Pith. sign in

REVIEW 2 major objections 7 minor 1 cited by

Position: Uncertainty Quantification Needs Reassessment for Large-language Model Agents

T0 review · 2 major / 7 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper argues that the aleatoric/epistemic uncertainty dichotomy breaks down in interactive LLM agent settings and should be replaced by underspecification detection, interactive learning, and verbalized output uncertainties.

desk verdict A defensible, useful position piece: the aleatoric/epistemic dichotomy is indeed a poor fit for interactive LLM agents, but the authors' strongest claim hinges on a modeler-relative semantics they don't fully justify. read the letter →

arxiv 2505.22655 v1 pith:QIS5JY5O submitted 2025-05-28 cs.LG cs.AIcs.CL

Michael Kirchhof , Gjergji Kasneci , Enkelejda Kasneci This is my paper
classification cs.LGcs.AIcs.CL
keywords uncertaintyquantificationaleatoricepistemiclargelanguagemodelsLLMagentsunderspecificationinteractivelearningoutputuncertainties
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

Large language models increasingly act as interactive agents, yet the standard toolkit for quantifying their uncertainty still reduces everything to one number, or to two numbers labeled aleatoric and epistemic. This position paper tries to establish that this dichotomy does not survive contact with multi-turn conversation: popular definitions of aleatoric and epistemic uncertainty contradict one another even in toy examples, and in a chat the same uncertainty can look epistemic when the agent may ask a follow-up question and aleatoric the moment it chooses to stop asking. The paper argues that uncertainty research for LLM agents should shift toward three new directions: underspecification uncertainties, where the user has not supplied enough information or even a fully defined task; interactive learning, where the agent asks clarifying questions to reduce that uncertainty; and output uncertainties, where the agent communicates what it is unsure about in words or speech rather than only as numbers. If the position holds, uncertainty quantification for LLM agents becomes less about estimating irreducible risk and more about detecting missing information, deciding whether and what to ask, and saying honestly what remains uncertain.

What carries the argument

The mechanism that carries the argument is the conversation itself viewed as a moving input point in a fixed input space. In a traditional setup the input $x$ is fixed and the uncertainty about the output $y$ is split into aleatoric and epistemic parts; in a chatbot, the input is the concatenated conversation string $x_t$, which changes with each turn, so the same question $y$ is predicted from different information at different times. Because the agent can choose to ask follow-up questions, any current uncertainty can be classified as reducible (epistemic) if the agent will keep asking, or irreducible (aleatoric) if it stops — yielding the paper's central claim that the labels flip and therefore lose objective meaning. The proposed alternative is a three-part structure: task- and context-underspecification uncertainties, interactive learning as a mechanism for reducing them, and output uncertainties as the communicative layer that replaces scalar scores with sets of possibilities, explanations, and verbal or prosodic expressions.

What would settle it

Build a controlled chat where the user's question is ambiguous but all further information is unavailable (the user answers "I don't know" to any follow-up). If an uncertainty estimator labels the state epistemic while follow-up questions are permitted and aleatoric as soon as the agent is forced to stop, while the available information is identical, the paper's agent-relative labeling is demonstrated. If the label instead stays stable, or tracks only the information actually supplied, the central breakdown claim is refuted.

Watch

Extended reading notes

Core claim

The paper's central claim is that the aleatoric/epistemic dichotomy, as commonly used in machine learning, is inapplicable to modern LLM agent interactions. To support this, it collects schools of thought that define epistemic uncertainty as the number of plausible models, as disagreement among models, or as distance from training data, and shows these give opposite answers to a simple two-Dirac belief example. It collects analogous conflicts for aleatoric uncertainty, where one school counts all residual risk of the Bayes-optimal model inside a fixed model class as aleatoric while another counts only the variance intrinsic to the data. It then argues that in a conversation the same uncertainty changes label depending on whether the agent continues to ask follow-up questions, so reducibility is a subjective modeling choice rather than a property of the world. The constructive part of the position is that research should develop three complementary notions — underspecification uncertainty, interactive learning, and output uncertainties — and that numerical probabilities remain useful mainly for machine-to-machine interfaces.

Load-bearing premise

The argument that epistemic uncertainty becomes aleatoric the moment an agent stops asking questions assumes that reducibility is judged relative to the agent's chosen stopping point rather than relative to the world; if more information exists that would settle the answer, the claimed label flip dissolves.

Editorial extensions

If this is right

  • Uncertainty quantification for LLM agents should be re-oriented from decomposing total uncertainty into aleatoric and epistemic parts toward detecting when a user's request is underspecified.
  • Agents should be built and evaluated on their ability to ask helpful follow-up questions, with an explicit trade-off between informativeness and user patience.
  • Uncertainty communication research should develop benchmarks and metrics for textual and speech-based expressions of uncertainty, including the coverage of listed possibilities.
  • Numerical uncertainty scores remain appropriate for automated consumers of agent output, such as abstention thresholds in pipelines, but are insufficient for human users.
  • Existing evidence that ambiguity appears in a large fraction of real user questions (for example, 56% of one open-domain QA benchmark) becomes a first target for the proposed underspecification direction.

Reading between the lines

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

  • One extension the paper leaves implicit: if reducibility labels are agent-relative, then the right formal object is not a label but a decision — the agent's stopping rule — and uncertainty could be expressed as the expected value of asking one more question, connecting to value-of-information theory.
  • A testable extension is to measure whether existing aleatoric and epistemic estimators, which the paper cites as almost perfectly correlated in static settings, predict the usefulness of follow-up questions in dialogue; if they do, the proposed research directions can reuse existing tools rather than requiring new ones.
  • The same reasoning applies beyond text chat to any interactive agent that can acquire information at a cost, such as robots or diagnostic systems, so the three proposed uncertainty types could serve as a general template for embodied uncertainty quantification.
  • A concrete experiment to sharpen the position would show users the same uncertainty expressed as a calibrated number versus as a short list of possibilities with reasons, then measure decision quality and trust; the paper predicts the verbal form will support human decision making better, but it does not itself run this study.
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

2 major / 7 minor

Summary. This position paper argues that the traditional aleatoric/epistemic uncertainty dichotomy is insufficient for LLM agents in open, interactive, multi-turn settings. It reviews conflicting definitions of the two uncertainty types (Sections 2.1–2.2), presents empirical and theoretical evidence that their estimators are entangled (Section 2.3), and argues that in chatbot conversations the epistemic/aleatoric label can flip depending on whether the agent continues to ask follow-up questions (Section 2.4). Based on these arguments, the paper proposes three research directions: underspecification uncertainties, interactive learning, and output uncertainties (Section 3), while acknowledging in Section 4 that traditional uncertainties and numerical scores remain useful in some contexts, such as automated downstream systems.

Significance. If the position is accepted, it would shift UQ research priorities for LLM agents away from scalar aleatoric/epistemic decomposition toward interactive clarification, underspecification detection, and verbalized output uncertainty. The paper provides a valuable synthesis of a fragmented literature, including a useful table (Table 1) of conflicting schools of thought and an explicit discussion of counter-positions. The three proposed research directions are concrete and lead to falsifiable benchmarks (e.g., ambiguity detection accuracy, helpfulness of follow-up questions). However, the paper's strongest claim—that the dichotomy is 'inapplicable' to LLM agents—rests substantially on the Section 2.4 argument, which is not fully justified.

major comments (2)
  1. [2.4] The epistemic-to-aleatoric flip argument relies on an agent-relative notion of reducibility, following Der Kiureghian and Ditlevsen (2009), but the paper does not independently justify why this semantics is appropriate for LLM-agent uncertainty. Under the standard feature-relative view (aleatoric uncertainty is H[Y | X=x_t], and reducibility is evaluated by whether some additional observation Z exists with H[Y | X=x_t, Z] < H[Y | X=x_t]), the agent's decision to stop asking questions merely changes the conditioning point; it does not change the label. Without such a justification, the claim in Section 1 that the dichotomy is 'inapplicable' is not established by Section 2.4. The paper should either defend the agent-relative semantics explicitly or weaken the claim to 'ambiguous relative to the modeler's chosen stopping point.'
  2. [2.1–2.3] The paper moves from 'different schools of thought define these terms differently' and 'estimators are correlated' to the conclusion that aleatoric and epistemic uncertainty 'cannot be split exactly.' The two-Diracs example in Section 2.1 shows that different axiomatic systems yield different numerical values, which is a lack of consensus rather than a logical contradiction. The empirical correlation results (Mucsányi et al. 2024; de Jong et al. 2024) show that current estimates are entangled, but they do not establish that the underlying quantities are inherently inseparable; the paper acknowledges the approximation-error objection in Section 2.3 but does not fully rebut it. The theoretical argument from Gruber et al. (2023) concerns a specific linear-model prediction interval and does not generalize to all settings. These considerations support the need for careful definitions and enrichment, but not the strong claim that the dichotomy is inapplicable.
minor comments (7)
  1. [Page 1] The 'Apple Confidential–Internal Use Only' header appears to be a version-control artifact and should be removed.
  2. [3.2] Typo: 'resemblence' should be 'resemblance'.
  3. [3.3] Typo: 'possiblities' should be 'possibilities'.
  4. [2.4] The notation 'P(Y)' is used without defining Y as the random variable for the answer token or string; please introduce it explicitly.
  5. [3.1] Equation (2) writes an integral over T, which appears to assume a continuous task space; if T is intended to be discrete (or a mixture), the equation should be written accordingly.
  6. [References] There are two entries for Belém et al. (2024) with slightly different spellings (Belem vs. Belém) and identical content; these should be consolidated.
  7. [Figure 2] The vertical axis label 'Q(θ)' is not explained as a probability density; the caption should state that Q is a second-order distribution.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the position is an externally grounded literature synthesis, not a derivation that reduces to its own inputs.

full rationale

The paper advances a position rather than a fitted derivation, and none of its load-bearing steps reduce to its own inputs by construction. The definitional-conflict examples in Sections 2.1-2.3 are presented as observed disagreements between named external schools of thought (e.g., Houlsby et al. 2011, Wimmer et al. 2023, Mukhoti et al. 2023) and are not derived from the paper's conclusion. The empirical disentanglement claim in Section 2.3 does cite the authors' own Mucsányi et al. (2024) benchmark, but that benchmark is an externally published, reproduced result and is corroborated by independent citations (de Jong et al. 2024; Bouvier et al. 2022; Gruber et al. 2023; Hüllermeier et al. 2022; Valdenegro-Toro & Mori 2022), so the self-citation is not load-bearing. The other self-citations (Kirchhof et al. 2023, 2025; Kasneci et al. 2023) appear only as examples or pointers and do not carry the argument. Section 2.4's epistemic-to-aleatoric flip is explicitly framed as 'one can argue... however, one can equally argue' and is attributed to the external modeler-relative convention of Der Kiureghian & Ditlevsen (2009); whether that convention is the right semantics for LLM agents is a substantive justification question, not a self-referential reduction. Equations (1) and (2) are standard decompositions and marginalizations cited from the literature rather than new predictions. The 'Apple Confidential' header is a metadata artifact and does not affect the argument chain.

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

The position depends on standard Bayesian formalizations, a contestable agent-relative notion of reducibility, a fixed input-space framing for conversations, and a value judgment about the importance of human-facing transparency. No fitted parameters or invented physical entities appear.

assumptions (5)
  • standard math Second-order (Bayesian) distributions formalize the epistemic state of a learner.
    Used throughout Section 2.1 and in Eq. (1) to define epistemic and aleatoric uncertainty; assumed without proof, but standard in the cited Bayesian ML literature.
  • ad hoc to paper Reducibility of uncertainty is relative to the agent's decision to stop gathering information.
    Section 2.4, following Der Kiureghian and Ditlevsen, treats the agent's choice to stop asking follow-up questions as converting epistemic uncertainty into aleatoric uncertainty. This is the load-bearing premise behind the claim that labels flip, and it is asserted rather than proved.
  • ad hoc to paper The input space of a chatbot is the set of all finite strings and remains fixed across conversation turns.
    Section 2.4's footnote argues that adding information via follow-up questions does not change the input space, which makes the epistemic/aleatoric flip appear to violate the traditional definitions.
  • domain assumption Calibrated language models must hallucinate, as shown by Kalai and Vempala.
    The introduction uses this result to motivate the importance of uncertainty quantification; it is accepted from the cited STOC paper rather than proved here.
  • domain assumption Human-interaction transparency is a primary goal for LLM-agent uncertainty quantification.
    This premise motivates the emphasis on output uncertainties over scalar scores; the authors acknowledge in Section 4.3 that calibrated numbers remain useful for automated downstream systems.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Position: Uncertainty Quantification Needs Reassessment for Large-language Model Agents." pith.science (2026). https://pith.science/paper/QIS5JY5O

@misc{pith2026250522655,
  author       = {Pith},
  title        = {Pith review of: Position: Uncertainty Quantification Needs Reassessment for Large-language Model Agents},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QIS5JY5O}},
  note         = {Machine review of arXiv:2505.22655}
}
read the original abstract

Large-language models (LLMs) and chatbot agents are known to provide wrong outputs at times, and it was recently found that this can never be fully prevented. Hence, uncertainty quantification plays a crucial role, aiming to quantify the level of ambiguity in either one overall number or two numbers for aleatoric and epistemic uncertainty. This position paper argues that this traditional dichotomy of uncertainties is too limited for the open and interactive setup that LLM agents operate in when communicating with a user, and that we need to research avenues that enrich uncertainties in this novel scenario. We review the literature and find that popular definitions of aleatoric and epistemic uncertainties directly contradict each other and lose their meaning in interactive LLM agent settings. Hence, we propose three novel research directions that focus on uncertainties in such human-computer interactions: Underspecification uncertainties, for when users do not provide all information or define the exact task at the first go, interactive learning, to ask follow-up questions and reduce the uncertainty about the current context, and output uncertainties, to utilize the rich language and speech space to express uncertainties as more than mere numbers. We expect that these new ways of dealing with and communicating uncertainties will lead to LLM agent interactions that are more transparent, trustworthy, and intuitive.

Figures

Figures reproduced from arXiv: 2505.22655 by the authors.

Figure 1
Figure 1. The traditional view on uncertainties suggests a clear black-and-white dichotomy between aleatoric and epistemic un￾certainty. We argue that recent developments show this dichotomy is not that simple, and not helpful for developing LLM agents. reiro et al., 2023), that is, to make up a response that is incorrect. Recent research has shown that this behaviour is rooted in their very generative nature, such that we ca… view at source ↗
Figure 2
Figure 2. In a binary prediction, the learner may have a belief that the Bernoulli probability is either high or low. Some schools of thought see this as a case of maximum epistemic uncertainty whereas other see it as nearly minimal epistemic uncertainty. 2. Where Traditional Uncertainties Fail This section gives an introduction to the traditional research on uncertainties. We critically review the popular dichotomy of aleato… view at source ↗
Figure 3
Figure 3. Using a too simple model class, like a linear model to fit quadratic data, leads to wide uncertainty estimates. The ques￾tion is whether this is irreducible, and thus aleatoric uncertainty. Bayes-optimality schools of thoughts would argue that yes, it is irreducible within the model class and thus aleatoric, whereas data￾uncertainty schools of thought would argue that it is reducible when choosing a better-suited mo… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: When estimating aleatoric and epistemic uncertainties, they can often not be disentangled. This plot is reproduced with permission from Mucsanyi et al. ´ (2024), where Equation (1) was used to split aleatoric and epistemic uncertainty of a deep ensemble trained on Imag…
Figure 5
Figure 5. Figure 5: ArXiv preprints in computer science, statistics, and math that include the terms ”aleatoric” or ”epistemic” in their title or abstract. The usage is at an all-time high, with roughly one paper being published each day in 2024. An LLM agent must also communicate which p…

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

  1. Uncertainty-Driven Reliability: Selective Prediction and Trustworthy Deployment in Modern Machine Learning

    cs.LG 2025-08 conditional novelty 5.0 of 10

    A training-dynamics abstention method matches deep ensembles at a fraction of the training cost, and a five-term error budget explains why selective classifiers still fall short of the oracle.

Reference graph

Works this paper leans on

24 extracted references · 6 canonical work pages · cited by 1 Pith paper

  1. [3]

    A multitask, mul- tilingual, multimodal evaluation of chatgpt on reason- ing, hallucination, and interactivity.arXiv preprint arXiv:2302.04023,

    Yejin Bang, Samuel Cahyawijaya, Nayeon Lee, Wenliang Dai, Dan Su, Bryan Wilie, Holy Lovenia, Ziwei Ji, Tiezheng Yu, Willy Chung, et al. A multitask, mul- tilingual, multimodal evaluation of chatgpt on reason- ing, hallucination, and interactivity.arXiv preprint arXiv:2302.04023,

    arXiv

  2. [7]

    How disentangled are your classifi- cation uncertainties?arXiv preprint arXiv:2408.12175,

    Ivo Pascal de Jong, Andreea Ioana Sburlea, and Matias Valdenegro-Toro. How disentangled are your classifi- cation uncertainties?arXiv preprint arXiv:2408.12175,

    arXiv

  3. [9]

    Ensembling over Classifiers: a Bias-Variance Perspective

    10 Position: Uncertainty Quantification Needs Reassessment for Large-language Model Agents Neha Gupta, Jamie Smith, Ben Adlam, and Zelda Mariet. Ensembling over classifiers: a bias-variance perspective. arXiv preprint arXiv:2206.10566,

    work page Pith review arXiv

  4. [11]

    Large language models must be taught to know what they don’t know.arXiv preprint arXiv:2406.08391,

    Sanyam Kapoor, Nate Gruver, Manley Roberts, Katherine Collins, Arka Pal, Umang Bhatt, Adrian Weller, Samuel Dooley, Micah Goldblum, and Andrew Gordon Wilson. Large language models must be taught to know what they don’t know.arXiv preprint arXiv:2406.08391,

    arXiv

  5. [13]

    Accessed on 04.08.2024

    URL https://twitter.com/BlackHC/ status/1817556167687569605. Accessed on 04.08.2024. Zahra Kolagar and Alessandra Zarcone. Aligning uncer- tainty: Leveraging llms to analyze uncertainty transfer in text summarization. InProceedings of the 1st Workshop on Uncertainty-Aware NLP (UncertaiNLP 2024), pp. 41– 61,

    work page arXiv 2024

  6. [14]

    Predictive uncertainty quantification via risk decompositions for strictly proper scoring rules.arXiv preprint arXiv:2402.10727,

    Nikita Kotelevskii and Maxim Panov. Predictive uncertainty quantification via risk decompositions for strictly proper scoring rules.arXiv preprint arXiv:2402.10727,

    arXiv

  7. [17]

    AmbigQA: Answering ambiguous open- domain questions

    Sewon Min, Julian Michael, Hannaneh Hajishirzi, and Luke Zettlemoyer. AmbigQA: Answering ambiguous open- domain questions. InProceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP),

    work page 2020

  8. [20]

    On uncertainty in natural language process- ing.arXiv preprint arXiv:2410.03446,

    Dennis Ulmer. On uncertainty in natural language process- ing.arXiv preprint arXiv:2410.03446,

    arXiv

Show all 24 references
  1. [21]

    Sayself: Teach- ing llms to express confidence with self-reflective ra- tionales, 2024a

    Tianyang Xu, Shujin Wu, Shizhe Diao, Xiaoze Liu, Xingyao Wang, Yangyi Chen, and Jing Gao. Sayself: Teach- ing llms to express confidence with self-reflective ra- tionales, 2024a. URL https://arxiv.org/abs/ 2405.20974. Ziwei Xu, Sanjay Jain, and Mohan Kankanhalli. Hallucina- ti...

    arXiv

  2. [22]

    Gal Yona, Roee Aharoni, and Mor Geva

    URLhttps://arxiv.org/abs/2505.16922. Gal Yona, Roee Aharoni, and Mor Geva. Can large language models faithfully express their intrinsic uncertainty in words? In Yaser Al-Onaizan, Mohit Bansal, and Yun- Nung Chen (eds.),Proceedings of the 2024 Conference on Empirical Methods in...

    2024

  3. [23]

    Boxuan Zhang and Ruqi Zhang

    URL https://arxiv.org/abs/ 2505.14489. Boxuan Zhang and Ruqi Zhang. CoT-UQ: Improving response-wise uncertainty quantification in llms with chain-of-thought,

  4. [24]

    Michael J.Q

    URL https://arxiv.org/ abs/2502.17214. Michael J.Q. Zhang and Eunsol Choi. Clarify when neces- sary: Resolving ambiguity through interaction with lms,

    arXiv

  5. [1990]

    Uncertainty in natural language generation: From theory to applications.arXiv preprint arXiv:2307.15703,

    Joris Baan, Nico Daheim, Evgenia Ilia, Dennis Ulmer, Haau- Sing Li, Raquel Fern´andez, Barbara Plank, Rico Sennrich, Chrysoula Zerva, and Wilker Aziz. Uncertainty in natural language generation: From theory to applications.arXiv preprint arXiv:2307.15703,

    arXiv

  6. [1997]

    Bayesian active learning for classification and preference learning.arXiv preprint arXiv:1112.5745,

    Neil Houlsby, Ferenc Husz ´ar, Zoubin Ghahramani, and M´at´e Lengyel. Bayesian active learning for classification and preference learning.arXiv preprint arXiv:1112.5745,

    arXiv

  7. [2004]

    Towards clear expectations for uncertainty estimation.arXiv preprint arXiv:2207.13341,

    Victor Bouvier, Simona Maggio, Alexandre Abraham, and L´eo Dreyfus-Schmidt. Towards clear expectations for uncertainty estimation.arXiv preprint arXiv:2207.13341,

    arXiv

  8. [2013]

    Know what you don’t know: Unanswerable questions for squad

    Pranav Rajpurkar, Robin Jia, and Percy Liang. Know what you don’t know: Unanswerable questions for squad. arXiv preprint arXiv:1806.03822,

    arXiv

  9. [2017]

    Sources of un- certainty in machine learning–a statisticians’ view.arXiv preprint arXiv:2305.16703,

    Cornelia Gruber, Patrick Oliver Schenk, Malte Schierholz, Frauke Kreuter, and G¨oran Kauermann. Sources of un- certainty in machine learning–a statisticians’ view.arXiv preprint arXiv:2305.16703,

    arXiv

  10. [2018]

    On information-theoretic measures of predictive uncertainty.arXiv preprint arXiv:2410.10786,

    Kajetan Schweighofer, Lukas Aichberger, Mykyta Ielan- skyi, and Sepp Hochreiter. On information-theoretic measures of predictive uncertainty.arXiv preprint arXiv:2410.10786,

    arXiv

  11. [2019]

    Perceptions of linguistic un- certainty by language models and humans.arXiv preprint arXiv:2407.15814,

    Catarina G Belem, Markelle Kelly, Mark Steyvers, Sameer Singh, and Padhraic Smyth. Perceptions of linguistic un- certainty by language models and humans.arXiv preprint arXiv:2407.15814,

    arXiv

  12. [2020]

    Teach- ing models to express their uncertainty in words.arXiv preprint arXiv:2205.14334,

    Stephanie Lin, Jacob Hilton, and Owain Evans. Teach- ing models to express their uncertainty in words.arXiv preprint arXiv:2205.14334,

    arXiv

  13. [2022]

    Dai, Jakob Uszkoreit, Quoc Le, and Slav Petrov

    Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Jacob Devlin, Kenton Lee, Kristina Toutanova, Llion Jones, Matthew Kelcey, Ming- Wei Chang, Andrew M. Dai, Jakob Uszkoreit, Quoc Le, and Sla...

    arXiv

  14. [2023]

    Phiseg: Capturing uncertainty in medical image segmentation

    Christian F Baumgartner, Kerem C Tezcan, Krishna Chai- tanya, Andreas M H¨otker, Urs J Muehlematter, Khoschy Schawkat, Anton S Becker, Olivio Donati, and Ender Konukoglu. Phiseg: Capturing uncertainty in medical image segmentation. InMedical Image Computing and Computer Assist...

    2019

  15. [2024]

    Llms will always hallucinate, and we need to live with this

    Sourav Banerjee, Ayushi Agarwal, and Saloni Singla. Llms will always hallucinate, and we need to live with this. arXiv preprint arXiv:2409.05746,

    arXiv

  16. [2025]

    Andreas Kirsch

    URL https: //arxiv.org/abs/2505.20295. Andreas Kirsch. (implicit) ensembles of ensembles: Epis- temic uncertainty collapse in large models.arXiv preprint arXiv:2409.02628,

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.