historical first Röntgen’s 1895 X-ray of a hand. Underlying radiograph: Wikimedia Commons (Wilhelm Röntgen, X-ray of Albert von Kölliker’s hand, 1896). degraded image and the off-domain question: the confidence signal registers that this input sits outside the model’s familiar distribution, which is exactly the behaviour a trustworthy uncertainty estimate should show. While these result… view at source ↗
Figure 4.7
Figure 4.7. Figure 4.7: Query with a synthetically generated x-ray image. The activation overlay is concentrated on the apple-shaped opacity that the model labels a prosthesis, confirming that the confident misinterpretation is driven by the anomalous object itself, and that the model substitutes its closest learnt concept for one it was never trained to recognise. 45 [PITH_FULL_IMAGE:figures/full_fig_p045_4_7.png] view at source ↗
Figure 6.1
Figure 6.1. Figure 6.1: The C6 framework. A draft report is scored along six clinical pillars — Coverage, Correctness, Comprehensiveness, Care, Coherence, and Conformity — each yielding a pillar score pi . A weighted sum with deployment-tunable weights wi produces a single Credibility score, P i wipi [PITH_FULL_IMAGE:figures/full_fig_p047_6_1.png] view at source ↗
read the original abstract

Imaging demand is growing faster than the radiology workforce can expand, and reporting backlogs cannot be resolved through training and recruitment alone. The most direct opportunity is reducing the time and effort radiologists spend producing reports, a task that requires interpreting images, integrating clinical history and prior studies, and drafting structured findings. We present Harrison.Rad 1.5 (HR1.5), a radiology-specific multimodal large language model that accepts interleaved text and visual inputs and generates structured and unstructured text across plain-film radiology, spanning computed radiography, chest, musculoskeletal, abdominal, spine, and pelvic x-rays, and mammography. HR1.5 is trained through a three-stage pipeline: domain adaptation of a base language model on radiology reports, contrastive vision-encoder training with curriculum-based hard negatives on ~6 million image-report instances, and visual-question-answering fine-tuning on multi-turn conversations. We evaluate it with a Findings-Diagnosis scoring framework that extends RadGraph-XL entity extraction with ontology-based synonym matching and polarity-contradiction detection, benchmarked on RadBench, a simulated FRCR 2B Short Case examination scored against Angoff-method thresholds, ReXGradient, and internal multi-modality datasets. HR1.5 is the only system evaluated to meet the simulated FRCR passing standard and achieves the highest accuracy on closed-format clinical questions, across anatomical regions, on internal multi-body-part and mammography reporting, and on the primary clinically-aligned score for public chest reporting. We further examine explainability and model behaviour, including question-sensitive Grad-CAM heatmaps, attention analysis, and confidence estimation, to support responsible future evaluation toward clinical use, and a framework for clinically grounded assessment of report quality.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

6 major / 12 minor

Summary. This technical report presents Harrison.Rad 1.5 (HR1.5), a radiology-specific multimodal large language model that drafts reports from X-ray images, priors, and clinical context. The model is trained via a three-stage pipeline (domain adaptation, contrastive vision-encoder training with curriculum hard negatives, and VQA fine-tuning) on approximately 6.5 million studies. Evaluation spans a simulated FRCR 2B Short Case examination, closed- and open-format VQA benchmarks (RadBench, ReXGradient, RadCoverage-VQA, CBIS-DDSM mammography), and qualitative case comparisons against frontier general-purpose models. The central claim is that HR1.5 is the only system evaluated to meet the simulated FRCR passing standard and achieves the highest accuracy on closed-format clinical questions across anatomical regions. The report also introduces a Findings-Diagnosis scoring framework, an explainability engine (concept tracing, attention-based localisation, confidence estimation), and outlines a future C6 evaluation framework.

Significance. The paper tackles an important problem (radiology workload scaling) and presents a broad evaluation across multiple body parts and modalities, which is a genuine strength over the many chest-only systems in the literature. The Findings-Diagnosis scoring framework, with its polarity-contradiction detection and ontology-based synonym matching, is a reasonable attempt to address well-known limitations of lexical-overlap metrics, and Table 1 effectively illustrates these deficiencies. The concept-tracing analysis (§4.1) with causal steering tests is a commendable depth of mechanistic analysis for a technical report. The authors are also transparent about several limitations, including the RadBench sample-size issue and the regression of HR1.5+ on certain benchmarks. However, the headline FRCR claim rests on an internally constructed simulation with an LLM-as-a-judge scorer whose properties are underspecified, and the statistical evidence at the claimed passing threshold is fragile.

major comments (6)
  1. §2.2.1, FRCR 2B Short Case scoring: The headline claim that HR1.5+ is the only system to meet the FRCR passing standard (Table 4: 62.5% pass rate) depends on an LLM-as-a-judge scoring 400 cases on a 0–5 rubric. The paper does not specify which LLM serves as judge, the prompt or rubric anchor text, whether model identity is visible to the judge, or the inter-rater agreement between the LLM judge and qualified FRCR examiners. This matters because HR1.5+ is explicitly described as having 'a greater tendency to produce detailed, characterising responses' (§2.2.1), and LLM-as-a-judge methods are documented to exhibit length and verbosity bias (Gu et al., 2024, the very citation used). The paper should specify the judge model, prompt, and any blinding, and ideally report inter-rater agreement with human FRCR examiners on a subset of cases.
  2. §2.2.1, Tables 2–4 and the verbosity confound: The pattern of results is internally consistent with a judge that rewards response length. HR1.5+ passes more Short Case sheets (all abnormal, rewarded for detail) but regresses on legacy Rapids (half normal, penalised for over-calling: 15.7% vs HR1.5's 24.3%) and on RadCoverage-VQA (46.6% vs HR1.5's 66.6%). The paper attributes the Rapids regression to verbosity but does not test whether the Short Case advantage is driven by the same verbosity bias rather than clinical precision. A controlled analysis — e.g., length-normalised re-scoring, or human expert re-scoring of a sample where HR1.5+ passes and a shorter competitor fails — would substantially strengthen the central claim. Without this, the verbosity-confound alternative explanation for the headline result remains open.
  3. §2.2.1, Table 3: The statistical evidence for the FRCR 2B Short Case passing claim is fragile. HR1.5+ has a mean score of 74.8 (95% CI: 61.4–87.2) against a mean Angoff cut-off of 73.2; the lower bound of the CI is well below the cut-off. The per-sheet pass rate of 62.5% (10 of 16 sheets) yields a 95% binomial CI of approximately [35%, 85%]. The Angoff cut-offs themselves were 'estimated internally' and 'spot-checked and verified by internal expert radiologists' — the estimation procedure, number of radiologists, and their agreement should be reported. The paper should acknowledge the uncertainty in the headline claim more explicitly and temper the conclusion accordingly.
  4. §2.2.3, Table 9: RadCoverage-VQA is an internal held-out set whose composition cannot be independently verified. HR1.5+ scores substantially lower than HR1.5 (46.6% vs 66.6% overall), which the authors attribute to the multi-turn dialogue format. This is a large and unexplained regression for the agentic configuration that is presented as the superior system on the FRCR headline. The paper should clarify whether this regression reflects a genuine limitation of the agentic framework or an artefact of the scoring methodology, and should reconcile the narrative that HR1.5+ is the preferred configuration with this result. At minimum, the abstract and discussion should not present HR1.5+ as uniformly superior without noting this caveat.
  5. §2.1, Findings-Diagnosis scoring: The scoring framework is described at a conceptual level but key operational details are missing — the fuzzy token matching threshold, the ontology source, the generic descriptor filter list, and the normal/abnormal classification fallback logic are unspecified. The paper acknowledges the framework 'can still credit a report that misses a subtle finding or accept a vague but incorrect descriptor as a match' (§2.1), but does not quantify the false-positive or false-negative rate of the scorer against a human reference. Given that this scorer underpins RadCoverage-VQA, ReXGradient, and mammography results, a validation study (even on a small sample) comparing F-D verdicts to expert radiologist verdicts would strengthen the reliability of all downstream claims.
  6. §2.2.1, FRCR simulation validity: The FRCR examination simulations are internally constructed using external data, not the official RCR sittings. The paper states it made 'best effort to follow the published examination guidelines' but does not report the case selection criteria, difficulty distribution, or body-part distribution in detail, nor how closely the simulation matches the official exam in these respects. Since the headline claim is that HR1.5+ meets the 'simulated FRCR passing standard,' the paper should be more explicit that this is a simulation that has not been validated against official RCR sittings, and should avoid language that could be read as claiming official FRCR-level competence.
minor comments (12)
  1. §1.1: The training pipeline is described at a high conceptual level, with 'internal architecture, stage count, and scale intentionally omitted' (Fig. 1.2 caption). While the rationale is understandable, the complete absence of model size, training compute, or architecture details makes independent reproduction impossible and limits scholarly assessment. Consider providing at least parameter count and total training compute.
  2. §1.2: The hard-negative curriculum and crafted dialogue generation are described conceptually but no quantitative details (curriculum schedule, oversampling rates, generator models) are provided. Adding even approximate figures would strengthen the methodology section.
  3. §2.1, Table 1: The table is effective but the column headers could be clearer — 'GT acc.' and 'F-D' are abbreviated without a footnote. A brief caption note explaining each column would aid readability.
  4. §2.2.1: The Angoff cut-off estimation procedure is referenced but not described. A brief description of the standard-setting method and the number of expert radiologists involved should be specified.
  5. §2.2.2, Table 6: CheXOne and CheXagent appear only in the CXR column with dashes elsewhere. A footnote or table note explaining that these are CXR-only models would improve readability.
  6. §2.2.3, Table 8: The note about RexGradient's structured reports affecting the automated F-D score is important context but appears after the table. Consider flagging this limitation before presenting the results or highlighting it more prominently.
  7. §4.1, Tables 13–14: The steering results are interesting but the classification criteria (bidirectional vs. subtractive) are described in text but not in the table. Adding a brief note or colour-coding would aid interpretation.
  8. §4.3: The confidence feature combines three signals with weights (w_e, w_d, w_cv) but the weight values are not reported. Stating the current values or noting that they are untuned would aid interpretation.
  9. §4.4: The OOD example (synthetic apple X-ray) is qualitatively interesting but the claim that the error is 'reasonable' is subjective. Consider framing this more neutrally as a limitation to be addressed.
  10. §6.1: The C6 framework is presented as future work but is described at a level of detail that suggests it is already partially implemented. Clarifying its current status relative to the current evaluation would help readers calibrate expectations.
  11. The paper would benefit from a consolidated table of all benchmarks, models, and metrics, as results are currently spread across Tables 2–10 with varying model subsets and metric sets. A summary table or figure would help readers navigate the evaluation.
  12. Several references are to non-peer-reviewed sources (arXiv preprints, model documentation pages). Where peer-reviewed alternatives exist, they should be preferred. Ensure all URLs are properly cited and accessible.

Simulated Author's Rebuttal

6 responses · 0 unresolved

We thank the referee for a careful and constructive review. The referee's comments are well-taken on several points, particularly regarding the underspecification of the LLM-as-a-judge scorer, the verbosity confound, the fragility of the statistical evidence at the FRCR passing threshold, and the need for validation of the Findings-Diagnosis scorer. We agree to revise the manuscript to address these. On the RadCoverage-VQA regression and the FRCR simulation validity, we agree that the manuscript needs to be more transparent and temper its claims. We provide point-by-point responses below.

read point-by-point responses

  1. Referee: §2.2.1, FRCR 2B Short Case scoring: The headline claim depends on an LLM-as-a-judge scoring 400 cases on a 0–5 rubric. The paper does not specify which LLM serves as judge, the prompt or rubric anchor text, whether model identity is visible to the judge, or the inter-rater agreement between the LLM judge and qualified FRCR examiners. The paper should specify the judge model, prompt, and any blinding, and ideally report inter-rater agreement with human FRCR examiners on a subset of cases.

    Authors: The referee is correct that these details are missing from the current manuscript. We will add them in the revision. Specifically, we will specify the judge model identity, the full rubric anchor text, and confirm that model identity was not visible to the judge (the judge received only the case image, the reference, and the candidate response, with no model identifiers). Regarding inter-rater agreement with human FRCR examiners: we agree this would substantially strengthen the claim. We have internal expert radiologists who verified a subset of the LLM judge's scores during development, and we will report the agreement rate on that subset. If the agreement data are sufficient for quantitative reporting (e.g., Cohen's kappa or percent agreement), we will include them; if not, we will at minimum report the qualitative validation procedure and its findings. We acknowledge that without formal inter-rater agreement data, the LLM-as-a-judge scoring remains a limitation that should be stated explicitly. revision: yes

  2. Referee: §2.2.1, Tables 2–4 and the verbosity confound: The pattern of results is internally consistent with a judge that rewards response length. HR1.5+ passes more Short Case sheets but regresses on legacy Rapids and RadCoverage-VQA. The paper attributes the Rapids regression to verbosity but does not test whether the Short Case advantage is driven by the same verbosity bias. A controlled analysis — e.g., length-normalised re-scoring, or human expert re-scoring of a sample where HR1.5+ passes and a shorter competitor fails — would substantially strengthen the central claim.

    Authors: This is a fair and important concern. We agree that the verbosity-confound alternative explanation is not currently ruled out by the evidence in the manuscript. We will address this in two ways in the revision. First, we will conduct a length-normalised re-scoring analysis on the FRCR 2B Short Case sheets, re-running the LLM judge with an explicit instruction to score based on clinical correctness regardless of response length, and report whether the pass rates change materially. Second, we will arrange for expert radiologist re-scoring of a sample of cases where HR1.5+ passes and a shorter competitor fails, to test whether the advantage holds under human judgement. If the re-scoring shows that the advantage is materially reduced or eliminated, we will temper the headline claim accordingly. If it holds, we will report the human validation alongside the LLM judge results. We acknowledge that as the manuscript currently stands, the verbosity confound remains an open alternative explanation, and we will state this limitation explicitly until the controlled analysis is completed. revision: yes

  3. Referee: §2.2.1, Table 3: The statistical evidence for the FRCR 2B Short Case passing claim is fragile. HR1.5+ has a mean score of 74.8 (95% CI: 61.4–87.2) against a mean Angoff cut-off of 73.2; the lower bound of the CI is well below the cut-off. The per-sheet pass rate of 62.5% (10 of 16 sheets) yields a 95% binomial CI of approximately [35%, 85%]. The Angoff cut-offs themselves were 'estimated internally' and 'spot-checked and verified by internal expert radiologists' — the estimation procedure, number of radiologists, and their agreement should be reported. The paper should acknowledge the uncertainty in the headline claim more explicitly and temper the conclusion accordingly.

    Authors: The referee's statistical assessment is correct. The 95% CI for HR1.5+'s mean score does include values below the Angoff cut-off, and the binomial CI on the pass rate is wide. We will revise the manuscript to (1) report the Angoff estimation procedure in more detail, including the number of radiologists involved and their level of agreement, (2) include the binomial CI on the per-sheet pass rate, and (3) temper the headline claim to acknowledge the statistical uncertainty. Specifically, we will revise the abstract and discussion to state that HR1.5+ meets the simulated passing standard on the majority of sheets but that the confidence intervals are wide and the result should be interpreted with appropriate caution. We will avoid language implying definitive FRCR-level competence. revision: yes

  4. Referee: §2.2.3, Table 9: RadCoverage-VQA is an internal held-out set whose composition cannot be independently verified. HR1.5+ scores substantially lower than HR1.5 (46.6% vs 66.6% overall), which the authors attribute to the multi-turn dialogue format. This is a large and unexplained regression for the agentic configuration that is presented as the superior system on the FRCR headline. The paper should clarify whether this regression reflects a genuine limitation of the agentic framework or an artefact of the scoring methodology, and should reconcile the narrative that HR1.5+ is the preferred configuration with this result. At minimum, the abstract and discussion should not present HR1.5+ as uniformly superior without noting this caveat.

    Authors: We agree that the manuscript currently presents HR1.5+ as superior without adequately flagging this regression, and that this is a genuine inconsistency in the narrative. The regression on RadCoverage-VQA is substantial and not fully explained. We believe it reflects a genuine limitation of the agentic configuration in the multi-turn dialogue setting — the agentic framework introduces additional processing steps that can degrade performance on certain dialogue-heavy tasks — but we cannot fully rule out a scoring methodology artefact. In the revision, we will (1) add a caveat to the abstract and discussion noting that HR1.5+ does not uniformly improve over HR1.5 and regresses on multi-turn dialogue evaluation, (2) provide additional analysis of the regression, and (3) present HR1.5+ as the preferred configuration for specific task types (e.g., structured report generation, single-pass examination-style reporting) rather than uniformly superior. We acknowledge that the internal nature of RadCoverage-VQA limits independent verification, and we will note this as well. revision: yes

  5. Referee: §2.1, Findings-Diagnosis scoring: The scoring framework is described at a conceptual level but key operational details are missing — the fuzzy token matching threshold, the ontology source, the generic descriptor filter list, and the normal/abnormal classification fallback logic are unspecified. The paper acknowledges the framework 'can still credit a report that misses a subtle finding or accept a vague but incorrect descriptor as a match' but does not quantify the false-positive or false-negative rate of the scorer against a human reference. A validation study comparing F-D verdicts to expert radiologist verdicts would strengthen the reliability of all downstream claims.

    Authors: The referee is correct that operational details are missing and that a validation study against expert radiologist verdicts would strengthen the reliability of downstream claims. We will add the operational details in the revision: the fuzzy token matching threshold, the ontology source (a radiology knowledge graph derived from standard radiology lexicons), the generic descriptor filter list (at least in summary form), and the normal/abnormal fallback logic. Regarding validation: we have internal data from radiologist review of F-D verdicts on a subset of cases (some of which is reflected in the GT-acceptable column of Table 1), and we will report the false-positive and false-negative rates of the F-D scorer against this human reference. If the validation sample is large enough, we will report quantitative agreement metrics; if not, we will report what we have and acknowledge the limitation. We agree that without this validation, the reliability of downstream claims resting on the F-D scorer is not fully established. revision: partial

  6. Referee: §2.2.1, FRCR simulation validity: The FRCR examination simulations are internally constructed using external data, not the official RCR sittings. The paper states it made 'best effort to follow the published examination guidelines' but does not report the case selection criteria, difficulty distribution, or body-part distribution in detail, nor how closely the simulation matches the official exam. The paper should be more explicit that this is a simulation that has not been validated against official RCR sittings, and should avoid language that could be read as claiming official FRCR-level competence.

    Authors: We agree. The manuscript already states that the simulations are internally constructed, but the language in the abstract and discussion could be read as implying a stronger claim than is warranted. We will revise to (1) add detail on case selection criteria, difficulty distribution, and body-part distribution, (2) explicitly state that the simulation has not been validated against official RCR sittings and that the official sittings are not publicly available, and (3) review all language referring to the FRCR to ensure it is consistently qualified as a simulation. We will ensure the abstract uses language such as 'simulated FRCR 2B Short Case examination' rather than language that could be read as claiming official FRCR-level competence. revision: yes

Circularity Check

0 steps flagged

No significant circularity found; central claims are evaluated against held-out data with externally-grounded scoring tools, though methodology concerns remain.

full rationale

Walking the paper's derivation chain, I find no step where a prediction or first-principles result reduces to its inputs by construction. The central claim — that HR1.5+ meets the simulated FRCR 2B Short Case passing standard (Table 4: 62.5%) — is evaluated on held-out exam sheets using an LLM-as-a-judge scoring rubric and Angoff-method thresholds. While the simulation, Angoff cut-offs, and expert verification are all internal to Harrison.ai (Section 2.2.1: 'We created a simulation of the FRCR examination, using external data that was not used for model training'), these are correctness/methodology risks (internal benchmark, unspecified judge model, small sample of 16 sheets), not circularity: the evaluation metric is not defined in terms of the model's output, and the model is not trained on the evaluation data. The Findings-Diagnosis scoring framework (Section 2.1) extends RadGraph-XL (Jain et al., 2021; Delbrouck et al., 2024) — an external tool — with ontology-based synonym matching and polarity-contradiction detection; it is a new evaluation metric, not a renaming of a known result or a self-definitional construct. RadBench and RadCoverage-VQA are internal datasets, but they are described as held-out evaluation sets, not training data repurposed as test data. The one self-citation (Harrison.ai, 2024) references prior HR1 results for context in the introduction and is not load-bearing for HR1.5's central claims. No uniqueness theorem is invoked, no ansatz is smuggled through self-citation, and no fitted parameter is renamed as a prediction. The concerns about LLM-as-a-judge verbosity bias, small confidence intervals, and internal benchmark construction are real but belong to correctness risk, not circularity.

Axiom & Free-Parameter Ledger

7 free parameters · 5 axioms · 4 invented entities

The paper introduces several internal frameworks and tools (Findings-Diagnosis scoring, HR1.5+ agentic system, confidence engine, C6 framework) whose parameters and implementation details are not specified. The training data, architecture, and scale are intentionally omitted. The FRCR simulation and Angoff thresholds are internally constructed and verified. The free parameters listed affect evaluation outcomes directly (Angoff thresholds, scoring thresholds, judge model choice) and are not independently reproducible.

free parameters (7)
  • Training data composition = ~6.5M studies, ~18M conversations
    Scale stated but composition (proportions by body part, pathology, oversampling rates) is not specified. Figure 1.3 states 'specific pipelines, prompts, generator models, and dataset scales are intentionally not shown.'
  • Hard negative curriculum schedule = Unspecified
    Section 1.2 describes breadth and depth of hard negatives but does not specify the curriculum schedule, difficulty thresholds, or how they are tuned.
  • Contrastive decoding parameters = Unspecified
    Section 1.3 describes an internally developed contrastive decoding approach 'adapted to our objective' but does not specify parameters.
  • Angoff cut-off estimates = Mean 73.2, range 69.3-77.3 per sheet
    Section 2.2.1: internally estimated and 'spot-checked and verified by internal expert radiologists.' The human expert cut-off came in at ~50%, significantly lower, but the Angoff threshold is used as the passing standard.
  • Findings-Diagnosis scoring thresholds = Unspecified
    Section 2.1: fuzzy token matching parameters, ontology synonym expansion rules, and polarity contradiction thresholds are not specified.
  • Confidence feature weights (w_e, w_d, w_cv) = Unspecified
    Section 4.3: U = w_e*H_attn + w_d*D_head + w_cv*CV. Weights for combining attention entropy, inter-head disagreement, and representation-trajectory instability are not stated.
  • Sink detection threshold = Unspecified
    Section 4.3: 'any position exceeding a fixed threshold is flagged as a sink' — the threshold value is not given.
axioms (5)
  • domain assumption RadGraph-XL entity extraction reliably extracts clinical entities (anatomy, observations, polarity) from both predicted and ground-truth radiology reports.
    Section 2.1: the entire Findings-Diagnosis scoring framework depends on RadGraph-XL extraction quality. The paper acknowledges this can be 'brittle' on structured reports (ReXGradient note).
  • domain assumption LLM-as-a-judge (Gu et al., 2024) provides a reliable adjudication of clinical correctness when automated finding-matching is insufficient.
    Section 2.1 and Table 3: the FRCR 2B Short Case is scored by an LLM-as-a-judge on a 0-5 rubric. The Judge F-D score on ReXGradient is presented as 'more representative of clinical correctness.' The specific judge model is not identified.
  • ad hoc to paper The internally constructed FRCR simulation faithfully represents the official RCR FRCR 2B Short Case examination in case selection, difficulty, and body-part distribution.
    Section 2.2.1: 'We created a simulation of the FRCR examination... we made our best effort to follow the published examination guidelines.' The official exams are not publicly available. The simulation uses 16 sheets of 25 cases each.
  • ad hoc to paper The Angoff cut-off estimated internally is a valid passing threshold for the simulated exam.
    Section 2.2.1: Angoff cut-offs were estimated and 'spot-checked and verified by internal expert radiologists.' The human expert cut-off was ~50%, significantly lower than the Angoff estimate (~73.2), which the authors interpret as holding models to a 'deliberately high standard.'
  • domain assumption Linear probe directions extracted from model activations correspond to clinically meaningful concepts that the model uses in its reasoning.
    Section 4.1: concept tracing assumes that a difference-of-means direction in the residual stream that separates positive from negative examples reflects a concept the model 'uses' rather than one that is merely decodable. The causal steering test provides partial evidence but is limited to logit shifts, not full output evaluation.
invented entities (4)
  • HR1.5+ agentic framework no independent evidence
    purpose: Extends the HR1.5 Core API with a system-of-experts architecture for report generation and chat interaction, consulting other harrison.ai models where appropriate.
    Section 1.3 describes the agentic configuration but does not specify which models are consulted, how routing decisions are made, or what the expert architecture looks like. No ablation isolating the agentic contribution is provided.
  • Findings-Diagnosis scoring framework no independent evidence
    purpose: Evaluates radiology report quality by comparing extracted findings with ontology-based synonym matching and polarity-contradiction detection.
    Section 2.1. Extends RadGraph-XL with an internal radiology knowledge graph. The knowledge graph, synonym expansion rules, and fuzzy matching parameters are not released or independently validated.
  • C6 evaluation framework no independent evidence
    purpose: Six-pillar clinical evaluation (Coverage, Correctness, Comprehensiveness, Care, Coherence, Conformity) for radiology report quality.
    Section 6.1. Proposed as future work; not yet implemented or validated. No results are presented.
  • Confidence estimation engine with sink correction no independent evidence
    purpose: Combines attention entropy, inter-head disagreement, and representation-trajectory instability into a single uncertainty score, corrected for attention sink artifacts.
    Section 4.3. The authors state this is 'an early implementation' that is 'not yet tuned/calibrated.' No quantitative evaluation of calibration (e.g., ECE, Brier score) is provided. Only qualitative worked examples are shown.

pith-pipeline@v1.1.0-glm · 35482 in / 4152 out tokens · 584099 ms · 2026-07-08T21:43:21.815808+00:00 · methodology

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

73 extracted references · 73 canonical work pages · 21 internal anchors

  1. [1]

    A Vision-Language Foundation Model to Enhance Efficiency of Chest X-ray Interpretation

    Chen, Zhihong and others , title =. 2024 , note =. 2401.12208 , archivePrefix =

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  2. [2]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    Alayrac, Jean-Baptiste and others , title =. Advances in Neural Information Processing Systems (NeurIPS) , year =

    work page

  3. [3]

    International Conference on Machine Learning (ICML) , year =

    Radford, Alec and others , title =. International Conference on Machine Learning (ICML) , year =

    work page

  4. [4]

    International Conference on Learning Representations (ICLR) , year =

    Yuksekgonul, Mert and others , title =. International Conference on Learning Representations (ICLR) , year =

    work page

  5. [5]

    and others , title =

    Selvaraju, Ramprasaath R. and others , title =. International Conference on Computer Vision (ICCV) , year =

    work page

  6. [6]

    Annual Meeting of the Association for Computational Linguistics (ACL) , year =

    Abnar, Samira and Zuidema, Willem , title =. Annual Meeting of the Association for Computational Linguistics (ACL) , year =

    work page

  7. [7]

    Conference on Computer Vision and Pattern Recognition (CVPR) , year =

    Chefer, Hila and Gur, Shir and Wolf, Lior , title =. Conference on Computer Vision and Pattern Recognition (CVPR) , year =

    work page

  8. [8]

    International Conference on Learning Representations (ICLR) , year =

    Galil, Ido and others , title =. International Conference on Learning Representations (ICLR) , year =

    work page

  9. [9]

    arXiv preprint arXiv:2603.05498 , year =

    Sun, Mingjie and others , title =. arXiv preprint arXiv:2603.05498 , year =

    work page arXiv

  10. [10]

    NeurIPS Datasets and Benchmarks , year =

    Jain, Saahil and others , title =. NeurIPS Datasets and Benchmarks , year =

    work page

  11. [11]

    Findings of the Association for Computational Linguistics (ACL) , year =

    Delbrouck, Jean-Benoit and others , title =. Findings of the Association for Computational Linguistics (ACL) , year =

    work page

  12. [12]

    A Survey on LLM-as-a-Judge

    Gu, Jiawei and others , title =. arXiv preprint arXiv:2411.15594 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  13. [13]

    Conference on Empirical Methods in Natural Language Processing (EMNLP) , year =

    Smit, Akshay and others , title =. Conference on Empirical Methods in Natural Language Processing (EMNLP) , year =

    work page

  14. [14]

    Text Summarization Branches Out, ACL Workshop , year =

    Lin, Chin-Yew , title =. Text Summarization Branches Out, ACL Workshop , year =

    work page

  15. [15]

    Annual Meeting of the Association for Computational Linguistics (ACL) , year =

    Papineni, Kishore and Roukos, Salim and Ward, Todd and Zhu, Wei-Jing , title =. Annual Meeting of the Association for Computational Linguistics (ACL) , year =

    work page

  16. [16]

    Patterns , year =

    Yu, Feiyang and others , title =. Patterns , year =

    work page

  17. [17]

    ReXGradient-160K: A Large-Scale Publicly Available Dataset of Chest Radiographs with Free-text Reports

    Zhang, Xiaoman and others , title =. arXiv preprint arXiv:2505.00228 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  18. [18]

    LLaVA-Med: Training a Large Language-and-Vision Assistant for Biomedicine in One Day

    Li, Chunyuan and others , title =. arXiv preprint arXiv:2306.00890 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  19. [19]

    MedGemma Technical Report

    Sellergren, Andrew and others , title =. arXiv preprint arXiv:2507.05201 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  20. [20]

    , title =

    Angoff, William H. , title =. Educational Measurement , year =

    work page

  21. [21]

    arXiv preprint arXiv:2604.00493 , year =

    Zhang, Xiaoman and others , title =. arXiv preprint arXiv:2604.00493 , year =

    work page arXiv

  22. [22]

    ICLR Workshop , year =

    Alain, Guillaume and Bengio, Yoshua , title =. ICLR Workshop , year =

    work page

  23. [23]

    The Geometry of Truth: Emergent Linear Structure in Large Language Model Representations of True/False Datasets

    Marks, Samuel and Tegmark, Max , title =. arXiv preprint arXiv:2310.06824 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  24. [24]

    Steering Language Models With Activation Engineering

    Turner, Alexander Matt and others , title =. arXiv preprint arXiv:2308.10248 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  25. [25]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    Li, Kenneth and others , title =. Advances in Neural Information Processing Systems (NeurIPS) , year =

    work page

  26. [26]

    Annual Meeting of the Association for Computational Linguistics (ACL) , year =

    Rimsky, Nina and others , title =. Annual Meeting of the Association for Computational Linguistics (ACL) , year =

    work page

  27. [27]

    , title =

    Jain, Sarthak and Wallace, Byron C. , title =. NAACL-HLT , year =

    work page

  28. [28]

    Conference on Empirical Methods in Natural Language Processing (EMNLP) , year =

    Wiegreffe, Sarah and Pinter, Yuval , title =. Conference on Empirical Methods in Natural Language Processing (EMNLP) , year =

    work page

  29. [29]

    International Conference on Learning Representations (ICLR) , year =

    Xiao, Guangxuan and others , title =. International Conference on Learning Representations (ICLR) , year =

    work page

  30. [30]

    Language Models (Mostly) Know What They Know

    Kadavath, Saurav and others , title =. arXiv preprint arXiv:2207.05221 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  31. [31]

    International Conference on Machine Learning (ICML) , year =

    Guo, Chuan and others , title =. International Conference on Machine Learning (ICML) , year =

    work page

  32. [32]

    Scientific Data , year =

    Abedeen, Iftekharul and others , title =. Scientific Data , year =

    work page

  33. [33]

    European Conference on Computer Vision (ECCV) , year =

    Boecking, Benedikt and others , title =. European Conference on Computer Vision (ECCV) , year =

    work page

  34. [34]

    and Nodine, Calvin F

    Kundel, Harold L. and Nodine, Calvin F. and Carmody, Dennis , title =. Investigative Radiology , year =

    work page

  35. [35]

    2024 , url =

    Clinical Radiology Workforce Census 2024 , institution =. 2024 , url =

    work page 2024

  36. [36]

    Journal of the American College of Radiology , year =

    Projected. Journal of the American College of Radiology , year =

    work page

  37. [37]

    and others , title =

    Krupinski, Elizabeth A. and others , title =. Journal of the American College of Radiology , volume =

    work page

  38. [38]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    Vaswani, Ashish and others , title =. Advances in Neural Information Processing Systems (NeurIPS) , year =

    work page

  39. [39]

    International Conference on Learning Representations (ICLR) , year =

    Dosovitskiy, Alexey and others , title =. International Conference on Learning Representations (ICLR) , year =

    work page

  40. [40]

    International Conference on Machine Learning (ICML) , year =

    Li, Junnan and others , title =. International Conference on Machine Learning (ICML) , year =

    work page

  41. [41]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    Liu, Haotian and others , title =. Advances in Neural Information Processing Systems (NeurIPS) , year =

    work page

  42. [42]

    arXiv preprint arXiv:2303.08774 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  43. [43]

    Nature Biomedical Engineering , volume =

    Tiu, Ekin and others , title =. Nature Biomedical Engineering , volume =. 2022 , url =

    work page 2022

  44. [44]

    BiomedCLIP: a multimodal biomedical foundation model pretrained from fifteen million scientific image-text pairs

    Zhang, Sheng and others , title =. arXiv preprint arXiv:2303.00915 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  45. [45]

    Towards Generalist Foundation Model for Radiology by Leveraging Web-scale 2D&3D Medical Data

    Wu, Chaoyi and others , title =. arXiv preprint arXiv:2308.02463 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  46. [46]

    ELIXR: Towards a general purpose X-ray artificial intelligence system through alignment of large language models and radiology vision encoders

    Xu, Shawn and others , title =. arXiv preprint arXiv:2308.01317 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  47. [47]

    MAIRA-1: A specialised large multimodal model for radiology report generation

    Hyland, Stephanie L. and others , title =. arXiv preprint arXiv:2311.13668 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  48. [48]

    MAIRA-2: Grounded Radiology Report Generation

    Bannur, Shruthi and others , title =. arXiv preprint arXiv:2406.04449 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  49. [49]

    Capabilities of Gemini Models in Medicine

    Saab, Khaled and others , title =. arXiv preprint arXiv:2404.18416 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  50. [50]

    Advancing Multimodal Medical Capabilities of Gemini

    Yang, Lin and others , title =. arXiv preprint arXiv:2405.03162 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  51. [51]

    Nature , year =

    Wang, others , title =. Nature , year =

    work page

  52. [52]

    2024 , url =

    Reimagining Medical AI:. 2024 , url =

    work page 2024

  53. [53]

    Annual Meeting of the Association for Computational Linguistics (ACL) , year =

    Li, Xiang Lisa and others , title =. Annual Meeting of the Association for Computational Linguistics (ACL) , year =

    work page

  54. [54]

    Conference on Computer Vision and Pattern Recognition (CVPR) , year =

    Leng, Sicong and others , title =. Conference on Computer Vision and Pattern Recognition (CVPR) , year =

    work page

  55. [55]

    SigLIP 2: Multilingual Vision-Language Encoders with Improved Semantic Understanding, Localization, and Dense Features

    Tschannen, Michael and others , title =. arXiv preprint arXiv:2502.14786 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  56. [56]

    ACM Symposium on Operating Systems Principles (SOSP) , year =

    Kwon, Woosuk and others , title =. ACM Symposium on Operating Systems Principles (SOSP) , year =

    work page

  57. [57]

    Qwen2-VL: Enhancing Vision-Language Model's Perception of the World at Any Resolution

    Wang, Peng and others , title =. arXiv preprint arXiv:2409.12191 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  58. [58]

    PaliGemma: A versatile 3B VLM for transfer

    Beyer, Lucas and others , title =. arXiv preprint arXiv:2407.07726 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  59. [59]

    American Journal of Roentgenology , year =

    Hong, others , title =. American Journal of Roentgenology , year =

    work page

  60. [60]

    Healthcare AI Challenge --- Results , year =

    work page

  61. [61]

    2025 , url =

    Clinical Radiology Workforce Census 2025 , institution =. 2025 , url =

    work page 2025

  62. [62]

    Radiology: Artificial Intelligence , volume =

    Reyes, Mauricio and others , title =. Radiology: Artificial Intelligence , volume =. 2020 , doi =

    work page 2020

  63. [63]

    Machine Learning for Healthcare (MLHC), PMLR 106 , year =

    Tonekaboni, Sana and others , title =. Machine Learning for Healthcare (MLHC), PMLR 106 , year =

    work page

  64. [64]

    European Conference on Computer Vision (ECCV) , year =

    Hendricks, Lisa Anne and others , title =. European Conference on Computer Vision (ECCV) , year =

    work page

  65. [65]

    2026 , howpublished =

    work page 2026

  66. [66]

    2025 , howpublished =

    work page 2025

  67. [67]

    2025 , howpublished =

    The. 2025 , howpublished =

    work page 2025

  68. [68]

    Annual Meeting of the Association for Computational Linguistics (ACL) , year =

    Xie, Yiqing and others , title =. Annual Meeting of the Association for Computational Linguistics (ACL) , year =

    work page

  69. [69]

    Findings of the Association for Computational Linguistics (EMNLP) , year =

    Ostmeier, Sophie and others , title =. Findings of the Association for Computational Linguistics (EMNLP) , year =

    work page

  70. [70]

    arXiv preprint arXiv:2604.24001 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  71. [71]

    Medical Image Computing and Computer Assisted Intervention (MICCAI) , year =

    Hamamci, Ibrahim Ethem and others , title =. Medical Image Computing and Computer Assisted Intervention (MICCAI) , year =

    work page

  72. [72]

    arXiv preprint arXiv:2406.07146 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  73. [73]

    , title =

    Ghassemi, Marzyeh and Oakden-Rayner, Luke and Beam, Andrew L. , title =. The Lancet Digital Health , volume =. 2021 , doi =

    work page 2021