Pith. sign in

REVIEW 4 major objections 5 minor 2 cited by

Learning Conformal Abstention Policies for Adaptive Risk Management in Large Language and Vision-Language Models

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

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

Pith's one-line read The paper claims that conformal abstention thresholds can be learned by reinforcement learning while preserving the 90% coverage guarantee, letting LLM/VLM systems decide per query between a single answer, a set, or abstention.

desk verdict A useful idea undone by test-set reward fitting and a coverage metric that counts abstention as coverage; needs major revision. read the letter →

arxiv 2502.06884 v1 pith:IHLKM47Z submitted 2025-02-08 cs.LG cs.AI

Sina Tayebati , Divake Kumar , Nastaran Darabi , Dinithi Jayasuriya , Ranganath Krishnan , Amit Ranjan Trivedi This is my paper · ORCID
classification cs.LGcs.AI
keywords conformalpredictionabstentionpolicyreinforcementlearninguncertaintyquantificationlargelanguagemodelsvision-languagehallucinationdetectionselectivegeneration
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper claims that the static thresholds of conformal prediction—the machinery that turns a model's confidence scores into a statistically guaranteed prediction set—can be replaced by learned thresholds without losing the guarantee. The proposed framework, CAP, uses reinforcement learning to pick two thresholds that divide each query into three outcomes: a single confident answer, a small set of plausible answers, or abstention. Across ten benchmarks and models from 2.7B to 34B parameters, the authors report that CAP holds the 90% coverage target while improving accuracy, hallucination detection, selective generation, and calibration relative to the LAC and APS baselines. If the claim holds, safety-critical LLM/VLM deployments could tune how often a system commits, hedges, or declines—while still bounding the rate at which the true answer is missed.

What carries the argument

The load-bearing object is the two-threshold conformal abstention partition. A single nonconformity score, $s(x) = 1 - \max_i p_i(x)$, is cut by two calibration quantiles: $ \hat{q}_{\text{predict}}$ from the $(1- \alpha)$ level and $ \hat{q}_{\text{abstain}}$ from the $(1- \beta)$ level, producing three regimes—single prediction, set prediction, abstention. A stochastic sigmoid policy (with scaling constant $c$) converts each score into probabilities over these three actions, and the α, β behind the thresholds are updated by REINFORCE using the reward $R( \alpha, \beta) = -C( \alpha, \beta)$, where $C$ trades off accuracy, average set size, abstention rate, coverage, and action-type diversity. The quantile construction is what carries the paper's 90% coverage claim; the RL loop is what adapts it to each task and model.

What would settle it

Freeze the trained α and β, run the policy on a fresh test set that never entered the reward computation, and measure marginal coverage over repeated splits; if the true-label inclusion rate systematically falls below 90%, the guarantee was an artifact of in-sample threshold tuning rather than a property of the two-threshold construction.

Watch

Extended reading notes

Core claim

The paper's central claim is that conformal abstention thresholds can be learned rather than fixed: CAP treats the two confidence levels α and β as actions sampled from a Gaussian policy and trains that policy with REINFORCE against a cost that balances accuracy, prediction-set size, abstention rate, coverage, and action diversity. Under the resulting two graduated thresholds, a test sample whose nonconformity score falls below $ \hat{q}_{\text{predict}}$ receives a single best prediction, a score between $ \hat{q}_{\text{predict}}$ and $ \hat{q}_{\text{abstain}}$ receives a set-valued prediction, and a score at or above $ \hat{q}_{\text{abstain}}$ triggers abstention. The authors claim this learned thresholding preserves the conformal coverage guarantee—the true label is included with probability at least $1- \alpha$, set to 90%—while outperforming LAC and APS: accuracy up to 3.2% higher, hallucination-detection AUROC up to 22.19% higher, selective-generation AUARC up to 21.17% higher, and expected calibration error reduced by 70%–85%, across ten datasets and multiple LLM/VLM families.

Load-bearing premise

The 90% coverage guarantee holds because the standard conformal rank argument is assumed to survive two extensions: adding a second abstention threshold, and tuning both thresholds with rewards computed on the test set—an assumption the appendix states but does not prove.

Editorial extensions

If this is right

  • Deployed LLM/VLM systems can fix a risk budget (e.g., 10% miss rate) and still adapt how often they answer outright, hedge with a set, or abstain, because thresholds move with the task and data distribution.
  • The same confidence scores that drive abstention double as hallucination detectors: the paper reports AUROC gains up to 22.19% over static conformal baselines.
  • Selective generation improves by up to 21.17% in AUARC, meaning systems reject more hard examples while retaining more correct answers on the examples they keep.
  • Expected calibration error drops by roughly 70–85% relative to LAC and APS, so the reported confidence of a prediction becomes a more honest estimate of its chance of being right.
  • The balance CAP strikes between set size and coverage avoids both failure modes of static methods: LAC's under-coverage from overly narrow sets and APS's uninformatively large sets.

Reading between the lines

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

  • The coverage guarantee should be read with caution: the paper tunes α and β with rewards computed on the test set, so the reported 90% coverage is an in-sample property unless the authors prove the rank argument survives test-set-dependent threshold selection—which they assert but do not demonstrate.
  • A clean way to test the method's honesty would be to freeze the learned thresholds and evaluate on a separate held-out split; if coverage degrades, the improvement over baselines is partly a selection effect.
  • The recipe generalizes beyond abstention: any conformal hyperparameter (score function, calibration skew, set-size penalty) could in principle be RL-tuned, but each such tuning would need its own coverage-preservation proof rather than an appeal to the classic exchangeability argument.
  • Because the policy is trained per model and per task, the method's practical cost is a tuning loop that must be rerun when the model or the data distribution changes; the paper's reported gains are for the tuned configuration.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes CAP, a reinforcement-learning framework that adaptively selects two conformal prediction thresholds (α and β) to produce single predictions, set predictions, or abstentions for LLM/VLM multiple-choice QA. A softmax-based nonconformity score and REINFORCE are used to tune the thresholds against a cost function combining accuracy, set size, abstention, coverage, and diversity. The method is evaluated on ten MCQA datasets across several LLMs and VLMs, with reported improvements over LAC and APS in AUROC, AUARC, accuracy, and ECE, and a claimed 90% coverage guarantee. Appendix A provides a standard single-threshold conformal coverage proof and asserts that the argument extends to the two-threshold RL setting. The code is publicly available.

Significance. If the central coverage guarantee and the empirical improvements held, CAP would be a practically useful contribution: it addresses a real limitation of static conformal thresholds, evaluates on a broad set of models and benchmarks, and ships code. The comparison against LAC and APS is a sensible baseline choice, and the paper is generally clearly written. However, the load-bearing claims are not supported. The reported coverage metric is defined to count abstentions as covered, which does not match the conformal guarantee P(Y_t ∈ C(X_t)) stated in the abstract. The thresholds are selected using rewards computed on the test set, so the headline AUROC, AUARC, accuracy, and ECE numbers are, to a substantial degree, fitted objectives rather than independent predictions. The two-threshold extension of the conformal proof is asserted, not proven, and the paper's own limitation paragraph concedes that learned policies may 'distort CP's theoretical guarantees.' These issues concern the central contribution, not peripheral details.

major comments (4)
  1. [Section IV-A and Section III] The coverage metric reported in Table II is not the conformal coverage guarantee claimed in the abstract and in Section IV-B. Section IV-A defines coverage as 'the fraction of instances where the correct label is included in the model's output' but immediately adds that 'in setups with abstention, it also accounts for instances where the model successfully avoids making an incorrect explicit guess,' while Section III defines coverage = 1 − abstention in the cost function. A system that abstains on 10% of examples and rarely includes the true label on the remaining 90% can still report 'coverage' above 90% under this definition. The conformal guarantee is P(Y_t ∈ C(X_t)), so abstentions cannot be counted as covered. Thus Table II and the statement 'CAP consistently achieves at least 90% coverage across all datasets' do not establish the claimed guarantee.
  2. [Section III and Algorithm 1 (Appendix B)] The RL policy is trained using rewards computed on the test set. The main text states that at each iteration the process involves 'evaluating performance on a test set, and using the performance-based cost as a reward signal to update πθ via REINFORCE,' and Algorithm 1 computes C(α, β) from accuracy, abstention, set size, coverage, and diversity evaluated on test samples. Consequently, the reported AUROC, AUARC, accuracy, ECE, and coverage values are in-sample metrics that directly drive the policy update. They are not independent evaluations of a fixed policy. A proper evaluation requires a held-out test set that is never used for reward computation, and the policy must be frozen before reporting final metrics.
  3. [Appendix A and Section IV-B] The claimed 90% coverage guarantee is not proven for the actual method. Appendix A proves a standard single-threshold coverage result under exchangeability, but the method uses two thresholds, stochastic action sampling, and thresholds learned by RL from test-set rewards. The appendix's statement that 'the rank-based argument holds equally under mild modifications when additional thresholds are introduced' is asserted without proof. Moreover, because the thresholds are selected by optimizing rewards computed on test labels, the exchangeability between calibration scores and test scores is broken, so standard conformal coverage need not hold. The Discussion section's own warning that learned policies 'may overfit, bias abstention strategies, or distort CP's theoretical guarantees' is a direct acknowledgment of this risk.
  4. [Section III] The action probabilities are not guaranteed to be valid probabilities. The paper defines psingle = σ(−c(s(x) − q_predict)), pabstain = σ(c(s(x) − q_abstain)), and pset = 1 − psingle − pabstain. Since psingle and pabstain are independent sigmoids, their sum can exceed 1, making pset negative. Additionally, α and β are sampled from unconstrained Gaussian distributions, so they can fall outside (0,1), producing invalid quantile levels and potentially reversing the ordering q_predict ≤ q_abstain. This undermines the stochastic decision procedure described in Section III and the threshold semantics used in the experiments.
minor comments (5)
  1. [Figures 1, 2, 6, 7 and Section IV-A] The method is called CAP, but several figure captions and one phrase in Section IV-A refer to 'ATCP'; the acronym should be made consistent.
  2. [Section IV-A and Appendix C-A] The dataset description says VLM datasets are 'standardized to four options,' while the prompting section and later standardization paragraphs describe six options (four content options plus 'I don't know' and 'None of the above'). Please clarify the final option count used in the reported experiments.
  3. [Section III and Algorithm 1] The cost function hyperparameters λ1–λ4, the sigmoid scaling constant c, the policy network architecture, the learning rate, and the number of episodes are not reported. These values are necessary for reproducibility and for assessing the sensitivity of the RL procedure.
  4. [Tables I–IV] No confidence intervals, standard errors, or multiple-seed results are reported for the RL-based method. Given the stochasticity of REINFORCE and the small reported margins in some accuracy comparisons, the empirical claims need variance estimates.
  5. [Appendix A] Theorem 1 uses the notation Quantile({s1,...,sn}, 1−α) without specifying the finite-sample correction ⌈(n+1)(1−α)⌉/n used in Section III; the two quantile definitions should be reconciled.

Circularity Check

2 steps flagged · score 7.0 of 10

Test-set reward fitting and a coverage metric that counts abstention as success make the headline accuracy/coverage claims fitted or definitional rather than predicted.

  1. fitted input called prediction [Section III, 'Reinforcement Learning and Abstention Policy' and 'Cost Function and Reward Design'; Algorithm 1 in Appendix B]
    "At each iteration, the process involves: (1) sampling α, β from the learned distribution, (2) computing the thresholds q̂predict, q̂abstain on a calibration set, (3) evaluating performance on a test set, and (4) using the performance-based cost as a reward signal to update πθ via REINFORCE. The cost function is defined as: C(α, β) = (1− acc) +λ1 avgSet +λ2 abstention − λ3 coverage − λ4 div."

    The policy is trained by maximizing R = −C on the test set, where C directly contains accuracy, average set size, abstention, and coverage. The abstract and Tables I–IV report accuracy, coverage, set size, and related metrics on that same test split, so these headline numbers are the optimized objective rather than independent predictions. The claim 'improving accuracy by up to 3.2%' and 'consistently meeting the 90% coverage target' states the outcome of fitting thresholds to the evaluation data. AUROC, AUARC, and ECE are not explicit reward terms, but they are measured after α and β are selected using test labels, so they inherit the same selection bias.

  2. self definitional [Section III, 'Cost Function and Reward Design'; Section IV-A, 'Coverage'; Section IV-B, 'Coverage Guarantee']
    "Additionally, coverage = 1− abstention, and div is an entropy-based term quantifying the balance among single predictions, set predictions, and abstentions. Coverage measures the fraction of instances where the correct label is included in the model’s output—either as a single prediction or within a prediction set. In setups with abstention, it also accounts for instances where the model successfully avoids making an incorrect explicit guess."

    The reward's coverage term is defined as 1 − abstention, and the evaluation metric counts abstention as a success ('avoids making an incorrect explicit guess'). Section IV-B claims the conformal guarantee that the true label is included with probability at least 1 − α = 0.9, but the reported 'coverage' is not P(Y_t ∈ C(X_t)). Because abstentions are counted as covered by definition, a policy that abstains on fewer than 10% of cases can report more than 90% 'coverage' while rarely including the true label. The central coverage guarantee is therefore definitionally disconnected from the metric used to demonstrate it.

full rationale

Two load-bearing reductions make the central claims circular or fitted. First, Algorithm 1 and Section III update the RL policy using rewards computed by evaluating on the test set, and the cost function explicitly contains accuracy, set size, abstention, and coverage. The abstract's headline accuracy, 90% coverage, and balanced set-size results are therefore the optimized objective, not independent predictions; no held-out protocol is described. Second, the coverage metric used for the guarantee is not the conformal probability asserted in Section IV-B: the cost defines coverage = 1 − abstention, and the evaluation section counts abstention as success, while the abstract and Section IV-B promise P(true label included) ≥ 0.9. A system that abstains on <10% of examples can satisfy the reported metric while never capturing the true label. Appendix A proves only single-threshold i.i.d. coverage and then asserts that the two-threshold RL extension 'holds equally under mild modifications' without proof; this is a missing-derivation gap rather than a circular reduction, but it removes the theoretical basis for the 90% claim. AUROC/AUARC/ECE gains are not direct reward terms, but they are measured on the same test set used to select α and β, so they inherit the fitting problem. Self-citations such as [13] and [40] are not load-bearing to these reductions. The score reflects that the central empirical and guarantee claims reduce, at least in part, to the fitted objective and to a definitionally different coverage metric.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The method rests on an unproven extension of conformal guarantees to RL-selected thresholds on the test set, a coverage proxy that differs from the reported metric, and standard exchangeability and softmax-score assumptions. Several hyperparameters (alpha, beta, lambda1-4, c) are free parameters with no reported values.

free parameters (5)
  • alpha (target coverage error) = not reported
    Sampled from a learned Gaussian distribution during RL and optimized on the test set; acts as the conformal coverage threshold.
  • beta (abstention threshold parameter) = not reported
    Sampled from a learned Gaussian distribution and optimized on the test set; controls the abstention threshold.
  • lambda1-4 (cost function weights) = not reported
    Weights in the cost function C(alpha, beta) = (1-acc) + lambda1 avgSet + lambda2 abstention - lambda3 coverage - lambda4 div; no values are given in the paper.
  • c (sigmoid scaling constant) = not reported
    Scaling constant in the sigmoid functions for psingle and pabstain; no value is given.
  • policy network parameters theta = not reported
    Neural network mapping state s to mean and standard deviation vectors for alpha and beta; fitted on test-set reward.
assumptions (5)
  • standard math Exchangeability or i.i.d. of calibration and test samples
    Invoked in Appendix A for the classic single-threshold conformal coverage proof.
  • ad hoc to paper The rank-based conformal argument extends to two-threshold RL-learned thresholds
    Appendix A asserts 'the rank-based argument holds equally under mild modifications' but no proof is given for the adaptive, test-set-optimized setting.
  • domain assumption Coverage = 1 - abstention in the cost function
    Section III defines coverage = 1 - abstention in C(alpha, beta), but the reported coverage in Table II is the fraction of instances where the true label is in the output, a different quantity.
  • domain assumption Softmax probabilities are suitable nonconformity scores
    Section III uses score(x) = 1 - max_i p_i(x) as the nonconformity measure; this is a heuristic inherited from LAC and APS.
  • domain assumption Action probabilities pset = 1 - psingle - pabstain are valid probabilities
    No constraint is placed on the sum psingle + pabstain, which can exceed 1 and make pset negative.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Learning Conformal Abstention Policies for Adaptive Risk Management in Large Language and Vision-Language Models." pith.science (2026). https://pith.science/paper/IHLKM47Z

@misc{pith2026250206884,
  author       = {Pith},
  title        = {Pith review of: Learning Conformal Abstention Policies for Adaptive Risk Management in Large Language and Vision-Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IHLKM47Z}},
  note         = {Machine review of arXiv:2502.06884}
}
read the original abstract

Large Language and Vision-Language Models (LLMs/VLMs) are increasingly used in safety-critical applications, yet their opaque decision-making complicates risk assessment and reliability. Uncertainty quantification (UQ) helps assess prediction confidence and enables abstention when uncertainty is high. Conformal prediction (CP), a leading UQ method, provides statistical guarantees but relies on static thresholds, which fail to adapt to task complexity and evolving data distributions, leading to suboptimal trade-offs in accuracy, coverage, and informativeness. To address this, we propose learnable conformal abstention, integrating reinforcement learning (RL) with CP to optimize abstention thresholds dynamically. By treating CP thresholds as adaptive actions, our approach balances multiple objectives, minimizing prediction set size while maintaining reliable coverage. Extensive evaluations across diverse LLM/VLM benchmarks show our method outperforms Least Ambiguous Classifiers (LAC) and Adaptive Prediction Sets (APS), improving accuracy by up to 3.2%, boosting AUROC for hallucination detection by 22.19%, enhancing uncertainty-guided selective generation (AUARC) by 21.17%, and reducing calibration error by 70%-85%. These improvements hold across multiple models and datasets while consistently meeting the 90% coverage target, establishing our approach as a more effective and flexible solution for reliable decision-making in safety-critical applications. The code is available at: {https://github.com/sinatayebati/vlm-uncertainty}.

Figures

Figures reproduced from arXiv: 2502.06884 by the authors.

Figure 1
Figure 1. Accuracy vs. Expected Calibration Error (ECE) comparison of CAP, APS, and LAC across various VLMs and five [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Accuracy versus Expected Calibration Error (ECE) comparison between ATCP, APS and LAC methods across different [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Performance comparison of CAP (Ours), APS, and LAC on Llava-v1.6-34B (VLM) and Yi-34B (LLM) across four [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (10 more)
Figure 4
Figure 4. Figure 4: Accuracy vs. Expected Calibration Error (ECE) comparison of CAP, APS, and LAC across various VLMs and five [PITH_FULL_IMAGE:figures/full_fig_p016_4.png]
Figure 5
Figure 5. Figure 5: Accuracy versus Expected Calibration Error (ECE) comparison between CAP, APS and LAC methods across different [PITH_FULL_IMAGE:figures/full_fig_p017_5.png]
Figure 6
Figure 6. Figure 6: Accuracy versus Expected Calibration Error (ECE) comparison between ATCP, APS and LAC methods across different [PITH_FULL_IMAGE:figures/full_fig_p018_6.png]
Figure 7
Figure 7. Figure 7: Accuracy versus Expected Calibration Error (ECE) comparison between ATCP, APS and LAC methods across different [PITH_FULL_IMAGE:figures/full_fig_p019_7.png]
Figure 8
Figure 8. Figure 8: Performance comparison of VLMs with different model sizes (2.7B to 34B) across various metrics. Figures from left [PITH_FULL_IMAGE:figures/full_fig_p019_8.png]
Figure 9
Figure 9. Figure 9: Performance comparison of additional VLMs with different model sizes (6B to 7B) across various metrics. Figures from [PITH_FULL_IMAGE:figures/full_fig_p020_9.png]
Figure 10
Figure 10. Figure 10: Performance comparison of LLMs with different model sizes (7B to 34B) across various metrics. Figures from left [PITH_FULL_IMAGE:figures/full_fig_p020_10.png]
Figure 11
Figure 11. Figure 11: Performance comparison of Llama-2 series LLMs with different model sizes (7B and 13B) across various metrics. [PITH_FULL_IMAGE:figures/full_fig_p020_11.png]
Figure 12
Figure 12. Figure 12: Performance comparison of Llama-2 series LLMs with different model sizes (7B and 13B) across various metrics. [PITH_FULL_IMAGE:figures/full_fig_p021_12.png]
Figure 13
Figure 13. Figure 13: Distribution of CAP’s prediction types for LLaVA-1.6-34B (VLM) and Yi-34B (LLM). The model’s responses are [PITH_FULL_IMAGE:figures/full_fig_p021_13.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Look Again Before You Abstain:Budgeted Conformal Evidence Acquisition for Reliable Vision-Language Model

    cs.CV 2026-06 conditional novelty 6.0 of 10

    Folding visual evidence acquisition into the conformal score and re-calibrating on post-acquisition scores preserves the hallucination-rate guarantee while recovering coverage in LVLM selective prediction.

  2. Operational Evidence Gaps for LLMs in Fraud Detection and Trust-and-Safety Workflows

    cs.CR 2026-07 conditional novelty 5.0 of 10

    A survey of 49 LLM fraud and trust-and-safety papers finds that fraud work reports almost no per-decision latency, cost, or calibration evidence, while moderation work reports more.

Reference graph

Works this paper leans on

47 extracted references · 21 canonical work pages · cited by 2 Pith papers

  1. [1]

    Vision language models in autonomous driving: A survey and outlook,

    X. Zhou, M. Liu, E. Yurtsever, B. L. Zagar, W. Zimmer, H. Cao, and A. C. Knoll, “Vision language models in autonomous driving: A survey and outlook,” IEEE Transactions on Intelligent Vehicles , 2024

    work page 2024

  2. [2]

    Multimodal healthcare ai: identifying and designing clinically relevant vision-language applications for radiology,

    N. Yildirim, H. Richardson, M. T. Wetscherek, J. Bajwa, J. Jacob, M. A. Pinnock, S. Harris, D. Coelho De Castro, S. Bannur, S. Hyland et al. , “Multimodal healthcare ai: identifying and designing clinically relevant vision-language applications for radiology,” in Proceedings of the CHI Conference on Human Factors in Computing Systems , 2024

    work page 2024

  3. [3]

    Securing large lan- guage models: Threats, vulnerabilities and responsible practices,

    S. Abdali, R. Anarfi, C. Barberan, and J. He, “Securing large lan- guage models: Threats, vulnerabilities and responsible practices,” arXiv preprint arXiv:2403.12503, 2024

    arXiv 2024

  4. [4]

    Mit- igating llm hallucinations via conformal abstention,

    Y . A. Yadkori, I. Kuzborskij, D. Stutz, A. Gy ¨orgy, A. Fisch, A. Doucet, I. Beloshapka, W.-H. Weng, Y .-Y . Yang, C. Szepesv ´ari et al. , “Mit- igating llm hallucinations via conformal abstention,” arXiv preprint arXiv:2405.01563, 2024

    arXiv 2024

  5. [5]

    Least ambiguous set-valued classifiers with bounded error levels,

    M. Sadinle, J. Lei, and L. Wasserman, “Least ambiguous set-valued classifiers with bounded error levels,”Journal of the American Statistical Association, vol. 114, no. 525, pp. 223–234, 2019

    2019

  6. [6]

    Classification with valid and adaptive coverage,

    Y . Romano, M. Sesia, and E. Candes, “Classification with valid and adaptive coverage,”Advances in Neural Information Processing Systems, vol. 33, pp. 3581–3591, 2020

    2020

  7. [7]

    V ovk, A

    V . V ovk, A. Gammerman, and G. Shafer, Algorithmic learning in a random world. Springer, 2005, vol. 29

    2005

  8. [8]

    Balasubramanian, S.-S

    V . Balasubramanian, S.-S. Ho, and V . V ovk, Conformal prediction for reliable machine learning: theory, adaptations and applications . Newnes, 2014

    work page 2014

Show all 47 references
  1. [9]

    A gentle introduction to confor- mal prediction and distribution-free uncertainty quantification,

    A. N. Angelopoulos and S. Bates, “A gentle introduction to confor- mal prediction and distribution-free uncertainty quantification,” arXiv preprint arXiv:2107.07511, 2021

    2021 arXiv

  2. [10]

    Dropout as a Bayesian approximation: Rep- resenting model uncertainty in deep learning,

    Y . Gal and Z. Ghahramani, “Dropout as a Bayesian approximation: Rep- resenting model uncertainty in deep learning,” International Conference on Machine Learning , pp. 1050–1059, 2016

    2016

  3. [11]

    Weight uncertainty in neural network,

    C. Blundell, J. Cornebise, K. Kavukcuoglu, and D. Wierstra, “Weight uncertainty in neural network,” in International conference on machine learning. PMLR, 2015

    2015

  4. [12]

    Simple and scalable predictive uncertainty estimation using deep ensembles,

    B. Lakshminarayanan, A. Pritzel, and C. Blundell, “Simple and scalable predictive uncertainty estimation using deep ensembles,” Advances in Neural Information Processing Systems , vol. 30, 2017

    2017

  5. [13]

    Improving model calibration with accu- racy versus uncertainty optimization,

    R. Krishnan and O. Tickoo, “Improving model calibration with accu- racy versus uncertainty optimization,” Advances in Neural Information Processing Systems, vol. 33, 2020

    2020

  6. [14]

    A survey of uncertainty in deep neural networks,

    J. Gawlikowski, C. R. N. Tassi, M. Ali, J. Lee, M. Humt, J. Feng, A. Kruspe, R. Triebel, P. Jung, R. Roscheret al., “A survey of uncertainty in deep neural networks,” Artificial Intelligence Review , vol. 56, no. Suppl 1, pp. 1513–1589, 2023

    2023

  7. [15]

    Verified uncertainty calibration,

    A. Kumar, P. S. Liang, and T. Ma, “Verified uncertainty calibration,” Advances in Neural Information Processing Systems , vol. 32, 2019

    2019

  8. [16]

    Benchmarking llms via uncertainty quantification,

    F. Ye, M. Yang, J. Pang, L. Wang, D. F. Wong, E. Yilmaz, S. Shi, and Z. Tu, “Benchmarking llms via uncertainty quantification,” arXiv preprint arXiv:2401.12794, 2024

    2024 arXiv

  9. [17]

    Sample efficient reinforcement learning with reinforce,

    J. Zhang, J. Kim, B. O’Donoghue, and S. Boyd, “Sample efficient reinforcement learning with reinforce,” in Proceedings of the AAAI conference on artificial intelligence , vol. 35, no. 12, 2021, pp. 10 887– 10 895

    2021

  10. [18]

    Mmbench: Is your multi-modal model an all-around player?

    Y . Liu, H. Duan, Y . Zhang, B. Li, S. Zhang, W. Zhao, Y . Yuan, J. Wang, C. He, Z. Liu et al., “Mmbench: Is your multi-modal model an all-around player?” in European conference on computer vision . Springer, 2025, pp. 216–233

    2025

  11. [19]

    Ood-cv: A benchmark for robustness to out- of-distribution shifts of individual nuisances in natural images,

    B. Zhao, S. Yu, W. Ma, M. Yu, S. Mei, A. Wang, J. He, A. Yuille, and A. Kortylewski, “Ood-cv: A benchmark for robustness to out- of-distribution shifts of individual nuisances in natural images,” in European conference on computer vision . Springer, 2022, pp. 163– 180

    2022

  12. [20]

    Learn to explain: Multimodal reasoning via thought chains for science question answering,

    P. Lu, S. Mishra, T. Xia, L. Qiu, K.-W. Chang, S.-C. Zhu, O. Tafjord, P. Clark, and A. Kalyan, “Learn to explain: Multimodal reasoning via thought chains for science question answering,” Advances in Neural Information Processing Systems , vol. 35, pp. 2507–2521, 2022

    2022

  13. [21]

    Seed-bench: Benchmarking multimodal llms with generative comprehension,

    B. Li, R. Wang, G. Wang, Y . Ge, Y . Ge, and Y . Shan, “Seed-bench: Benchmarking multimodal llms with generative comprehension,” arXiv preprint arXiv:2307.16125, 2023

    2023 arXiv

  14. [22]

    A diagram is worth a dozen images,

    A. Kembhavi, M. Salvato, E. Kolve, M. Seo, H. Hajishirzi, and A. Farhadi, “A diagram is worth a dozen images,” in Computer Vision– ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11–14, 2016, Proceedings, Part IV 14 . Springer, 2016, pp. 235–251

    2016

  15. [23]

    Measuring massive multitask language understanding,

    D. Hendrycks, C. Burns, S. Basart, A. Zou, M. Mazeika, D. Song, and J. Steinhardt, “Measuring massive multitask language understanding,” arXiv preprint arXiv:2009.03300 , 2020

    2009 arXiv

  16. [24]

    Cosmos qa: Ma- chine reading comprehension with contextual commonsense reasoning,

    L. Huang, R. L. Bras, C. Bhagavatula, and Y . Choi, “Cosmos qa: Ma- chine reading comprehension with contextual commonsense reasoning,” arXiv preprint arXiv:1909.00277 , 2019

    1909 arXiv

  17. [25]

    Hel- laswag: Can a machine really finish your sentence?

    R. Zellers, A. Holtzman, Y . Bisk, A. Farhadi, and Y . Choi, “Hel- laswag: Can a machine really finish your sentence?” arXiv preprint arXiv:1905.07830, 2019

    1905 arXiv

  18. [26]

    Halueval: A large-scale hallucination evaluation benchmark for large language models,

    J. Li, X. Cheng, W. X. Zhao, J.-Y . Nie, and J.-R. Wen, “Halueval: A large-scale hallucination evaluation benchmark for large language models,” arXiv preprint arXiv:2305.11747 , 2023

    2023 arXiv

  19. [27]

    Improved baselines with visual instruction tuning,

    H. Liu, C. Li, Y . Li, and Y . J. Lee, “Improved baselines with visual instruction tuning,” 2023

    2023

  20. [28]

    Moe-llava: Mixture of experts for large vision-language models,

    B. Lin, Z. Tang, Y . Ye, J. Cui, B. Zhu, P. Jin, J. Zhang, M. Ning, and L. Yuan, “Moe-llava: Mixture of experts for large vision-language models,” arXiv preprint arXiv:2401.15947 , 2024

    2024 arXiv

  21. [29]

    Monkey: Image resolution and text label are important things for large multi-modal models,

    Z. Li, B. Yang, Q. Liu, Z. Ma, S. Zhang, J. Yang, Y . Sun, Y . Liu, and X. Bai, “Monkey: Image resolution and text label are important things for large multi-modal models,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2024, pp. 26 763–26 773

    2024

  22. [30]

    Internlm-xcomposer2: Mastering free-form text-image composition and comprehension in vision-language large model,

    X. Dong, P. Zhang, Y . Zang, Y . Cao, B. Wang, L. Ouyang, X. Wei, S. Zhang, H. Duan, M. Cao et al. , “Internlm-xcomposer2: Mastering free-form text-image composition and comprehension in vision-language large model,” arXiv preprint arXiv:2401.16420 , 2024

    2024 arXiv

  23. [31]

    Yi: Open foundation models by 01. ai,

    A. Young, B. Chen, C. Li, C. Huang, G. Zhang, G. Zhang, H. Li, J. Zhu, J. Chen, J. Chang et al., “Yi: Open foundation models by 01. ai,” arXiv preprint arXiv:2403.04652, 2024

    2024 arXiv

  24. [32]

    Cogagent: A visual language model for gui agents,

    W. Hong, W. Wang, Q. Lv, J. Xu, W. Yu, J. Ji, Y . Wang, Z. Wang, Y . Dong, M. Ding et al. , “Cogagent: A visual language model for gui agents,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2024, pp. 14 281–14 290

    2024

  25. [33]

    Qwen-vl: A versatile vision-language model for un- derstanding, localization, text reading, and beyond,

    J. Bai, S. Bai, S. Yang, S. Wang, S. Tan, P. Wang, J. Lin, C. Zhou, and J. Zhou, “Qwen-vl: A versatile vision-language model for un- derstanding, localization, text reading, and beyond,” arXiv preprint arXiv:2308.12966, vol. 1, no. 2, p. 3, 2023

    2023 arXiv

  26. [34]

    Uncertainty- aware evaluation for vision-language models,

    V . Kostumov, B. Nutfullin, O. Pilipenko, and E. Ilyushin, “Uncertainty- aware evaluation for vision-language models,” arXiv preprint arXiv:2402.14418, 2024

    2024 arXiv

  27. [35]

    The relationship between precision-recall and roc curves,

    J. Davis and M. Goadrich, “The relationship between precision-recall and roc curves,” in Proceedings of the 23rd international conference on Machine learning, 2006, pp. 233–240

    2006

  28. [36]

    Obtaining well calibrated probabilities using bayesian binning,

    M. P. Naeini, G. Cooper, and M. Hauskrecht, “Obtaining well calibrated probabilities using bayesian binning,” in Proceedings of the AAAI conference on artificial intelligence , vol. 29, 2015

    2015

  29. [37]

    Visual instruction tuning,

    H. Liu, C. Li, Q. Wu, and Y . J. Lee, “Visual instruction tuning,” Advances in neural information processing systems , vol. 36, 2024

    2024

  30. [38]

    Judging llm-as-a-judge with mt-bench and chatbot arena,

    L. Zheng, W.-L. Chiang, Y . Sheng, S. Zhuang, Z. Wu, Y . Zhuang, Z. Lin, Z. Li, D. Li, E. Xing et al. , “Judging llm-as-a-judge with mt-bench and chatbot arena,” Advances in Neural Information Processing Systems, vol. 36, pp. 46 595–46 623, 2023

    2023

  31. [39]

    Detecting hallucinations in large language models using semantic entropy,

    S. Farquhar, J. Kossen, L. Kuhn, and Y . Gal, “Detecting hallucinations in large language models using semantic entropy,” Nature, vol. 630, no. 8017, pp. 625–630, 2024

    2024

  32. [40]

    Enhancing trust in large language models with uncertainty-aware fine-tuning,

    R. Krishnan, P. Khanna, and O. Tickoo, “Enhancing trust in large language models with uncertainty-aware fine-tuning,” arXiv preprint arXiv:2412.02904, 2024

    2024 arXiv

  33. [41]

    Mobilevlm v2: Faster and stronger baseline for vision language model,

    X. Chu, L. Qiao, X. Zhang, S. Xu, F. Wei, Y . Yang, X. Sun, Y . Hu, X. Lin, B. Zhang et al. , “Mobilevlm v2: Faster and stronger baseline for vision language model,” arXiv preprint arXiv:2402.03766 , 2024

    2024 arXiv

  34. [42]

    mplug-owl2: Revolutionizing multi-modal large language model with modality collaboration,

    Q. Ye, H. Xu, J. Ye, M. Yan, A. Hu, H. Liu, Q. Qian, J. Zhang, and F. Huang, “mplug-owl2: Revolutionizing multi-modal large language model with modality collaboration,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2024, pp. 13 040–13 051

    2024

  35. [43]

    Llama 2: Open foundation and fine-tuned chat models,

    H. Touvron, L. Martin, K. Stone, P. Albert, A. Almahairi, Y . Babaei, N. Bashlykov, S. Batra, P. Bhargava, S. Bhosale et al. , “Llama 2: Open foundation and fine-tuned chat models,” arXiv preprint arXiv:2307.09288, 2023. APPENDIX A FORMAL PROOF OF CONFORMAL COVERAGE GUARANTEE ...

    2023 arXiv

  36. [44]

    AUROC vs AUARC, coverage rate vs set size, and accuracy vs ex- pected calibration error respectively

    Results of VLMs Additional results in Table VII, Table VIII, and Table IX demonstrate the performance of multiple VLMs mentioned in Appendix C in terms of uncertainty quantification i.e. AUROC vs AUARC, coverage rate vs set size, and accuracy vs ex- pected calibration error re...

  37. [45]

    AUROC vs AUARC, coverage rate vs set size, and accuracy vs expected calibration error respectively

    Results of LLMs Additional results in Table X, Table XI, and Table XII demonstrate the performance of Llama-2 series models (7B and 13B) discussed in Appendix C in terms of uncertainty quantification i.e. AUROC vs AUARC, coverage rate vs set size, and accuracy vs expected cali...

  38. [46]

    Lower ECE values indicate better calibration, signifying that confidence scores are more reliable indicators of prediction accuracy

    Accuracy vs ECE: Figure 5 shows the results of accuracy vs ECE achieved using CAP versus APS and LAC across multiple VLMs. Lower ECE values indicate better calibration, signifying that confidence scores are more reliable indicators of prediction accuracy. As shown in these fig...

  39. [47]

    Effect of Model Scale To examine the impact of model scale, we analyzed the performance of our CAP method across models of varying TABLE VII: Evaluation of uncertainty quantification: Comparative analysis of the proposed CAP (Ours) meth with standard Least Ambiguous set-valued...

    2002

Pith tools

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