arxiv: 2604.02543 · v1 · submitted 2026-04-02 · 💻 cs.CV · cs.LG

Recognition: 1 theorem link

· Lean Theorem

Overconfidence and Calibration in Medical VQA: Empirical Findings and Hallucination-Aware Mitigation

Ji Young Byun , Young-Jin Park , Jean-Philippe Corbeil , Asma Ben Abacha

Authors on Pith no claims yet

Pith reviewed 2026-05-13 21:21 UTC · model grok-4.3

classification 💻 cs.CV cs.LG

keywords medical VQAconfidence calibrationoverconfidencehallucination detectionvision-language modelsPlatt scalingAUROCpost-hoc calibration

0 comments

The pith

Post-hoc calibration such as Platt scaling reduces overconfidence in medical vision-language models more effectively than prompting, and hallucination signals further lift both calibration and AUROC.

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

Vision-language models for medical visual question answering remain overconfident regardless of model size or prompting techniques such as chain-of-thought. Simple post-hoc methods like Platt scaling lower calibration error and beat prompt-based approaches across three model families and multiple benchmarks. These monotonic adjustments leave the models' ranking ability, measured by AUROC, unchanged. Incorporating vision-grounded hallucination detection signals into calibration improves both error metrics and discrimination, with the largest benefits on open-ended questions. The work positions calibrated confidence estimates as essential for safe clinical deployment of these models.

Core claim

Overconfidence persists across model families, scales, and prompting strategies in medical VQA. Post-hoc calibration approaches such as Platt scaling reduce calibration error and outperform prompt-based methods, but their strict monotonicity prevents AUROC improvement. Hallucination-aware calibration that adds vision-grounded hallucination signals as complementary inputs improves both calibration and AUROC, with the biggest gains on open-ended questions.

What carries the argument

Hallucination-aware calibration (HAC), which refines confidence estimates by treating vision-grounded hallucination detection signals as additional inputs.

If this is right

Post-hoc calibration becomes standard practice for medical VLM deployment instead of raw confidence estimates.
Hallucination signals enable more reliable use of VLMs in medical VQA by lifting both calibration and discrimination.
Prompting strategies alone, including chain-of-thought and verbalized confidence, leave overconfidence unresolved.
Open-ended questions show the largest calibration and AUROC gains from hallucination-aware adjustments.

Where Pith is reading between the lines

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

Deploying these models in hospitals would benefit from routine post-hoc recalibration on site-specific data before clinical use.
The same hallucination signals could be tested as features in downstream tasks such as uncertainty-guided referral to human experts.
Combining HAC with non-monotonic recalibration techniques might further raise AUROC while preserving calibration gains.

Load-bearing premise

The hallucination detection signals are reliable, vision-grounded, and supply independent value without introducing bias.

What would settle it

A new medical VQA benchmark where applying hallucination-aware calibration produces no reduction in calibration error and no AUROC increase compared with raw or Platt-scaled outputs.

Figures

Figures reproduced from arXiv: 2604.02543 by Asma Ben Abacha, Jean-Philippe Corbeil, Ji Young Byun, Young-Jin Park.

**Figure 2.** Figure 2: ACE across sampling-based and verbalized confidence extraction methods and [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

**Figure 3.** Figure 3: Calibration errors (ECE and ACE) before and after post-hoc calibration (Platt [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

**Figure 4.** Figure 4: Sample size ablation for 2B models. ECE (↓), ACE (↓), and AUROC (↑) as a function of N′ , averaged across question types. Shaded bands show std over 1,000 simulations. 22 [PITH_FULL_IMAGE:figures/full_fig_p022_4.png] view at source ↗

**Figure 5.** Figure 5: Sample size ablation for 7/8B models. Same setup as [PITH_FULL_IMAGE:figures/full_fig_p023_5.png] view at source ↗

**Figure 6.** Figure 6: Sample size ablation for 30B+ models. Same setup as [PITH_FULL_IMAGE:figures/full_fig_p023_6.png] view at source ↗

**Figure 7.** Figure 7: ECE across different confidence extraction methods, including sampling and [PITH_FULL_IMAGE:figures/full_fig_p024_7.png] view at source ↗

**Figure 8.** Figure 8: AUROC across different confidence extraction methods, including sampling and [PITH_FULL_IMAGE:figures/full_fig_p024_8.png] view at source ↗

**Figure 9.** Figure 9: Ablation on hallucination detection metrics used in HAC-Platt. The pooled dataset [PITH_FULL_IMAGE:figures/full_fig_p034_9.png] view at source ↗

**Figure 10.** Figure 10: Cross-dataset ACE transfer for HAC-Platt. Each cell shows the ACE when [PITH_FULL_IMAGE:figures/full_fig_p035_10.png] view at source ↗

**Figure 11.** Figure 11: Cross-dataset ECE transfer for HAC-Platt. Same layout as Figure [PITH_FULL_IMAGE:figures/full_fig_p036_11.png] view at source ↗

**Figure 12.** Figure 12: Cross-dataset AUROC transfer for HAC-Platt. Same layout as Figure [PITH_FULL_IMAGE:figures/full_fig_p037_12.png] view at source ↗

read the original abstract

As vision-language models (VLMs) are increasingly deployed in clinical decision support, more than accuracy is required: knowing when to trust their predictions is equally critical. Yet, a comprehensive and systematic investigation into the overconfidence of these models remains notably scarce in the medical domain. We address this gap through a comprehensive empirical study of confidence calibration in VLMs, spanning three model families (Qwen3-VL, InternVL3, LLaVA-NeXT), three model scales (2B--38B), and multiple confidence estimation prompting strategies, across three medical visual question answering (VQA) benchmarks. Our study yields three key findings: First, overconfidence persists across model families and is not resolved by scaling or prompting, such as chain-of-thought and verbalized confidence variants. Second, simple post-hoc calibration approaches, such as Platt scaling, reduce calibration error and consistently outperform the prompt-based strategy. Third, due to their (strict) monotonicity, these post-hoc calibration methods are inherently limited in improving the discriminative quality of predictions, leaving AUROC at the same level. Motivated by these findings, we investigate hallucination-aware calibration (HAC), which incorporates vision-grounded hallucination detection signals as complementary inputs to refine confidence estimates. We find that leveraging these hallucination signals improves both calibration and AUROC, with the largest gains on open-ended questions. Overall, our findings suggest post-hoc calibration as standard practice for medical VLM deployment over raw confidence estimates, and highlight the practical usefulness of hallucination signals to enable more reliable use of VLMs in medical VQA.

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.

Desk Editor's Note private letter to a colleague

Post-hoc calibration like Platt scaling beats prompting for medical VLMs, and HAC adds AUROC lift on open-ended questions, but the hallucination signals need to prove they stay independent at deployment time.

read the letter

The paper's core message is straightforward: overconfidence sticks around in medical VLMs no matter the size or prompting tricks, simple post-hoc calibration cuts the error reliably, and their hallucination-aware calibration (HAC) can improve both calibration and AUROC by folding in vision-grounded signals. That last part is the main addition worth noting, especially the bigger gains on open-ended questions where raw confidence is weakest.

Referee Report

2 major / 2 minor

Summary. The paper presents a comprehensive empirical study of overconfidence and calibration in vision-language models (VLMs) for medical visual question answering, evaluating three model families (Qwen3-VL, InternVL3, LLaVA-NeXT) across scales from 2B to 38B and multiple prompting strategies on three benchmarks. It reports that overconfidence persists despite scaling and prompting, that post-hoc methods such as Platt scaling reduce calibration error more effectively than prompting but cannot improve AUROC due to monotonicity, and that a proposed hallucination-aware calibration (HAC) method leveraging vision-grounded hallucination signals improves both calibration and AUROC, with largest gains on open-ended questions.

Significance. The broad empirical scope across models, scales, and benchmarks provides useful practical guidance for VLM deployment in clinical settings. If the HAC signals are shown to be computable at inference time without label leakage, the work could establish post-hoc calibration as standard practice and demonstrate the additive value of hallucination detection for improving both calibration and discriminative quality.

major comments (2)

[HAC method description] HAC method description: the manuscript must explicitly detail the inference-time procedure for computing vision-grounded hallucination signals (e.g., cross-attention or consistency checks) and confirm that no test-benchmark answer labels were used for tuning or evaluation, as any such dependence would render the reported AUROC gains non-generalizable to deployment.
[Experimental results section] Experimental results section: the claim that post-hoc methods 'consistently outperform' prompting requires reporting of exact calibration-error values, AUROC numbers, and statistical significance tests across all model-benchmark combinations; without these, the comparative findings remain difficult to assess.

minor comments (2)

[Abstract] The abstract should specify the exact calibration metrics (e.g., ECE) and the three benchmarks by name.
[Tables] Tables comparing methods should include standard deviations or confidence intervals over multiple runs to convey variability.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. We address each major comment point by point below and describe the revisions that will be incorporated into the next version of the manuscript.

read point-by-point responses

Referee: [HAC method description] HAC method description: the manuscript must explicitly detail the inference-time procedure for computing vision-grounded hallucination signals (e.g., cross-attention or consistency checks) and confirm that no test-benchmark answer labels were used for tuning or evaluation, as any such dependence would render the reported AUROC gains non-generalizable to deployment.

Authors: We appreciate the referee's emphasis on transparency regarding the HAC procedure. In the revised manuscript we will add a dedicated subsection (Section 4.3) that explicitly describes the inference-time computation: vision-grounded hallucination signals are extracted from (i) cross-attention maps between image patch tokens and generated text tokens and (ii) consistency scores obtained by running the model on two lightly perturbed versions of the same image (random crop + color jitter) and measuring token-level disagreement. All operations use only the model's internal activations and the input image at inference time; no ground-truth answers from any test benchmark are accessed during signal extraction, hyper-parameter selection, or evaluation. We will also include pseudocode and confirm that the reported AUROC improvements therefore remain valid for real-world deployment without label leakage. revision: yes
Referee: [Experimental results section] Experimental results section: the claim that post-hoc methods 'consistently outperform' prompting requires reporting of exact calibration-error values, AUROC numbers, and statistical significance tests across all model-benchmark combinations; without these, the comparative findings remain difficult to assess.

Authors: We agree that exact numerical reporting and statistical tests are necessary for rigorous comparison. In the revised manuscript we will replace the current summary statements with two new tables (Table 3 and Table 4) that list, for every model-scale-benchmark triplet: (a) exact Expected Calibration Error (ECE) and Brier score for raw, prompt-based, Platt-scaled, and HAC confidences; (b) AUROC values; and (c) p-values from paired Wilcoxon signed-rank tests comparing post-hoc methods against prompting. These tables will be accompanied by a short statistical-methods paragraph. The added numbers will directly support the claim that post-hoc calibration consistently outperforms prompting while also documenting the further gains from HAC. revision: yes

Circularity Check

0 steps flagged

No circularity: purely empirical application of standard post-hoc methods

full rationale

The paper reports experimental results on VLM confidence calibration across models and benchmarks. It applies established techniques such as Platt scaling (fitted on held-out data) and augments them with hallucination detection signals as additional inputs. No equations, derivations, or predictions are presented that reduce to self-definition, fitted inputs renamed as outputs, or load-bearing self-citations. The central claims rest on observed performance differences rather than any closed logical loop.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The central claims rest on standard ML evaluation assumptions and the new HAC construct; no free parameters are introduced in a fitting sense, and the invented entity is the HAC method itself.

axioms (1)

domain assumption Medical VQA benchmarks are representative of clinical scenarios and standard metrics (ECE, AUROC) validly measure calibration and discrimination.
Invoked throughout the empirical study design and result interpretation.

invented entities (1)

Hallucination-aware calibration (HAC) no independent evidence
purpose: Refine confidence estimates by incorporating vision-grounded hallucination detection signals as additional inputs.
Introduced as a novel mitigation strategy motivated by the empirical findings.

pith-pipeline@v0.9.0 · 5605 in / 1202 out tokens · 61640 ms · 2026-05-13T21:21:03.766240+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

simple post-hoc calibration approaches, such as Platt scaling, reduce calibration error... HAC, which incorporates vision-grounded hallucination detection signals as complementary inputs to refine confidence estimates

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

36 extracted references · 36 canonical work pages · 6 internal anchors

[1]

Qwen-VL: A Versatile Vision-Language Model for Understanding, Localization, Text Reading, and Beyond

Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond.arXiv preprint arXiv:2308.12966,

work page internal anchor Pith review Pith/arXiv arXiv
[2]

InWorking Notes of CLEF 2019, volume 2380 ofCEUR Workshop Proceedings, Lugano, Switzerland, September 9-12

work page 2019
[3]

Assessing gpt-4 multimodal performance in radiological image analysis

Dana Brin, Vera Sorin, Yiftach Barash, Eli Konen, Benjamin S Glicksberg, Girish N Nadkarni, and Eyal Klang. Assessing gpt-4 multimodal performance in radiological image analysis. European Radiology, 35(4):1959–1965,

work page 1959
[4]

arXiv:2405.19538 (2024)

Pierre Chambon, Jean-Benoit Delbrouck, Thomas Sounack, Shih-Cheng Huang, Zhihong Chen, Maya Varma, Steven QH Truong, Chu The Chuong, and Curtis P Langlotz. Chex- pert plus: Augmenting a large chest x-ray dataset with text radiology reports, patient demographics and additional image formats.arXiv preprint arXiv:2405.19538,

work page arXiv
[5]

Mind the confidence gap: Overconfidence, cal- ibration, and distractor effects in large language models

Prateek Chhikara. Mind the confidence gap: Overconfidence, calibration, and distractor effects in large language models.arXiv preprint arXiv:2502.11028,

work page arXiv
[6]

Calibration of pre-trained transformers

Shrey Desai and Greg Durrett. Calibration of pre-trained transformers. InProceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pp. 295–302. Association for Computational Linguistics,

work page 2020
[7]

arXiv preprint arXiv:2508.15260 , year=

Yichao Fu, Xuewei Wang, Yuandong Tian, and Jiawei Zhao. Deep think with confidence. arXiv preprint arXiv:2508.15260,

work page arXiv
[8]

Hedge: Hallucination estima- tion via dense geometric entropy for vqa with vision-language models.arXiv preprint arXiv:2511.12693,

Sushant Gautam, Michael A Riegler, and P ˚al Halvorsen. Hedge: Hallucination estima- tion via dense geometric entropy for vqa with vision-language models.arXiv preprint arXiv:2511.12693,

work page arXiv
[9]

A survey of confidence estimation and calibration in large language models

Jiahui Geng, Fengyu Cai, Yuxia Wang, Heinz Koeppl, Preslav Nakov, and Iryna Gurevych. A survey of confidence estimation and calibration in large language models. InProceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pp. 6577–6595,

work page 2024
[10]

Overconfidence is key: Verbalized uncertainty evaluation in large language and vision-language models

Tobias Groot and Matias Valdenegro-Toro. Overconfidence is key: Verbalized uncertainty evaluation in large language and vision-language models. InProceedings of the 4th Workshop on Trustworthy Natural Language Processing (TrustNLP 2024), pp. 145–171,

work page 2024
[11]

DeBERTa: Decoding-enhanced BERT with Disentangled Attention

Pengcheng He, Xiaodong Liu, Jianfeng Gao, and Weizhu Chen. Deberta: Decoding- enhanced bert with disentangled attention.arXiv preprint arXiv:2006.03654,

work page internal anchor Pith review Pith/arXiv arXiv 2006
[12]

Calibrating long-form generations from large language models

Yukun Huang, Yixin Liu, Raghuveer Thirukovalluru, Arman Cohan, and Bhuwan Dhingra. Calibrating long-form generations from large language models. In Yaser Al-Onaizan, Mohit Bansal, and Yun-Nung Chen (eds.),Findings of the Association for Computational Linguistics: EMNLP 2024, pp. 13441–13460, Miami, Florida, USA, November

work page 2024
[13]

doi: 10.18653/v1/2024.findings-emnlp.785

Associ- ation for Computational Linguistics. doi: 10.18653/v1/2024.findings-emnlp.785. URL https://aclanthology.org/2024.findings-emnlp.785/. Jeremy Irvin, Pranav Rajpurkar, Michael Ko, Yifan Yu, Silviana Ciurea-Ilcus, Chris Chute, Henrik Marklund, Behzad Haghgoo, Robyn Ball, Katie Shpanskaya, et al. Chexpert: A large chest radiograph dataset with uncerta...

work page doi:10.18653/v1/2024.findings-emnlp.785 2024
[14]

Language Models (Mostly) Know What They Know

Saurav Kadavath, Tom Conerly, Amanda Askell, Tom Henighan, Dawn Drain, Ethan Perez, Nicholas Schiefer, Zac Hatfield-Dodds, Nova DasSarma, Eli Tran-Johnson, et al. Language models (mostly) know what they know.arXiv preprint arXiv:2207.05221,

work page internal anchor Pith review Pith/arXiv arXiv
[15]

LLaVA-NeXT-Interleave: Tackling Multi-image, Video, and 3D in Large Multimodal Models

Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava-next-interleave: Tackling multi-image, video, and 3d in large multimodal models. arXiv preprint arXiv:2407.07895,

work page internal anchor Pith review Pith/arXiv arXiv
[16]

Generating with confidence: Uncertainty quantification for black-box large language models.arXiv preprint arXiv:2305.19187,

12 Preprint. Under review. Zhen Lin, Shubhendu Trivedi, and Jimeng Sun. Generating with confidence: Uncertainty quantification for black-box large language models.arXiv preprint arXiv:2305.19187,

work page arXiv
[17]

Xiaoou Liu, Tiejin Chen, Longchao Da, Chacha Chen, Zhen Lin, and Hua Wei

URL https://llava-vl.github.io/blog/2024-01-30-llava-next/. Xiaoou Liu, Tiejin Chen, Longchao Da, Chacha Chen, Zhen Lin, and Hua Wei. Uncertainty quantification and confidence calibration in large language models: A survey. InProceed- ings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V . 2, pp. 6107–6117,

work page 2024
[18]

When do llms need retrieval augmen- tation? mitigating llms’ overconfidence helps retrieval augmentation

Shiyu Ni, Keping Bi, Jiafeng Guo, and Xueqi Cheng. When do llms need retrieval augmen- tation? mitigating llms’ overconfidence helps retrieval augmentation. InFindings of the Association for Computational Linguistics: ACL 2024, pp. 11375–11388,

work page 2024
[19]

Know what you don’t know: Uncertainty calibration of process reward models.arXiv preprint arXiv:2506.09338,

Young-Jin Park, Kristjan Greenewald, Kaveh Alim, Hao Wang, and Navid Azizan. Know what you don’t know: Uncertainty calibration of process reward models.arXiv preprint arXiv:2506.09338,

work page arXiv
[20]

Large language model uncer- tainty measurement and calibration for medical diagnosis and treatment.medRxiv, pp

Thomas Savage, John Wang, Robert Gallo, Abdessalem Boukil, Vishwesh Patel, Seyed Amir Ahmad Safavi-Naini, Ali Soroush, and Jonathan H Chen. Large language model uncer- tainty measurement and calibration for medical diagnosis and treatment.medRxiv, pp. 2024–06,

work page 2024
[21]

Aviv Slobodkin, Omer Goldman, Avi Caciularu, Ido Dagan, and Shauli Ravfogel

Maohao Shen, Subhro Das, Kristjan Greenewald, Prasanna Sattigeri, Gregory Wornell, and Soumya Ghosh. Thermometer: Towards universal calibration for large language models. arXiv preprint arXiv:2403.08819,

work page arXiv
[22]

Qwen3 Technical Report

URLhttps://arxiv.org/abs/2505.09388. 13 Preprint. Under review. Alberto Testoni and Iacer Calixto. Mind the gap: Benchmarking LLM uncertainty and calibration with specialty-aware clinical QA and reasoning-based behavioural features. In Vera Demberg, Kentaro Inui, and Llu´ıs Marquez (eds.),Proceedings of the 19th Conference of the European Chapter of the A...

work page internal anchor Pith review Pith/arXiv arXiv
[23]

ISBN 979-8-89176-380-7

Association for Computational Linguistics. ISBN 979-8-89176-380-7. doi: 10.18653/v1/2026.eacl-long.106. URL https: //aclanthology.org/2026.eacl-long.106/. Katherine Tian, Eric Mitchell, Allan Zhou, Archit Sharma, Rafael Rafailov, Huaxiu Yao, Chelsea Finn, and Christopher D Manning. Just ask for calibration: Strategies for eliciting calibrated confidence s...

work page doi:10.18653/v1/2026.eacl-long.106 2026
[24]

Towards generalist biomedical ai.Nejm Ai, 1(3):AIoa2300138, 2024a

Tao Tu, Shekoofeh Azizi, Danny Driess, Mike Schaekermann, Mohamed Amin, Pi-Chuan Chang, Andrew Carroll, Charles Lau, Ryutaro Tanno, Ira Ktena, et al. Towards generalist biomedical ai.Nejm Ai, 1(3):AIoa2300138, 2024a. Weijie Tu, Weijian Deng, Dylan Campbell, Stephen Gould, and Tom Gedeon. An empirical study into what matters for calibrating vision-language...

work page 2024
[25]

Self-Consistency Improves Chain of Thought Reasoning in Language Models

Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. Self-consistency improves chain of thought reasoning in language models.arXiv preprint arXiv:2203.11171,

work page internal anchor Pith review Pith/arXiv arXiv
[26]

Seeing is believing, but how much? a comprehensive analysis of verbalized calibration in vision- language models

Weihao Xuan, Qingcheng Zeng, Heli Qi, Junjue Wang, and Naoto Yokoya. Seeing is believing, but how much? a comprehensive analysis of verbalized calibration in vision- language models. In Christos Christodoulopoulos, Tanmoy Chakraborty, Carolyn Rose, and Violet Peng (eds.),Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processin...

work page 2025
[27]

emnlp-main.184/

Association for Computational Linguistics. ISBN 979-8-89176-332-6. doi: 10.18653/v1/2025.emnlp-main

work page doi:10.18653/v1/2025.emnlp-main 2025
[28]

Yasin Abbasi Yadkori, Ilja Kuzborskij, Andr´as Gy¨orgy, and Csaba Szepesv´ari

URLhttps://aclanthology.org/2025.emnlp-main.74/. Yasin Abbasi Yadkori, Ilja Kuzborskij, Andr´as Gy¨orgy, and Csaba Szepesv´ari. To believe or not to believe your llm.arXiv preprint arXiv:2406.02543,

work page arXiv 2025
[29]

Obtaining calibrated probability estimates from decision trees and naive bayesian classifiers

Bianca Zadrozny and Charles Elkan. Obtaining calibrated probability estimates from decision trees and naive bayesian classifiers. InIcml, volume 1, pp. 2001,

work page 2001
[30]

Radflag: A black-box hallucination detection method for medical vision language models.arXiv preprint arXiv:2411.00299, 2024a

Serena Zhang, Sraavya Sambara, Oishi Banerjee, Julian Acosta, L John Fahrner, and Pranav Rajpurkar. Radflag: A black-box hallucination detection method for medical vision language models.arXiv preprint arXiv:2411.00299, 2024a. Sheng Zhang, Yanbo Xu, Naoto Usuyama, Hanwen Xu, Jaspreet Bagga, Robert Tinn, Sam Preston, Rajesh Rao, Mu Wei, Naveen Valluri, et ...

work page arXiv
[31]

Under review

15 Preprint. Under review. A Further Related Work: Calibration in Deep Learning and Generative Models. Modern deep neural networks are often observed to be miscalibrated: while achieving high accuracy, they tend to produce overconfident predictions. Guo et al. (2017) provided the characterization of this phenomenon and introducedtemperature scalingas a po...

work page 2017
[32]

Which organ is abnormal, heart or lung?

Note that the SLAKE dataset provides its own CLOSED/OPEN labels, where CLOSED includes not only yes/no questions but alsoclosed-choicequestions such as “Which organ is abnormal, heart or lung?” Similarly, VQA-Med-2019 contains comparable closed-choice questions (e.g., “Is this image modality T1, T2, or FLAIR?”) without explicitly providing such labels. Me...

work page 2019
[33]

to perform semantic clustering, its NLI-based training on general-domain corpora may not generalize well to domain-specific equivalences; we instead use Qwen/Qwen3-4B-Instruct-2507, which demonstrates competitive performance on medical benchmarks (Team, 2025). Verbalized confidence.We evaluate six variants of verbalized confidence prompting strategy, each...

work page 2025
[34]

C.1.3 LLM-as-a-Judge Prompt We use an LLM judge to evaluate whether predicted answers are semantically equivalent to the ground truth

Term Probability Almost certain 0.95 Highly likely 0.90 Very good chance 0.85 Probable 0.75 Likely 0.70 Better than even 0.60 About even 0.50 Unlikely 0.30 Improbable 0.20 Very good chance not 0.15 Highly unlikely 0.10 Almost certainly not 0.05 Table 6: Mapping from linguistic confidence terms to numerical probabilities. C.1.3 LLM-as-a-Judge Prompt We use...

work page 2025
[35]

We consistently observed systematic overconfidence across model families and datasets; scaling did not mitigate the overconfidence or miscali- bration

E.1 Calibration across Model Families and Varying Scales Tables 8–10 present the full results. We consistently observed systematic overconfidence across model families and datasets; scaling did not mitigate the overconfidence or miscali- bration. While accuracy generally increased with larger models, reliability metrics such as ECE, ACE, and AUROC did not...

work page arXiv 2019
[36]

For open-ended ques- tions, we reconfirmed that HAC consistently improved AUROC

F.1 Predictive Discrimination Gains Table 12 extends Table 2 to per-dataset and per-model breakdowns. For open-ended ques- tions, we reconfirmed that HAC consistently improved AUROC. For closed-ended questions, performance varied across datasets, with the largest gain on VQA-Med. Exploring a more robust HAC formulation that generalizes across question typ...

work page 1900