CoChat
Sign Up
Cited by 0Multimodal Machine Learning ApplicationsArtificial Intelligence in Healthcare and EducationCOVID-19 diagnosis using AI

Swin-Qwen3: a three-stage vision–language framework for automated radiology report generation with multi-agent verification

Open access
Abdaoui, Hamida, Barbaria, Sabri, Al Kuwaiti, Ahmed, Raouafi, Noureddine, Ben Salah, Ridha, Boussi Rahmouni, Hanene·2026·Frontiers in Radiology
DOI 10.3389/fradi.2026.1875905PMC PMC13458744
Open this paper in CoChat — free accountOpen PDF

Abstract

Background: Automated chest x-ray reporting could substantially reduce the burden on radiology services worldwide; however, the implementation of current vision–language models (VLMs) in clinical workflows remains limited by factual errors, hallucinations, and inadequate clinical reliability. Bridging this implementation gap requires frameworks that are not only technically sound but also designed for safe integration into real-world healthcare settings. Objective: This study aimed to improve the clinical accuracy and factual consistency of radiology report generation by introducing an agent-inspired, three-stage reasoning framework and evaluating its feasibility as research prototype for potential implementation in resource-constrained and high-throughput clinical environments. Methods: We propose the Swin-Qwen3 vision–language architecture, which integrates a Swin Transformer visual encoder with a Q-Former-style query-driven cross-attention module aligned with a large language model (Qwen3-0.6B). A three-stage generation strategy—comprising initial report drafting, clinical verification, and structured refinement—was guided by role-specific prompts to simulate distinct clinical reasoning behaviours. Parameter-efficient fine-tuning via low-rank adaptation (LoRA) enabled training within standard GPU constraints. The framework was evaluated on the full IU x-Ray test set (321 Samples) using lexical, semantic, clinical, and factuality metrics and compared with one- and two-stage ablations. Computational feasibility and inference overhead were assessed for offline or batch processing contexts. Results: The three-stage framework showed modest but consistent improvements over ablation baselines. CheXpert-F1 reached 0.7038 (compared to 0.7154 for one-stage and 0.7123 for two-stage), Clinical-F1 reached 0.5401 (compared to 0.5478 for one-stage and 0.5449 for two-stage), and RadGraph-F1 scored 0.5467 (compared to 0.5478 for one-stage and 0.5445 for two-stage). The factuality score reached 0.843 (compared to 0.845 for one-stage and 0.845 for two-stage), while the hallucination rate remained at 0.6116 (compared to 0.6109 for one-stage and 0.6109 for two-stage). Clinical recall decreased from 0.8361 (one-stage) and 0.8337 (two-stage) to 0.8044 (three-stage), reflecting a trade-off between sensitivity and precision. However, the clinical verifier demonstrated substantial effectiveness, correcting 85.6% of identified hallucinations. Lexical quality improved: BLEU-4 was 0.0628 (+10.9% vs. one-stage; +13.1% vs. two-stage), ROUGE-2-F was 0.0957 (+10.2% vs. one-stage; +13.1% vs. two-stage), and METEOR was 0.3728 (+4.5% vs. one-stage; +6.3% vs. two-stage). Improvements in METEOR ( p < 0.001) and ROUGE-2-F ( p < 0.05) were statistically significant. A 14.8 × inference overhead was identified, indicating suitability for offline reporting contexts. Conclusions: The Swin-Qwen3 three-stage framework demonstrates that explicit clinical verification and iterative refinement can modestly enhance the clinical reliability of VLMs for automated radiology reporting. The parameter-efficient design is scalable across diverse clinical settings, including resource-limited environments, and supports a practical deployment pathway aligned with SDGs 3 (Good Health and Well-being), 9 (Industry, Innovation and Infrastructure), and 17 (Partnerships for the Goals). However, the residual hallucination rate of 61.2% indicates that the current system is a promising research prototype rather than a clinically deployable tool. Future work should prioritize multi-institutional validation on larger datasets (e.g., MIMIC-CXR) with radiologist expert review, regulatory evaluation, and prospective clinical integration studies.

AI snapshot

Written from the paper’s full text.

TL;DR

Swin-Qwen3 is a three-stage vision–language framework combining a frozen Swin Transformer, Q-Former-style cross-attention, and LoRA-adapted Qwen3-0.6B for chest x-ray report generation, evaluated on the full IU x-Ray test set (321 samples). It improved lexical metrics and corrected 85.6% of identified hallucinations, but the residual hallucination rate remained 61.2%, so the authors describe it as a research prototype rather than a clinically deployable tool.

Key findings

  • The three-stage system improved lexical quality over one- and two-stage ablations: BLEU-4 = 0.0628 (+10.9% vs. one-stage; +13.1% vs. two-stage), ROUGE-2-F = 0.0957 (+10.2%; +13.1%), and METEOR = 0.3728 (+4.5%; +6.3%). METEOR and ROUGE-2-F improvements were statistically significant relative to the two-stage baseline (p < 0.001 and p < 0.05, respectively).
  • Clinical and factuality metrics were mixed or slightly lower: CheXpert-F1 = 0.7038 vs. 0.7154 one-stage and 0.7123 two-stage; Clinical-F1 = 0.5401 vs. 0.5478 and 0.5449; RadGraph-F1 = 0.5467 vs. 0.5478 and 0.5445; Clinical Recall decreased to 0.8044 from 0.8361 and 0.8337. Factuality Score = 0.843 vs. 0.845 and 0.845, and Hallucination Rate = 0.6116 vs. 0.6109 and 0.6109.
  • The clinical verifier corrected 85.6% of identified hallucinations, with precision, recall, and F1 of 0.8563; it corrected an average of 5.52 true hallucinations per sample while making 0.86 false corrections. Critical hallucinations (1,186) outnumbered moderate ones (855), and 99.7% of samples contained residual hallucinations.
  • A 14.8× inference overhead was identified, indicating suitability for offline or batch processing contexts rather than real-time point-of-care use.

Methods

  • Dataset: IU x-Ray, filtered to 2,134 diagnostic reports paired with frontal and lateral chest radiographs (4,268 images total), split at report level 75%/15%/15% into 1,493 training, 320 validation, and 321 test samples. Images were resized to 224 × 224, normalized with ImageNet statistics, and augmented during training; reports were tokenized with the Qwen3 tokenizer at maximum length 128.
  • Architecture and training: a frozen Swin Transformer encoded dual-view images; a Q-Former-style query-driven cross-attention module produced visual prefixes for Qwen3-0.6B. LoRA was applied to attention projections (q_proj, k_proj, v_proj, o_proj) with rank r = 8, α = 16, dropout 0.1, and 0.53% trainable parameters. Training used AdamW for eight epochs, learning rate 1 × 10⁻⁴, weight decay 0.01, effective batch size four via gradient accumulation, and FP16.
  • Generation and evaluation: a three-stage inference pipeline used role-specific prompts for draft generation, clinical verification, and refinement. Evaluation on the 321-sample IU x-Ray test set used lexical, semantic, clinical, and factuality metrics, compared against one- and two-stage ablations, with Shapiro–Wilk, Wilcoxon/paired t-tests, Cohen’s d, McNemar tests, and Cohen’s Kappa.

Limitations

  • Stated: evaluation was limited to the IU x-Ray test set (321 samples); external validation on MIMIC-CXR or CheXpert Plus is needed. The study relied primarily on automated text-comparison metrics rather than expert clinical review, and prospective clinical studies are required.
  • Stated: the residual hallucination rate of 61.2% indicates the system is a research prototype, not clinically deployable; Clinical Recall decreased by −3.8% (p < 0.001), reflecting a sensitivity–precision trade-off. The 14.8× inference overhead may limit real-time use, and the generalised verifier may lack specialised knowledge for rare pathologies or complex multi-system findings.
  • Evident from the text: despite the verifier’s 85.6% correction rate, hallucination rate showed no improvement over baselines, 99.7% of samples had residual hallucinations, and aggregate clinical metrics such as CheXpert-F1 and Clinical-F1 were slightly lower than one- and two-stage baselines.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.3389/fradi.2026.1875905

Cites

55 of 56 references verified0 citation errors · 1 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

32 of 40 cited sources checked · 13 supported · 4 flagged · accuracy not yet measured

  • Findingcitation

    Our BLEU-4 score of 0.0628 is lower than R2Gen (0.103) and R2GenCMN (0.106) on IU x-ray.

    “Our BLEU-4 score of 0.0628 is lower than R2Gen (20) and R2GenCMN (48), which reported BLEU-4 scores of 0.103 and 0.106, respectively, on IU x-ray.”
    Quote in the paperNot a substantial claim2 of 2 sources checked
    • cites [20] Chen, 2020, Generating radiology reports via memory-driven transformer source

      Table 3 (Section 3.2) reports a BLEU-4 score of 0.165 for the paper’s model on IU X-RAY. The claimed 0.103 is its score on MIMIC-CXR, not IU X-RAY. The citing authors’ own score and R2GenCMN’s score are not attributed to this source. (LLM, conf=0.90)

      Checked by a reasoning model — may be wrong.

      Experimental — accuracy not yet measured
    • cites [48] Chen, 2022, Cross-modal memory networks for radiology report generation source

      The results table in §4.1 (p. 6) reports R2GEN’s BLEU-4 score as 0.165 on IU X-RAY, not 0.103; it lists 0.103 for R2GEN on MIMIC-CXR. The claim assigns that score to the wrong dataset, and the cited passage does not establish the claimed 0.106 IU X-RAY score for R2GenCMN. (LLM, conf=0.88)

      Checked by a reasoning model — may be wrong.

      Experimental — accuracy not yet measured
  • Premisecitation

    These delays are particularly pronounced in resource-constrained environments, such as rural hospitals, emergency departments, and low-income regions, where access to experienced radiologists is limited or inconsistent.

    “These delays are particularly pronounced in resource-constrained environments, such as rural hospitals, emergency departments, and low-income regions, where access to experienced radiologists is limited or inconsistent (3, 4).”
    Quote in the paperNot a substantial claim1 of 2 sources checked
    • cites [3] Miles, 2025, Advancing health care delivery in community, rural, and safety-net settings, J Am Coll Radiol source

      The Introduction (page unavailable) links access barriers to delayed or missed care, and “Shared Challenges” (page unavailable) identifies radiologist shortages in community, rural, and safety-net settings. The paper does not establish that delays are particularly pronounced in those settings, or that limited access to experienced radiologists causes delays in emergency departments specifically. (LLM, conf=0.84)

      Experimental — accuracy not yet measured
    • cites [4] Rohatgi, 2015, After-hours radiology: challenges and strategies for the radiologist, Am J Roentgenol source
  • Premisecitation

    Many existing vision-language models rely on simplistic feature concatenation strategies that fail to capture fine-grained correspondences between localised radiographic findings and their semantic descriptions.

    “these models often rely on simplistic feature concatenation strategies that fail to capture fine-grained correspondences between localised radiographic findings and their semantic descriptions (11, 16).”
    Quote in the paperSubstantial claim2 of 2 sources checked
    • cites [11] Artsi, 2025, “Large language models in radiology reporting—a systematic review of performance, limitations, and clinical implications” source

      The Discussion addresses input quality and limitations in radiology reporting, but not feature concatenation or fine-grained image–text correspondence. Study Selection and Characteristics describes one model combining regional image features with knowledge-graph features; it does not say this was simplistic concatenation or that it failed to link localised findings to descriptions. Searches for those terms found no supporting passage. Page numbers were unavailable in our copy. (LLM, conf=0.75)

      Experimental — accuracy not yet measured
    • cites [16] Li, 2025, Automatic radiology report generation with deep learning: a comprehensive review of methods and advances, Artif Intell Rev source

      Section 4.2.2 describes some report-generation methods that use straightforward concatenation, and §3.1 identifies a deficiency in precise image–text correspondence. The review does not say that many vision-language models use concatenation or establish that concatenation causes the failure to align localized findings with their descriptions. Page numbers were not available in our copy. (LLM, conf=0.84)

      Experimental — accuracy not yet measured
  • Premisecitation

    Direct comparisons are difficult due to differing evaluation settings.

    “This is lower than MRGAgents' (35) BLEU-4 score of 0.191 on IU x-ray; however, as ReXrank (46) notes, direct comparisons are difficult due to differing evaluation settings.”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [46] Zhang, 2024, ReXrank: a public leaderboard for AI-powered radiology report generation source
  • Findingcitation

    The BLEU-4 score of 0.0628 is lower than MRGAgents' BLEU-4 score of 0.191 on IU x-ray.

    “This is lower than MRGAgents' (35) BLEU-4 score of 0.191 on IU x-ray; however, as ReXrank (46) notes, direct comparisons are difficult due to differing evaluation settings.”
    Quote in the paperNot a substantial claim1 of 2 sources checked
    • cites [35] Wang, 2025

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites [46] Zhang, 2024, ReXrank: a public leaderboard for AI-powered radiology report generation source
  • May combine two claims. 45
  • Classified as probably not a substantial claim. 30

  • May combine two claims.

    Premisecitation

    These models do not provide comprehensive clinical metrics like CheXpert-F1 or RadGraph-F1, preventing a direct comparison on clinically relevant dimensions.

    “However, these models do not provide comprehensive clinical metrics like CheXpert-F1 or RadGraph-F1, preventing a direct comparison on clinically relevant dimensions.”
    Quote in the paperSubstantial claim2 of 2 sources checked
    • cites [20] Chen, 2020, Generating radiology reports via memory-driven transformer source

      Section 3.2 reports clinical-efficacy precision, recall, and F1 using CheXpert labels, including F1 scores for models on MIMIC-CXR. This contradicts the claim that the paper does not provide CheXpert-F1 or permit comparison on that clinical dimension. The paper notes that these metrics do not apply to IU X-RAY; the cited passages do not report RadGraph-F1. (LLM, conf=0.90)

      Checked by a reasoning model — may be wrong.

      Experimental — accuracy not yet measured
    • cites [48] Chen, 2022, Cross-modal memory networks for radiology report generation source

      Section 3.2 (p. 5) says the paper evaluates clinical efficacy using CheXpert labels and reports precision, recall, and F1. The results table in Section 4.1 (p. 6) gives those clinical metrics for MIMIC-CXR models, contrary to the claim that they provide no CheXpert-F1 for clinical comparison. The paper does not report RadGraph-F1, and its CheXpert-based metrics do not apply to IU X-RAY. (LLM, conf=0.90)

      Checked by a reasoning model — may be wrong.

      Experimental — accuracy not yet measured

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    Chest radiography is one of the most frequently performed diagnostic examinations and plays a key role in the detection and monitoring of pulmonary diseases, cardiac abnormalities, infections, and thoracic trauma.

    “Chest radiography is one of the most frequently performed diagnostic examinations and plays a key role in the detection and monitoring of pulmonary diseases, cardiac abnormalities, infections, and thoracic trauma (5, 6).”
    Quote in the paperNot a substantial claim1 of 2 sources checked
    • cites [5] Miró Catalina, 2024, Real-world testing of an artificial intelligence algorithm for the analysis of chest x-rays in primary care settings, Sci Rep source

      The Introduction (page unavailable) calls chest X-rays one of the most common tests and says they are used to detect pulmonary and cardiovascular diseases. The copy does not establish their role in monitoring disease or specifically in detecting infections and thoracic trauma; the Results list abnormalities such as consolidation and abnormal ribs without making those broader claims. (LLM, conf=0.85)

      Experimental — accuracy not yet measured
    • cites [6] Sargent, 2023, The sensitivity of chest x-ray (CXR) for the detection of significant thoracic injury in children exposed to blast, Injury source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    In emergency and intensive care settings, timely interpretation of chest radiographs is essential for guiding clinical decisions, including ventilation management, antibiotic initiation, and imaging referrals.

    “In emergency and intensive care settings, timely interpretation of chest radiographs is essential for guiding clinical decisions, including ventilation management, antibiotic initiation, and imaging referrals (7, 8).”
    Quote in the paperNot a substantial claim1 of 2 sources checked
    • cites [7] Tanabe, 2024, Lung imaging in COPD and asthma, Respir Investig source
    • cites [8] Bonaffini, 2023, Chest x-ray at emergency admission and potential association with barotrauma in mechanically ventilated patients: experience from the Italian core of the first pandemic peak, Tomography source

      The Discussion says chest X-ray is an early emergency-room investigation and examines whether admission findings identify patients at risk of barotrauma. Methods §2.3 says barotrauma suspected or detected on CXR could prompt CT, supporting the imaging-referral example. The paper does not establish that timely interpretation guides ventilation management or antibiotic initiation. (LLM, conf=0.87)

      Experimental — accuracy not yet measured

  • Classified as probably not a substantial claim. May combine two claims.

    Premisecitation

    The growing gap between imaging demand and radiologist availability has intensified the challenge of radiology reporting, leading to delayed diagnoses and increased workloads across healthcare systems worldwide.

    “The growing gap between imaging demand and radiologist availability has intensified this challenge, leading to delayed diagnoses and increased workloads across healthcare systems worldwide (1, 2).”
    Quote in the paperSubstantial claim2 of 2 sources checked
    • cites [1] Pesapane, 2023, Advancements in standardizing radiological reports: a comprehensive review, Medicina (B Aires) source

      The Introduction describes the importance of radiology reports, and Section 4 discusses the effort required to standardize them and mentions workload guidelines. Neither passage addresses a growing gap between imaging demand and radiologist availability or links such a gap to delayed diagnoses and increased workloads worldwide. Searches for those terms and for reporting backlogs found no supporting passage in our copy. (LLM, conf=0.76)

      Experimental — accuracy not yet measured
    • cites [2] Ouis, 2024, ChestBioX-Gen: contextual biomedical report generation from chest x-ray images using BioGPT and co-attention mechanism, Front Imaging source

      The Abstract (page unavailable) says timely radiology reporting matters for diagnosis and describes reducing radiologists’ burden and enabling faster diagnoses. It does not say that imaging demand is outpacing radiologist availability or that this gap has caused delayed diagnoses and increased workloads worldwide. (LLM, conf=0.78)

      Experimental — accuracy not yet measured

  • May combine two claims.

    Premisecitation

    Key barriers to translation include the generation of clinically plausible but factually incorrect statements (hallucinations) and the lack of structured frameworks that decompose generation into auditable, verifiable steps amenable to clinical governance.

    “Key barriers include the generation of clinically plausible but factually incorrect statements—so-called hallucinations (11, 12)—and the lack of structured frameworks that decompose generation into auditable, verifiable steps amenable to clinical governance.”
    Quote in the paperSubstantial claim2 of 2 sources checked
    • cites [11] Artsi, 2025, “Large language models in radiology reporting—a systematic review of performance, limitations, and clinical implications” source

      Split into 2 atoms — 1 confirmed, 1 partially supported, 0 disputed, 0 misattributed, 0 unverifiable. Expand the claim to see per-atom verdicts.

      Experimental — accuracy not yet measured
    • cites [12] Salehi, 2025, Agentic AI and large language models in radiology: opportunities and hallucination challenges, Bioengineering source

  • May combine two claims.

    Premisecitation

    The high volume of chest x-rays, combined with the need for detailed narrative reporting, places significant pressure on radiologists, especially during night shifts and peak admission periods.

    “The high volume of chest x-rays, combined with the need for detailed narrative reporting, places significant pressure on radiologists, especially during night shifts and peak admission periods (9, 10).”
    Quote in the paperSubstantial claim2 of 2 sources checked
    • cites [9] Naidji, 2024, Automatic detection of COVID-19 from chest x-ray images using EfficientNet-B7 CNN model with channel-wise attention, Int J Com Dig Sys source

      The Introduction (page not provided) says the rapid spread of COVID-19 imposed a significant workload on radiologists and discusses faster interpretation of chest radiographs. It does not attribute that workload to detailed narrative reporting or identify night shifts and peak admission periods as especially demanding. (LLM, conf=0.88)

      Experimental — accuracy not yet measured
    • cites [10] Schalekamp, 2024, Performance of AI to exclude normal chest radiographs to reduce radiologists' Workload, Eur Radiol source

      The Introduction describes increasing radiologist workload and the potential to save time by reducing chest-radiograph reporting; the Discussion explicitly cites the high volume of chest radiographs. Neither section says that detailed narrative reports are required or that pressure is especially high during night shifts or peak admission periods. Page numbers are not available in our copy. (LLM, conf=0.88)

      Experimental — accuracy not yet measured

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Recent multi-agent frameworks demonstrated that task decomposition improves both precision and comprehensiveness.

    “Our results align with recent multi-agent frameworks (35, 49), which demonstrated that task decomposition improves both precision and comprehensiveness.”
    Quote in the paperNot a substantial claim1 of 2 sources checked
    • cites [35] Wang, 2025

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites [49] Yi, 2025 source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    This is similar to challenges observed in RAG systems where retrieved information is not effectively incorporated.

    “This is similar to challenges observed in RAG systems where retrieved information is not effectively incorporated (50, 51).”
    Quote in the paperNot a substantial claim2 of 2 sources checked
    • cites [50] Xia, 2024, Rule: reliable multimodal RAG for factuality in medical vision language models source
    • cites [51] MDocAgent: a multi-modal multi-agent framework for document understanding Han S Xia P Zhang R Sun T Li Y Zhu H 2025 source

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    ReXrank findings show that even top-performing models struggle with abnormal cases, achieving error rates of 45% for abnormal studies compared to over 90% for normal studies.

    “This aligns with ReXrank (46)'s findings that even top-performing models struggle with abnormal cases, achieving error rates of 45% for abnormal studies compared to over 90% for normal studies.”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [46] Zhang, 2024, ReXrank: a public leaderboard for AI-powered radiology report generation source

      The Results section compares models across datasets using report-generation metrics, but does not report error rates for abnormal versus normal studies. Searches for abnormal studies and the claimed 45% and over-90% rates found no passage addressing that comparison; a mention of abnormalities in §3.5 describes a model’s architecture, not its results. (LLM, conf=0.77)

      Experimental — accuracy not yet measured

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Yi et al. noted that incorporating retrieval and refinement can introduce redundancy, affecting overall consistency.

    “Similarly, Yi et al. (49) noted that incorporating retrieval and refinement can introduce redundancy, affecting overall consistency.”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [49] Yi, 2025 source

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    EviAgent demonstrated that while evidence-driven frameworks outperform MLLMs, they still face challenges with complex clinical scenarios.

    “Similarly, EviAgent (52) demonstrated that while evidence-driven frameworks outperform MLLMs, they still face challenges with complex clinical scenarios.”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [52] Qi, 2026 source

      Section 4.2 and the Conclusion report that EviAgent outperforms standard MLLM baselines. But Section 4.5 uses a complex ICU case to illustrate EviAgent’s precision, not a challenge it faces in complex clinical scenarios. A separate error analysis describes mislocalization and over-diagnosis caused by its perception tools, without tying those errors to complex scenarios. (LLM, conf=0.86)

      Experimental — accuracy not yet measured

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    The explicit separation of generation, verification, and refinement stages represents a departure from end-to-end approaches.

    “The explicit separation of generation, verification, and refinement stages represents a departure from end-to-end approaches (20, 48).”
    Quote in the paperNot a substantial claim2 of 2 sources checked
    • cites [20] Chen, 2020, Generating radiology reports via memory-driven transformer source
    • cites [48] Chen, 2022, Cross-modal memory networks for radiology report generation source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    EviAgent trades inference speed for clinical accuracy.

    “ReXrank (46) similarly noted that inference efficiency remains a challenge for multi-stage and agent-based systems, as also observed in EviAgent (52) which trades inference speed for clinical accuracy.”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [52] Qi, 2026 source

  • May combine two claims.

    Premisecitation

    ReXrank demonstrated that even top-performing models struggle with abnormal cases, with no significant error rates below 45% for abnormal studies.

    “ReXrank (46) demonstrated that even top-performing models struggle with abnormal cases, with no significant error rates below 45% for abnormal studies.”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [46] Zhang, 2024, ReXrank: a public leaderboard for AI-powered radiology report generation source

      The Results section compares model performance across datasets but does not report results separately for abnormal studies. Searches for “abnormal,” “45%,” and “significant error” found no abnormal-study error-rate analysis or 45% threshold in our copy; the Metrics section only describes measures that can identify clinically significant errors. (LLM, conf=0.76)

      Experimental — accuracy not yet measured

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Yi et al. showed that multi-agent collaboration, incorporating retrieval and refinement, produces more clinically grounded reports.

    “Similarly, Yi et al. (49) showed that multi-agent collaboration, incorporating retrieval and refinement, produces more clinically grounded reports.”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [49] Yi, 2025 source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    ReXrank noted that inference efficiency remains a challenge for multi-stage and agent-based systems.

    “ReXrank (46) similarly noted that inference efficiency remains a challenge for multi-stage and agent-based systems, as also observed in EviAgent (52) which trades inference speed for clinical accuracy.”
    Quote in the paperNot a substantial claim2 of 2 sources checked
    • cites [46] Zhang, 2024, ReXrank: a public leaderboard for AI-powered radiology report generation source

      The Introduction discusses reporting efficiency and benchmarking, while §3.5 describes models used sequentially; neither addresses inference efficiency. Searches for inference speed, computational cost, latency, and agent-based systems found no passage in our copy stating that inference efficiency remains a challenge for those systems. Page numbers were not available in the indexed passages. (LLM, conf=0.72)

      Experimental — accuracy not yet measured
    • cites [52] Qi, 2026 source

  • May combine two claims.

    Findingown data

    The factuality score for the three-stage framework was 0.843, compared to 0.845 for one-stage and 0.845 for two-stage.

    “The factuality score reached 0.843 (compared to 0.845 for one-stage and 0.845 for two-stage)”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    ROUGE-2-F for the three-stage framework was 0.0957, +10.2% vs one-stage and +13.1% vs two-stage.

    “ROUGE-2-F was 0.0957 (+10.2% vs. one-stage; +13.1% vs. two-stage)”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The hallucination rate for the three-stage framework was 0.6116, compared to 0.6109 for one-stage and 0.6109 for two-stage.

    “the hallucination rate remained at 0.6116 (compared to 0.6109 for one-stage and 0.6109 for two-stage)”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    RadGraph-F1 for the three-stage framework was 0.5467, compared to 0.5478 for one-stage and 0.5445 for two-stage.

    “RadGraph-F1 scored 0.5467 (compared to 0.5478 for one-stage and 0.5445 for two-stage)”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    METEOR for the three-stage framework was 0.3728, +4.5% vs one-stage and +6.3% vs two-stage.

    “METEOR was 0.3728 (+4.5% vs. one-stage; +6.3% vs. two-stage)”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    CheXpert-F1 for the three-stage framework was 0.7038, compared to 0.7154 for one-stage and 0.7123 for two-stage.

    “CheXpert-F1 reached 0.7038 (compared to 0.7154 for one-stage and 0.7123 for two-stage)”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Clinical-F1 for the three-stage framework was 0.5401, compared to 0.5478 for one-stage and 0.5449 for two-stage.

    “Clinical-F1 reached 0.5401 (compared to 0.5478 for one-stage and 0.5449 for two-stage)”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    BLEU-4 for the three-stage framework was 0.0628, +10.9% vs one-stage and +13.1% vs two-stage.

    “BLEU-4 was 0.0628 (+10.9% vs. one-stage; +13.1% vs. two-stage)”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    METEOR improved by +4.5% vs one-stage and +6.3% vs two-stage.

    “METEOR improved by +4.5% vs. one-stage and +6.3% vs. two-stage”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    BLEU-3 improved by +5.0% vs one-stage and +7.8% vs two-stage.

    “and BLEU-3 (+5.0% vs. one-stage; +7.8% vs. two-stage)”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    ROUGE-2-F demonstrated a relative improvement of +10.2% compared to one-stage and +13.1% compared to two-stage.

    “ROUGE-2-F demonstrated a substantial relative improvement of +10.2% compared to the one-stage baseline and +13.1% compared to the two-stage baseline”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    McNemar tests confirmed 3-Stage superiority for CheXpert-F1 (p < 0.001), with 49 vs. 16 and 51 vs. 16 improvements over baselines.

    “McNemar tests confirmed 3-Stage superiority for CheXpert-F1 (p < 0.001), with 49 vs. 16 and 51 vs. 16 improvements over baselines.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    The hallucination rate persisted at 61.2%, showing no improvement over baselines.

    “However, the hallucination rate persisted at 61.2%, showing no improvement over baselines.”
    • Also stated in: § Architectural performance and clinical utility · plain
      “the hallucination rate persisted at 61.2%, showing no improvement over baselines”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The 1-Stage generator produces a fluent but incomplete report, omitting the lobulated opacity and hyperexpanded lungs while underestimating severity.

    “The 1-Stage generator produces a fluent but incomplete report, omitting the lobulated opacity and hyperexpanded lungs while underestimating severity.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    The three-stage system attained METEOR = 0.3728, a +4.5% improvement vs. one-stage (p < 0.001).

    “Our three-stage framework achieved modest improvements in lexical quality, as evidenced by BLEU-4 (+10.9%), ROUGE-2-F (+10.2%, p < 0.05), and METEOR (+4.5%, p < 0.001).”
    • Also stated in: § Conclusion · plain
      “METEOR=0.3728 (+4.5%, p < 0.001)”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The 2-Stage refiner offers modest terminology improvements but does not address these omissions.

    “Although the 2-Stage refiner offers modest terminology improvements, it does not address these omissions.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    In the top 5 failure cases, correction rates ranged from 0% to 33.3%, significantly below the average of 85.6%.

    “In the top 5 failure cases, correction rates ranged from 0% to 33.3%, significantly below the average of 85.6%.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Compared with the one-stage model, 49 cases improved vs. 16 deteriorated (p < 0.001).

    “Compared with the one-stage model, 49 cases improved vs. 16 deteriorated (p < 0.001).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The three-stage system attained BLEU-4 = 0.0628, a +10.9% improvement vs. one-stage.

    “Our three-stage framework achieved modest improvements in lexical quality, as evidenced by BLEU-4 (+10.9%), ROUGE-2-F (+10.2%, p < 0.05), and METEOR (+4.5%, p < 0.001).”
    • Also stated in: § Conclusion · plain
      “the three-stage system attained BLEU-4 = 0.0628 (+10.9% vs. one-stage)”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The 3-Stage framework achieves the closest lexical and structural alignment with the ground truth but introduces a hallucinated finding ("Mild hyperinflation" instead of "hyperexpanded") and still misses the lobulated opacity.

    “The 3-Stage framework achieves the closest lexical and structural alignment with the ground truth but introduces a hallucinated finding (“Mild hyperinflation” instead of “hyperexpanded”) and still misses the lobulated opacity.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    CheXpert-F1 reached 0.7038 (compared to 0.7154 for one-stage and 0.7123 for two-stage).

    “CheXpert-F1 reached 0.7038 (compared to 0.7154 for one-stage and 0.7123 for two-stage), Clinical-F1 reached 0.5401, and RadGraph-F1 remained stable at 0.5467.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data

    BERTScore-F1 showed a slight decrease of −0.2%.

    “BERTScore-F1 showed a slight decrease of −0.2%”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    The three-stage system attained ROUGE-2-F = 0.0957, a +10.2% improvement vs. one-stage (p < 0.05).

    “Our three-stage framework achieved modest improvements in lexical quality, as evidenced by BLEU-4 (+10.9%), ROUGE-2-F (+10.2%, p < 0.05), and METEOR (+4.5%, p < 0.001).”
    • Also stated in: § Conclusion · plain
      “ROUGE-2-F = 0.0957 (+10.2%, p < 0.05)”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Compared with the two-stage model, 51 cases improved vs. 16 deteriorated (p < 0.001).

    “Likewise, compared with the two-stage model, 51 cases improved vs. 16 deteriorated (p < 0.001).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Hallucination Rate persisted at 61.2%, showing no improvement over baselines (61.1%).

    “Hallucination Rate persisted at 61.2%, showing no improvement over baselines (61.1%)”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Two primary failure modes were identified: verifier omission and verifier hallucination.

    “Two primary failure modes were identified: (i) verifier omission, where existing hallucinations were not identified, and (ii) verifier hallucination, where new errors were introduced.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    BLEU-4 improved by +10.9% vs one-stage and +13.1% vs two-stage.

    “Gains were particularly evident in BLEU-4 (+10.9% vs. one-stage; +13.1% vs. two-stage)”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The verifier corrects an average of 5.52 true hallucinations per sample while making only 0.86 false corrections.

    “It corrects an average of 5.52 true hallucinations per sample while making only 0.86 false corrections.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Our framework, with a much smaller model (0.6B vs. 8B), demonstrates competitive clinical performance with Semb: 0.656 (scaled: 65.6) and RadCliQ: 0.688 (scaled inverse: ∼145).

    “Our framework, with a much smaller model (0.6B vs. 8B), demonstrates competitive clinical performance with Semb: 0.656 (scaled: 65.6) and RadCliQ: 0.688 (scaled inverse: ∼145).”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Findingnone given

    The parameter-efficient design (Qwen3-0.6B with LoRA adaptation r = 8, α = 16 and a frozen Swin encoder) requires only commodity GPU hardware for both training and inference.

    “First, the parameter-efficient design, Qwen3-0.6B with LoRA adaptation (r = 8, α = 16) and a frozen Swin encoder, requires only commodity GPU hardware for both training and inference, lowering the infrastructure barrier for adoption in resource-limited healthcare settings, including rural hospitals and low-to-middle-income country facilities.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Interpretationown data

    The 14.8× inference overhead indicates suitability for offline reporting contexts.

    “indicating suitability for offline reporting contexts”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given

    The parameter-efficient design is scalable across diverse clinical settings, including resource-limited environments.

    “The parameter-efficient design is scalable across diverse clinical settings, including resource-limited environments”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data

    The residual hallucination rate of 61.2% indicates that the current system is a promising research prototype rather than a clinically deployable tool.

    “the residual hallucination rate of 61.2% indicates that the current system is a promising research prototype rather than a clinically deployable tool”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data

    The persistent 61.2% hallucination rate confirms that this remains a research prototype requiring substantial further validation before clinical deployment.

    “The persistent 61.2% hallucination rate confirms that this remains a research prototype requiring substantial further validation before clinical deployment.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing). May combine two claims.

    Interpretationown data

    The susceptibility to unsupported predictions may stem from limitations in the verifier's encoded clinical knowledge or the inherent ambiguity of certain medical presentations.

    “This susceptibility may stem from limitations in the verifier's encoded clinical knowledge or the inherent ambiguity of certain medical presentations.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

    Interpretationnone given

    The phased approach aligns with established implementation science frameworks for health technology adoption.

    “This phased approach aligns with established implementation science frameworks for health technology adoption and supports potential SDG 17 (Partnerships for the Goals) by enabling collaborative, multi-institutional rollout strategies, pending prospective clinical validation.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Interpretationreasoning

    Real-time applications at the point of care may find the 14.8× inference overhead limiting.

    “Real-time applications at the point of care may find this overhead limiting; however, the framework remains well-suited for batch processing scenarios (e.g., offline preliminary report generation for research purposes).”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Premisecitation

    MRGAgents observed that their multi-agent framework, despite improving comprehensiveness, resulted in lower ROUGE-L scores due to structural differences in generated reports.

    “MRGAgents (53) also observed that their multi-agent framework, despite improving comprehensiveness, resulted in lower ROUGE-L scores due to structural differences in generated reports.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [53] Wang, 2025, Mrgagents: a multi-agent framework for improved medical report generation with med-lvlms

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Probably not a substantial claim: may not pass S2 (checkable). May combine two claims.

    Interpretationreasoning

    Prioritizing clinical accuracy is a justified design choice, supporting the use of clinical metrics over lexical metrics as the primary evaluation criteria for implementation-focused systems.

    “These results underscore that prioritizing clinical accuracy is a justified design choice, supporting the use of clinical metrics over lexical metrics as the primary evaluation criteria for implementation-focused systems.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim. May combine two claims.

    Interpretationreasoning

    The explicit verification and refinement stages allow researchers to audit where errors occur, facilitating targeted improvements.

    “The explicit verification and refinement stages allow researchers to audit where errors occur—whether in the generator, verifier, or refiner—thus facilitating targeted improvements.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

    Interpretationreasoning

    The modular three-stage architecture facilitates a staged research roadmap.

    “Fourth, the modular three-stage architecture facilitates a staged research roadmap.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    The regression in Clinical Recall reflects a sensitivity-precision trade-off, where the verifier prioritizes removing unsupported findings over exhaustive enumeration.

    “The regression in Clinical Recall (−3.8%, p < 0.001) reflects a sensitivity-precision trade-off, where the verifier prioritizes removing unsupported findings over exhaustive enumeration.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: In decision-support contexts, false positives are more harmful than missed findings.

  • Classified as probably not a substantial claim.

    Interpretationreasoning

    The three-stage pipeline's explicit separation of drafting, verification, and refinement creates natural audit trails and interpretable intermediate outputs.

    “Second, the three-stage pipeline's explicit separation of drafting, verification, and refinement creates natural audit trails and interpretable intermediate outputs.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data

    The three-stage architecture enhances linguistic quality and structural coherence, but challenges persist in detecting subtle mediastinal abnormalities and controlling hallucinations.

    “These results demonstrate that while the three-stage architecture enhances linguistic quality and structural coherence, challenges persist in detecting subtle mediastinal abnormalities and controlling hallucinations.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

    Interpretationreasoning

    The phased approach supports potential SDG 17.

    “This phased approach aligns with established implementation science frameworks for health technology adoption and supports potential SDG 17 (Partnerships for the Goals) by enabling collaborative, multi-institutional rollout strategies, pending prospective clinical validation.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

    Interpretationnone given

    The generalised clinical verifier may lack the specialised knowledge required for rare pathologies or complex multi-system findings.

    “The generalised clinical verifier may lack the specialised knowledge required for rare pathologies or complex multi-system findings.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationreasoning

    This supports research reproducibility and transparency.

    “This supports research reproducibility and transparency, which are essential for scientific validation and regulatory evaluation.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data

    The high Cohen's Kappa agreement levels indicate systematic differences and support the reliability of the comparative evaluation.

    “These high agreement levels indicate systematic differences and support the reliability of the comparative evaluation.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationreasoning

    The framework remains well-suited for batch processing scenarios.

    “Real-time applications at the point of care may find this overhead limiting; however, the framework remains well-suited for batch processing scenarios (e.g., offline preliminary report generation for research purposes).”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based). May combine two claims.

    Interpretationnone given

    The framework advances research in AI-assisted radiology reporting with potential alignment with SDGs 3, 9, and 17.

    “This framework advances research in AI-assisted radiology reporting with potential alignment with SDGs 3, 9, and 17, pending prospective clinical validation.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationreasoning

    This design supports potential alignment with SDG 3.

    “This design supports potential alignment with SDG 3 (Good Health and Well-being) by making AI-assisted radiological research more accessible, pending further validation.”
    Quote in the paperNot a substantial claim
67 of 67

  • The paper relies on this without support.

    Automated report generation systems can ease the burden on radiologists by providing preliminary diagnostic descriptions that support clinical workflows without replacing expert judgment.

    “Automated report generation systems can ease this burden by providing preliminary diagnostic descriptions that support clinical workflows without replacing expert judgment.”

  • The paper relies on this without support.

    In decision-support contexts, false positives are more harmful than missed findings.

    “This pattern aligns with clinical priorities in decision-support contexts, where false positives are more harmful than missed findings.”

  • The paper relies on this without support.

    The full IU x-ray test set (321 samples) provides sufficient statistical power for the reported comparisons.

    “The evaluation was based on the full IU x-ray test set (321 samples), which provides sufficient statistical power for the reported comparisons.”
3 of 3

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

  • Abstract claims consistent improvements despite lower clinical and factuality scores

    Major

    The abstract calls the three-stage results consistent improvements over the ablations, but its own reported CheXpert-F1, Clinical-F1, factuality score, and hallucination rate are worse than both baselines.

  • METEOR significance claim conflicts with the one-stage comparison

    Major

    The discussion and conclusion attach p < 0.001 to the +4.5% METEOR gain over the one-stage baseline, but Table 3 reports p = 0.063 for that comparison. The p < 0.001 result applies to the two-stage comparison.

  • Training-set percentage does not match the reported counts

    Minor

    The reported 1,493 training samples are 69.96% of the 2,134 reports, not 75%. The three stated split percentages also total 105%.

    Reported 75 · recomputed 69.962512

  • Hallucination-rate standard deviation conflicts with its table entry

    Minor

    The discussion describes a hallucination-rate standard deviation of about 23%, but Table 2 prints zero dispersion for the three-stage hallucination rate.

4 of 4

Cites

Flagged first, then by citations

Authors
Abdaoui, H., Barbaria, S., Al Kuwaiti, A., Raouafi, N., Ben Salah, R., & Boussi Rahmouni, H.
Affiliations
  • Tunis El Manar University (TN)
  • Department of Dental Education, College of Dentistry, Deanship of Quality and Academic Accreditation, Imam Abdulrahman Bin Faisal University
  • Imam Abdulrahman Bin Faisal University (SA)
  • Laboratory of Biophysics and Medical Technologies, Higher Institute of Medical Technologies of Tunis (ISTMT), University of Tunis El Manar
Corresponding author
Barbaria, Sabri
ORCID
Venue
Frontiers in Radiology
Publisher
Frontiers Media SA
Published
2026-07-28
Volume
6
Pages
1875905-1875905
ISSN
2673-8740 (ISSN-L)
DOI
10.3389/fradi.2026.1875905
PMC
PMC13458744
OpenAlex
W7171535071
Semantic Scholar
2b447ab99e153b8eae054b1e8acb193ff0b63618
Cited by
0
References
32
FWCI
0.00
Type
journal article
Language
en
Open access
Licence
Locations
  • Frontiers in Radiology (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • PubMed Central (submittedVersion, cc-by)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion)
Topics
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Multimodal Machine Learning Applications
  • Health Sciences › Medicine › Health Informatics › Artificial Intelligence in Healthcare and Education
  • Health Sciences › Medicine › Radiology, Nuclear Medicine and Imaging › COVID-19 diagnosis using AI
Keywords
automated radiology report generation, vision-language models, Swin Transformer, Q-Former, hallucination mitigation, factuality, large language models, low-rank adaptation

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-07.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-07.

Cited sources are checked by gpt-6-sol.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review