Identification and Learning of Semantic Observation Kernels: Partial Observation, Uniform Recovery, & Minimax Limits

Matthew Francis Dixon

REVIEW 2 major objections 5 minor 18 references

A posterior over meaningful states can be recovered from grouped language probabilities whenever the induced observation law is injective and stable.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 03:30 UTC pith:76O47CJZ

load-bearing objection A serious and unusually candid methods paper that builds the right statistical scaffolding for turning LLM phrase probabilities into auditable state posteriors, provided the explicitly flagged sufficiency assumption is accepted. the 2 major comments →

arxiv 2607.23130 v1 pith:76O47CJZ submitted 2026-07-25 stat.ME stat.ML

Identification and Learning of Semantic Observation Kernels: Partial Observation, Uniform Recovery, & Minimax Limits

Matthew Francis Dixon This is my paper

classification stat.ME stat.ML MSC 62B1062G0562C2062M20

keywords semantic observation kernelsinverse problemspartial identificationnonparametric regressionminimax theorylarge language modelsposterior recoverystability certificates

verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper asks whether the observable conditional probabilities a text generator assigns to phrases can be turned into a posterior over declared states, rather than read as a model's self-reported confidence. It argues yes, under explicit conditions: group meaning-equivalent continuations into semantic cells, learn a calibration map on held-out pairs of grouped probabilities and reference posteriors, and freeze that map for deployment. The central results give an inverse modulus that separates existence, uniqueness, and error amplification; uniform nonparametric learning rates under dependent calibration data; a minimax boundary showing that certifying repeated-use stability is fundamentally harder than predicting well; and a set-valued treatment of missing probability mass. If true, this makes LLM phrase probabilities usable as an auditable external measurement, without claiming to interpret internal representations.

Core claim

The central claim is that the observable conditional probabilities a text generator assigns to complete continuations can serve as a measurement channel for a posterior over declared states, provided meanings are first grouped into prespecified semantic cells. The paper formalizes the channel as a finite pushforward of the language kernel through a measurable semantic partition, and proves that if this observation map is injective and continuous on a compact supported domain, the inverse exists and its finite-sample error is bounded by an inverse modulus applied to observation and optimization error. It then establishes that the calibration surface m0(s)=E(Theta|S=s) can be recovered from he

What carries the argument

The central object is the semantic observation law g_u(theta) = (Q_u{phi_u(Y)=j | theta})_j, the finite pushforward of an unrestricted language kernel through a declared semantic partition phi_u, together with the observable experiment E_O = {P^O_theta}. The inverse modulus omega_D(epsilon) = sup{||theta-theta'|| : d_E(P^O_theta,P^O_theta') <= epsilon} is the key quantity: it encodes both point identification (omega(0)=0) and stable conditioning (omega(eps)->0). The calibration map m0(s)=E(Theta|S=s) is the identifiable predictive inverse, and the derivative functional kappa(F0)=sup ||D_z F0||_op is the contraction certificate for repeated use. Each carries a distinct part of the argument: t

Load-bearing premise

The load-bearing premise is that the grouped language probability vector S is a sufficient statistic for the reference posterior—that Theta = m0(S) almost surely on the supported domain—together with injectivity of the induced observation map; without sufficiency, the learned map is a conditional mean, not a state posterior.

What would settle it

Find two calibration scenarios on the same supported domain whose observed grouped language probability vectors are equal (or within sampling error) but whose reference posteriors differ materially; if such collisions occur in held-out data beyond expected noise, then S is not sufficient and the predictive inverse is not a structural state measurement. Similarly, observe a small perturbation of a grouped probability vector that produces a posterior jump larger than the modeled inverse modulus, which would falsify stable recovery.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

On a declared state space, evidence class, prompt family, and supported calibration domain, an LLM's grouped phrase probabilities can be converted into a state posterior with quantified sampling error, and the conversion is auditable because it uses only observable probabilities.
Prompt rewording is harmless only when it leaves the grouped semantic vector unchanged; when it changes that vector, presentation has altered the measurement and must be modeled as observation noise.
Unreturned continuation probability mass must be represented as an identified set; assigning zero would manufacture certainty, and the sequential width of such sets is bounded by contraction of the state transition plus a persistent term.
One-shot predictive accuracy is insufficient for recursive deployment: stable repeated updating requires a simultaneous band on the derivative of the update, and no estimator can separate contraction from noncontraction faster than n^{-(s-1)/(2s+d)}.
A conformal region around the recovered posterior gives distribution-free marginal coverage but cannot establish identification; identification and support must be checked before coverage is interpreted.

Where Pith is reading between the lines

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

Extension: the same measurement construction could be applied to any fitted probabilistic generator with a defensible probability record, such as clinical prediction models or speech recognition systems, since the theorems do not depend on a transformer architecture.
Extension: the structural-inversion assumption can be tested empirically by checking whether held-out reference coordinates still vary after conditioning on the semantic probability summary; the paper's own table marks global injectivity as unverifiable, so a deployment guide would need a threshold for when a conditional-mean output is acceptable as a state measurement.
Extension: because the prompt is part of the observation design, an active measurement strategy that samples multiple information-equivalent prompts and checks agreement in calibrated outputs could strengthen evidence for identification beyond a single frozen prompt family.
Extension: if the minimax derivative boundary is intrinsic, high-stakes recursive use should budget sample sizes according to the desired resolution of the contraction coefficient near one, rather than only the predictive accuracy of the update.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

A serious and unusually candid methods paper that builds the right statistical scaffolding for turning LLM phrase probabilities into auditable state posteriors, provided the explicitly flagged sufficiency assumption is accepted.

read the letter

Two things to know. First, this is a serious and unusually honest methods paper: the inverse-modulus identification, compatible-set partial identification, and minimax derivative gate are genuinely useful, and the proofs in the appendix look correct. Second, the object being learned is the conditional mean m0(s)=E(Theta|S=s); it becomes a structural posterior only when Theta=m0(S) almost surely. The paper says this in Section 3 and Table 1 but never tests it, so the empirical "recovery" is regression performance on a self-constructed benchmark, not proof that language probabilities identify the state.

What the paper does well: it separates existence, identification, learnability, and stability, and insists on the distinction between predictive and structural inversion. The inverse modulus is a clean way to quantify when one-to-one maps are still unstable. The partial-identification treatment of missing candidate mass is honest—absent candidates are not replaced with zeros. The derivative stability gate, with its minimax boundary, is a real contribution and the simulations align with the theorems. The empirical LLM studies are clearly framed as illustrations on a declared support, with frozen prompts and disjoint partitions. Table 1 is the most candid assumption audit I have seen in a while.

The soft spots: the load-bearing assumption is that the grouped language probabilities S are sufficient for the reference posterior, and that the smooth calibration map persists on deployment. The paper marks both as unverifiable, then proceeds. That's fair for a methods paper, but it means the headline "held-out recovery" is conditional on a structural assumption the authors do not and cannot check from one archive. A sensitivity analysis that intentionally coarsens S would quantify what is lost. The beta-mixing assumption is also theoretical; the financial archive is cross-sectional, so it is not actually established there. The Gaussian-sieve exact band is verified only in simulation, not in the language illustration. These are minor given the paper's scope, not fatal.

Who should read it: statisticians working on semantic uncertainty, inverse problems, or measurement error, and quantitative practitioners who want to know when LLM probabilities can serve as measurements. It is a theory/methods paper, and a good one—the central argument holds up, with the caveat stated above.

For peer review: send it out. It deserves a serious referee. My own verdict would be conditional acceptance, with the condition that the authors add the coarsening sensitivity analysis and make the code and data available with a stable identifier.

Referee Report

2 major / 5 minor

Summary. This paper develops a semiparametric inverse-problem framework for converting observable grouped language probabilities from a frozen probabilistic text generator into a posterior over declared states. The main theoretical objects are an inverse modulus separating identification from conditioning (Theorem 2.2), a finite-replication/CLT result (Corollary 2.3), an impossibility bound under observational equivalence (Theorem 2.4), uniform nonparametric recovery rates for the predictive inverse under beta-mixing and estimated covariates (Theorem 3.2), a derivative-based contraction gate for sequential stability with an exact Gaussian-sieve band (Corollary 3.4) and a minimax lower bound (Theorem 3.5), and a partial-identification propagation theorem for truncated candidate probabilities (Theorem 4.1). Simulations are paired one-for-one with the theorems, and two frozen LLM studies on a 520-scenario financial benchmark illustrate held-out recovery and conformal coverage.

Significance. If the framework's assumptions hold, it provides a principled way to use LLM phrase probabilities as auditable state measurements without interpreting internal representations, separating semantic measurement from internal belief. The paper is unusually careful: Table 1 explicitly distinguishes verified-by-design conditions from assumed conditions, the proofs are standard, and the simulation design tests each theorem with a reproducibility supplement and fixed seeds. The main scientific value is the formalization of the inverse-modulus/derivative-gate distinction, which clarifies why predictive calibration cannot by itself certify structural recovery or recursive stability. The central caveat is that the learned object is the conditional mean E(Theta|S=s), so the measurement interpretation depends on a sufficiency/structural-identification condition that is plausible but not empirically verified in the presented archive.

major comments (2)

[Section 3, Table 1, Section 5.2] The identified object is m0(s)=E(Theta|S=s) (Section 3, 'Predictive versus structural inversion'), and all recovery rates in Theorem 3.2 are rates for this regression surface. The interpretation as recovery of the reference posterior requires the additional structural/sufficiency condition Theta_i=m0(S_i) a.s. Table 1 candidly marks global injectivity as 'not identified' and the calibration map as 'assumed and stress-tested, not globally verifiable.' Consequently, the held-out JS divergences in Table 5 demonstrate predictive calibration, not structural posterior recovery; the 520-scenario archive never varies the semantic partition or evidence set to test whether S is a sufficient coarsening. The abstract and Section 5.2 should either report a direct sufficiency check (e.g., nested semantic partitions or evidence ablations) or explicitly restrict the central claim to predictive calibrati
[Theorem 3.2 and Appendix A.4] The 'estimated semantic probabilities' extension asserts an additional O_P(C_G sqrt(log n/R_min)) function error and its h_n^{-1} multiple for derivatives, based on perturbing the local-polynomial normal equations. But when the covariate S_i itself is estimated, the regression surface identified from the estimated covariate is generally a convolution of m0 with the measurement-error distribution unless further conditions are imposed; a uniform bound on max_i ||hat S_i - S_i|| does not by itself yield the displayed perturbation argument. A formal generated-covariate expansion in the style of Mammen et al. (2012), with the required conditions stated, is needed before this rate claim can be considered established.

minor comments (5)

[Section 2, Definition 2.1] The notation L[...] is used without definition; please define it as 'the law of' at first use.
[Section 2, worked equivalence example] 'Theorem 2' in the worked example should be 'Theorem 2.4.'
[Section 3, 'Simple rate calculation'] The phrase 'reduces function error by about 32^{-2/5}=1/4' is mathematically misleading: 32^{-2/5}=0.25 is the remaining factor, so the error is reduced to about 1/4, not by 1/4.
[Table 5] Coverage values .940 and .900 are based on single 100-scenario test partitions; report the calibration and validation sizes and give a binomial interval for coverage so the reader can judge sampling variability.
[Appendix A.2] The phrase 'absorbing the conventional factor between total variation and L1 distance into the displayed modulus' is vague; state the factor explicitly (e.g., TV = (1/2)||.||_1) and adjust the modulus accordingly.

Circularity Check

0 steps flagged

No significant circularity: the held-out 'recovery' is an out-of-sample regression check, explicitly distinguished from structural inversion; all theorems follow from stated assumptions.

full rationale

The paper's central object is m0(s)=E(Theta|S=s), which it explicitly labels the 'identified predictive inverse' and states 'it equals a structural inverse only when Theta_i=m0(S_i) almost surely' (Section 3, 'Predictive versus structural inversion'). The held-out empirical study fits this regression on a calibration partition and evaluates it on a disjoint test partition, so the reported JS divergence is an honest out-of-sample prediction error, not a fitted parameter renamed as a discovery. Theorems 2.2, 2.4, 3.2, 3.3, 3.5, and 4.1 are proved from explicit assumptions (compact support, injectivity, Holder smoothness, beta-mixing, Gaussian noise, etc.); none of them reduces to its conclusion by construction. The paper repeatedly discloses that global injectivity is 'not identified' and that the smooth calibration map is 'assumed and stress-tested, not globally verifiable' (Table 1). That is a scope limitation, not circularity. There are no self-citations, no imported uniqueness theorems from the authors, and no ansatz smuggled in via citation. The sufficiency condition Theta=m0(S) is indeed untested, but the paper does not claim to verify it; instead it explicitly conditions the structural interpretation on that assumption. Consequently, no circular step can be quoted from the text.

Axiom & Free-Parameter Ledger

3 free parameters · 7 axioms · 0 invented entities

The paper's results rest on explicit statistical assumptions (Assumption 3.1, Table 1) rather than new physical entities. The main free/tuning choices are the bandwidth and finite-difference step, plus the prompt-offset parameter in the ablation. No invented entities are introduced.

free parameters (3)

Local-polynomial bandwidth h_n
Chosen by asymptotic balance h_n ≍ (log m_n/m_n)^{1/(2s+d)} in Theorem 3.2; no data-driven selector is specified, so a practitioner must choose it.
Finite-difference step h for the derivative gate
Theorem 3.3 sets h ≍ r_n(α)^{1/2}; the constant is not specified, affecting the radius ρ_n.
Prompt-offset correction
Table 4's final row estimates a prompt offset from calibration data and freezes it; the fitted value is not reported, and it is a fitted parameter in the empirical ablation.

axioms (7)

domain assumption Assumption 3.1: calibration pairs (S_i, Θ_i) are stationary, geometrically beta-mixing, with bounded design density, sub-Gaussian errors, and m0 in a Hölder ball C^s(S0; B).
Invoked in Theorem 3.2 for uniform nonparametric rates; unverifiable from the frozen cross-sectional archive (Table 1).
domain assumption The observation map θ ↦ P^O_θ is continuous and injective on the supported domain.
Theorem 2.2 requires injectivity for point identification; Table 1 states global injectivity outside the design is not identified.
domain assumption Reference posterior π*(E) is specified independently of the fitted language law.
Defines the estimand throughout; the paper notes this is verified by design in the experiments, not in the world.
domain assumption Semantic map φ_u is measurable and prespecified; no post-test ontology selection.
Required for the pushforward g_u(θ) to exist; Table 1 'Verified by experimental design'.
ad hoc to paper Corollary 3.4: F0 lies in a correctly specified p-dimensional Gaussian linear sieve with fixed full-rank design and normal errors.
The exact simultaneous band is finite-sample exact only under this correctly specified model; the paper notes it is verified only in the derivative-band simulation, not in the language illustration.
domain assumption For generated regressors, each S_i is estimated from R_min repeats via a locally C_G-Lipschitz inverse.
Used in Theorem 3.2's repeated-measurement error term; exact in simulation, operational approximation for fitted systems.
domain assumption State transition F_t is L_t-Lipschitz and the set-valued update U_t is Hausdorff nonexpansive.
Theorem 4.1's width propagation requires these structural assumptions on the external state-space model.

pith-pipeline@v1.3.0-alltime-deepseek · 18748 in / 14953 out tokens · 147231 ms · 2026-08-01T03:30:36.031986+00:00 · methodology

0 comments

read the original abstract

Probabilistic text generators supply conditional distributions over tokens and complete verbal continuations, whereas scientific use often requires a posterior over a finite state. Large language models are the leading example: phrase probabilities depend on prompt wording, and model-printed percentages are generated text rather than state posteriors. More generally, we ask when an observable language law can support a reproducible posterior over declared states. A semantic map groups meaning-equivalent continuations; held-out cases with reference posteriors identify a semiparametric inverse from grouped language probabilities to state probabilities. The language law remains nonparametric and no hidden model quantity is used. This is principally a theory and methods paper and makes several contributions. On the theoretical side, we derive conditions for existence, identification, stable recovery, and sequential updating; concentration, asymptotic, and nonparametric rates; identified sets under truncated probabilities; and a minimax boundary for uniform stability. On the empirical side, theorem- directed simulations verify recovery rates, compatible-set coverage, and stability gates, while two frozen language-model studies illustrate held-out recovery and conformal coverage. The results specify when observable language probabilities can provide an auditable state measurement without being interpreted as internal belief.

Figures

Figures reproduced from arXiv: 2607.23130 by Matthew Francis Dixon.

**Figure 2.** Figure 2: Inverse-recovery error against the number [PITH_FULL_IMAGE:figures/full_fig_p017_2.png] view at source ↗

**Figure 3.** Figure 3: Finite-sample behavior of the derivative stability certificate. Left: the probability of a [PITH_FULL_IMAGE:figures/full_fig_p018_3.png] view at source ↗

**Figure 4.** Figure 4: Oracle classification power under smooth alternatives below, at, and above the theoretical [PITH_FULL_IMAGE:figures/full_fig_p018_4.png] view at source ↗

discussion (0)

Reference graph

Works this paper leans on

18 extracted references · 4 canonical work pages

[17]

doi: 10.18653/v1/2025.findings-acl.1101. Bin Yu. Rates of convergence for empirical processes of stationary mixing sequences.The Annals of Probability, 22(1):94–116,

doi:10.18653/v1/2025.findings-acl.1101 2025
[1953]

Mark Braverman, Xinyi Chen, Sham Kakade, Karthik Narasimhan, Cyril Zhang, and Yi Zhang

doi: 10.1214/aoms/1177729032. Mark Braverman, Xinyi Chen, Sham Kakade, Karthik Narasimhan, Cyril Zhang, and Yi Zhang. Calibration, entropy rates, and memory in language models. InProceedings of the 37th Interna- tional Conference on Machine Learning, volume 119 ofProceedings of Machine Learning Research, pages 1089–1099,

arXiv
[1964]

François Le Gland and Laurent Mevel

doi: 10.1214/aoms/1177700372. François Le Gland and Laurent Mevel. Exponential forgetting and geometric ergodicity in hidden markov models.Mathematics of Control, Signals, and Systems, 13(1):63–93,

arXiv
[1982]

Anastasios N

doi: 10.1111/j.2517-6161.1982.tb01195.x. Anastasios N. Angelopoulos and Stephen Bates. Conformal prediction: A gentle introduction. Foundations and Trends in Machine Learning, 16(4):494–591,

arXiv 1982
[1994]

doi: 10.1214/aop/1176988849. 28

arXiv
[1996]

Sebastian Farquhar, Jannik Kossen, Lorenz Kuhn, and Yarin Gal

doi: 10.1007/978-94-009-1740-8. Sebastian Farquhar, Jannik Kossen, Lorenz Kuhn, and Yarin Gal. Detecting hallucinations in large language models using semantic entropy.Nature, 630:625–630,

doi:10.1007/978-94-009-1740-8
[2003]

doi: 10.1023/A:1023818214614. Heinz W. Engl, Martin Hanke, and Andreas Neubauer.Regularization of Inverse Problems. Kluwer Academic Publishers,

doi:10.1023/a:1023818214614
[2006]

Laurent Cavalier

doi: 10.1201/9781420010138. Laurent Cavalier. Nonparametric statistical inverse problems.Inverse Problems, 24(3):034004,

doi:10.1201/9781420010138
[2008]

Victor Chernozhukov, Han Hong, and Elie Tamer

doi: 10.1088/0266-5611/24/3/034004. Victor Chernozhukov, Han Hong, and Elie Tamer. Estimation and confidence regions for parameter sets in econometric models.Econometrica, 75(5):1243–1284,

work page doi:10.1088/0266-5611/24/3/034004
[2009]

Ramon van Handel

doi: 10.1007/ b13794. Ramon van Handel. Observability and nonlinear filtering.Probability Theory and Related Fields, 145:35–74, 2009a. doi: 10.1007/s00440-008-0161-y. Ramon van Handel. Uniform observability of hidden markov models and filter stability for unstable signals.The Annals of Applied Probability, 19(3):1172–1199, 2009b. doi: 10.1214/08-AAP576. Z...

work page doi:10.1007/s00440-008-0161-y
[2010]

Katherine Tian, Eric Mitchell, Allan Zhou, Archit Sharma, Rafael Rafailov, Huaxiu Yao, Chelsea Finn, and Christopher D

doi: 10.1146/annurev.economics.050708.143401. 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 scores from language models fine-tuned with human feedback. InProceedings of EMNLP, pages 5433–5442,

arXiv
[2011]

David Blackwell

doi: 10.3982/ECTA8680. David Blackwell. Equivalent comparisons of experiments.The Annals of Mathematical Statistics, 24 (2):265–272,

work page doi:10.3982/ecta8680
[2012]

Charles F

doi: 10.1214/12-AOS995. Charles F. Manski.Partial Identification of Probability Distributions. Springer, New York,

work page doi:10.1214/12-aos995
[2014]

Juan José Egozcue, Vera Pawlowsky-Glahn, Glòria Mateu-Figueras, and Carles Barceló-Vidal

doi: 10.1214/14-AOS1230. Juan José Egozcue, Vera Pawlowsky-Glahn, Glòria Mateu-Figueras, and Carles Barceló-Vidal. Isometric logratio transformations for compositional data analysis.Mathematical Geology, 35(3): 279–300,

doi:10.1214/14-aos1230
[2018]

Enno Mammen, Christoph Rothe, and Melanie Schienle

doi: 10.1080/01621459.2017.1307116. Enno Mammen, Christoph Rothe, and Melanie Schienle. Nonparametric regression with non- parametrically generated covariates.The Annals of Statistics, 40(2):1132–1170,

arXiv 2017
[2021]

Lorenz Kuhn, Yarin Gal, and Sebastian Farquhar

doi: 10.1162/tacl_a_00407. Lorenz Kuhn, Yarin Gal, and Sebastian Farquhar. Semantic uncertainty: Linguistic invariances for uncertainty estimation in natural language generation. InInternational Conference on Learning Representations,

doi:10.1162/tacl_a_00407
[2023]

Alexandre B

doi: 10.18653/v1/2023.emnlp-main.330. Alexandre B. Tsybakov.Introduction to Nonparametric Estimation. Springer,

doi:10.18653/v1/2023.emnlp-main.330 2023
[2025]

Zhiqiu Xia, Jinxuan Xu, Yuqian Zhang, and Hang Liu

URLhttps://proceedings.mlr.press/v258/wang25i.html. Zhiqiu Xia, Jinxuan Xu, Yuqian Zhang, and Hang Liu. A survey of uncertainty estimation methods on large language models. InFindings of the Association for Computational Linguistics: ACL 2025, pages 21381–21396,

2025

[1] [17]

doi: 10.18653/v1/2025.findings-acl.1101. Bin Yu. Rates of convergence for empirical processes of stationary mixing sequences.The Annals of Probability, 22(1):94–116,

doi:10.18653/v1/2025.findings-acl.1101 2025

[2] [1953]

Mark Braverman, Xinyi Chen, Sham Kakade, Karthik Narasimhan, Cyril Zhang, and Yi Zhang

doi: 10.1214/aoms/1177729032. Mark Braverman, Xinyi Chen, Sham Kakade, Karthik Narasimhan, Cyril Zhang, and Yi Zhang. Calibration, entropy rates, and memory in language models. InProceedings of the 37th Interna- tional Conference on Machine Learning, volume 119 ofProceedings of Machine Learning Research, pages 1089–1099,

arXiv

[3] [1964]

François Le Gland and Laurent Mevel

doi: 10.1214/aoms/1177700372. François Le Gland and Laurent Mevel. Exponential forgetting and geometric ergodicity in hidden markov models.Mathematics of Control, Signals, and Systems, 13(1):63–93,

arXiv

[4] [1982]

Anastasios N

doi: 10.1111/j.2517-6161.1982.tb01195.x. Anastasios N. Angelopoulos and Stephen Bates. Conformal prediction: A gentle introduction. Foundations and Trends in Machine Learning, 16(4):494–591,

arXiv 1982

[5] [1994]

doi: 10.1214/aop/1176988849. 28

arXiv

[6] [1996]

Sebastian Farquhar, Jannik Kossen, Lorenz Kuhn, and Yarin Gal

doi: 10.1007/978-94-009-1740-8. Sebastian Farquhar, Jannik Kossen, Lorenz Kuhn, and Yarin Gal. Detecting hallucinations in large language models using semantic entropy.Nature, 630:625–630,

doi:10.1007/978-94-009-1740-8

[7] [2003]

doi: 10.1023/A:1023818214614. Heinz W. Engl, Martin Hanke, and Andreas Neubauer.Regularization of Inverse Problems. Kluwer Academic Publishers,

doi:10.1023/a:1023818214614

[8] [2006]

Laurent Cavalier

doi: 10.1201/9781420010138. Laurent Cavalier. Nonparametric statistical inverse problems.Inverse Problems, 24(3):034004,

doi:10.1201/9781420010138

[9] [2008]

Victor Chernozhukov, Han Hong, and Elie Tamer

doi: 10.1088/0266-5611/24/3/034004. Victor Chernozhukov, Han Hong, and Elie Tamer. Estimation and confidence regions for parameter sets in econometric models.Econometrica, 75(5):1243–1284,

work page doi:10.1088/0266-5611/24/3/034004

[10] [2009]

Ramon van Handel

doi: 10.1007/ b13794. Ramon van Handel. Observability and nonlinear filtering.Probability Theory and Related Fields, 145:35–74, 2009a. doi: 10.1007/s00440-008-0161-y. Ramon van Handel. Uniform observability of hidden markov models and filter stability for unstable signals.The Annals of Applied Probability, 19(3):1172–1199, 2009b. doi: 10.1214/08-AAP576. Z...

work page doi:10.1007/s00440-008-0161-y

[11] [2010]

Katherine Tian, Eric Mitchell, Allan Zhou, Archit Sharma, Rafael Rafailov, Huaxiu Yao, Chelsea Finn, and Christopher D

doi: 10.1146/annurev.economics.050708.143401. 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 scores from language models fine-tuned with human feedback. InProceedings of EMNLP, pages 5433–5442,

arXiv

[12] [2011]

David Blackwell

doi: 10.3982/ECTA8680. David Blackwell. Equivalent comparisons of experiments.The Annals of Mathematical Statistics, 24 (2):265–272,

work page doi:10.3982/ecta8680

[13] [2012]

Charles F

doi: 10.1214/12-AOS995. Charles F. Manski.Partial Identification of Probability Distributions. Springer, New York,

work page doi:10.1214/12-aos995

[14] [2014]

Juan José Egozcue, Vera Pawlowsky-Glahn, Glòria Mateu-Figueras, and Carles Barceló-Vidal

doi: 10.1214/14-AOS1230. Juan José Egozcue, Vera Pawlowsky-Glahn, Glòria Mateu-Figueras, and Carles Barceló-Vidal. Isometric logratio transformations for compositional data analysis.Mathematical Geology, 35(3): 279–300,

doi:10.1214/14-aos1230

[15] [2018]

Enno Mammen, Christoph Rothe, and Melanie Schienle

doi: 10.1080/01621459.2017.1307116. Enno Mammen, Christoph Rothe, and Melanie Schienle. Nonparametric regression with non- parametrically generated covariates.The Annals of Statistics, 40(2):1132–1170,

arXiv 2017

[16] [2021]

Lorenz Kuhn, Yarin Gal, and Sebastian Farquhar

doi: 10.1162/tacl_a_00407. Lorenz Kuhn, Yarin Gal, and Sebastian Farquhar. Semantic uncertainty: Linguistic invariances for uncertainty estimation in natural language generation. InInternational Conference on Learning Representations,

doi:10.1162/tacl_a_00407

[17] [2023]

Alexandre B

doi: 10.18653/v1/2023.emnlp-main.330. Alexandre B. Tsybakov.Introduction to Nonparametric Estimation. Springer,

doi:10.18653/v1/2023.emnlp-main.330 2023

[18] [2025]

Zhiqiu Xia, Jinxuan Xu, Yuqian Zhang, and Hang Liu

URLhttps://proceedings.mlr.press/v258/wang25i.html. Zhiqiu Xia, Jinxuan Xu, Yuqian Zhang, and Hang Liu. A survey of uncertainty estimation methods on large language models. InFindings of the Association for Computational Linguistics: ACL 2025, pages 21381–21396,

2025