REVIEW 3 major objections 4 minor 59 references
Which Model Is Actually Serving You? IRIS: Budgeted Black-Box Auditing of Model Substitution and Routing Dilution in LLM Gateways
T0 review · 3 major / 4 minor · reviewed 2026-08-01 · deepseek-v4-flash
Pith's one-line read A text-only audit can detect when an LLM gateway serves a cheaper model, name the substitute, and estimate the diluted fraction before spending a query budget it sizes itself.
desk verdict IRIS nails the text-only dilution-detection task on well-separated pairs, but the abstract oversells the self-sizing budget and the pooled false-positive rate; the authors themselves disclose the gaps in the appendix. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The central object is the random-generation probe family c_{n,L} (e.g., c_{2,16} for 16 random bits, c_{10,8} for 8 random digits) paired with a 179-dimensional visible-string feature vector and a multiclass classifier's negative log-posterior score. The load-bearing mechanism is the 'estimate-then-budget' pilot: a cheap labeled pilot (m <= 4 queries) fits the exponential rank-error decay log(1-AUROC) ~ const - I_auc * m, and the fitted exponent (with a lower-confidence bound) is used to freeze the live-query budget and select the probe before any suspect traffic is queried. For dilution, response-level 'tell' rates (reference-atypical responses) are calibrated on a separate split, and the t
What would settle it
A concrete observation that would settle the central claim: take a fresh honest endpoint of the same model not used in calibration, run the frozen IRIS plan (same probe, same budget, same thresholds), and measure the realized false-positive rate across many independent audit runs. If the per-pair FPR exceeds the claimed 0.017 (or the pre-registered alpha) on a material fraction of pairs, or if the realized power at the pre-committed budget falls below the target on a substantial minority of margin-qualified pairs, then the calibration-based budget guarantee fails. The paper itself reports such
Extended reading notes
Core claim
The paper's central claim is that visible-string biases in responses to random-generation prompts—such as asking for digits, bits, or short sequences—are backend-specific fingerprints that can be used to audit LLM gateways. IRIS turns each response into a 179-dimensional feature vector, trains a multiclass classifier on enrolled endpoints, and scores responses by the reference negative log-posterior. It then calibrates telling thresholds and, critically, estimates the query budget before any suspect query is issued: a pilot fits the exponential decay of rank error (1 - AUROC) and the dilution tell-rate regime, yielding a pre-committed budget. The paper argues that evidence accumulates across
Load-bearing premise
The honest reference calibration window is representative and large enough that the thresholds and budgets frozen on it control type-I error at the stated level and meet the power target on a new audit stream.
Editorial extensions
If this is right
- A client can audit an LLM gateway unilaterally from returned text alone, with no privileged access, and with a budget that is estimated before the audit begins.
- The audit can not only decide whether dilution is happening but also name the substitute and estimate the routing fraction, making the result actionable for billing disputes or price renegotiation.
- The measured 1/epsilon budget law means that even low-rate dilution (e.g., epsilon=0.1) is detectable within a realistic number of queries when a separating tail exists, and the paper's tail-exponent measurement can decide which pairs fall into the cheap 1/epsilon regime versus the expensive 1/epsilon^2 wall.
- Because the probes are reusable and task-agnostic, the same audit pipeline transfers across endpoints without target-specific probe design, unlike prior knowledge-boundary or border-input audits.
- A real cross-provider audit of the same nominal model served by different providers found 14 of 15 pairs distinguishable, indicating that the method can flag genuine quantization/kernel deviations in the wild.
Reading between the lines
- If IRIS's calibration guarantees hold in deployment, a practical consequence is that gateway providers can be held accountable for silent downgrades: a customer could routinely audit a fraction of their traffic and obtain evidence that stands up enough to contest billing, even without provider cooperation.
- The paper's distinction between a pure temperature retune (second-order, near-invisible) and a base-model swap (first-order, detectable) suggests that the audit's 'blind spot' is intentionally aligned with economic incentives: tempering alone does not create extra gateway profit under token-count billing, so the audit focuses on the substitution threat that matters financially.
- A testable extension is to turn the audit into an anytime-valid sequential test using the same tell-count machinery; the paper's own appendix compares a betting e-process and SPRT, so a deployment could stop early on clearly honest or clearly diluted streams while preserving the pre-committed budget contract.
- The measured tail exponent kappa is reported as a pair-specific property; if a gateway adaptively switches between multiple diluents or uses sticky routing, the i.i.d. tell-count interval degrades, and the paper's block-bootstrap repair suggests that per-session or longer audit horizons would be needed to maintain calibrated uncertainty on the routing fraction.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper presents IRIS, an output-only (visible-string) black-box auditor for LLM gateways. It uses random-generation probes, a 179-dimension visible-string feature vector, and a random-forest posterior to detect whole-stream substitution and fractional dilution, attribute the served backend, estimate the routing fraction ε, and pre-commit a query budget estimated from a cheap pilot. The theoretical core (Section 5, App. S28) proves exponential decay of mean-score verification error, a response-level tell-rate characterization of the dilution budget, and a Θ(ε^{-2}) all-test lower bound under a χ²-regularity condition, with an interpolating tail-exponent law (Prop. 6). Experiments on a local Qwen3 ladder, a 17- and 45-endpoint OpenRouter library, and a live cross-provider audit report strong detection, attribution, and ε-recovery numbers, supported by a released code/data bundle. The appendices are unusually candid about finite-sample limitations, including per-pair FPR tails, under-coverage of the δ-method interval, and the fact that the prescribed budget meets the power target on only a minority of pairs at the calibrated sample size.
Significance. If the claims hold, IRIS is a substantial advance: a text-only auditor that detects dilution, names the diluent, estimates ε, and predicts its own query cost before issuing suspect queries would close a real gap in gateway auditing. The paper's strengths are concrete: machine-checkable-style complete proofs of the main propositions in App. S28; a large frozen response dataset and code release; head-to-head comparisons against FLIPS, MET, and RUT on a shared probe; and a clear identification of the 1/ε versus 1/ε² regimes as a measured tail-exponent phenomenon rather than an unconditional law. The paper also deserves credit for explicitly labeling its own reliability limits in App. S5 (Table 3: the prescribed budget is a 'feasibility diagnostic, not a coverage guarantee') and for reporting FPR tails and held-out margin-selection checks rather than only pooled means. However, the central abstract claims—'a query budget it sizes itself' and a '0.017 false-positive rate'—are stronger than what the paper's own finite-calibration results support, and this is load-bearing for the paper's main novelty.
major comments (3)
- [Abstract; §1 Contributions; App. S5] The central reliability claims are not supported as stated. App. S5's end-to-end check at (α=0.05, δ=0.1, ε=0.3) shows realized power averaging 0.73 but meeting the pre-registered 0.9 target on only 17.8% of feasible pairs (12.6% at ε=0.2), and per-pair type-I FPR with mean 0.013 but maximum 0.475 before hardening and 0.39 after exact-binomial hardening. Table 3 of the appendix explicitly labels the prescribed budget a 'feasibility diagnostic, not a coverage guarantee.' The abstract's '0.017 false-positive rate' is the pooled mean, not a per-audit guarantee, and 'a query budget it sizes itself' is not a reliability contract. Revision must either provide a calibration-window size/composition rule that restores the (α, δ) guarantee at the stated level, or re-scope the claim to 'feasibility diagnostic' and report the per-pair FPR distribution rather than the pooled mean.
- [§4.1 Eq. (7), Eq. (12); App. S4] The estimate-then-budget pilot is not calibration-robust. App. S4 shows that a 1000-sample reference calibration overshoots the realized total FPR to 0.09–0.14 on c2,16 (≤0.04 on c10,8), and the certifiable budget is only m≤16 at N=1000 and m≤50 at N=3000 for α=0.05. Since the deployed pipeline calibrates on roughly 33 held-out responses per pair (App. S5), the m⋆ produced by Eqs. (7) and (12) does not in general meet the stated (α, δ) targets. The paper should state the honest-window size required to certify a given m⋆ and qualify the 1/ε law as finite-ε and calibration-bounded, exactly as App. S4 already does.
- [§6.2; App. S3] The headline detection figure is conditioned on margin-qualified pairs in a way not fully transparent in the body. App. S3 reports that over the 218 margin-qualified pairs the per-pair honest FPR has median 0.000, 95th percentile 0.090, maximum 0.557 (0.39 after exact-binomial hardening), with 7.8% of pairs above 0.05; the body reports only the pooled mean 0.017. The held-out margin-selection power drops to 0.66 vs. 0.85 in-sample; the matched control attributes this to bank size, but it still shows sensitivity to the calibration split. The body should present the FPR distribution, the overall reliable-detection rate (≈0.51 over all 272 ordered pairs), and the caveat that a substantial minority of pairs are returned as indeterminate/low-margin rather than flagged.
minor comments (4)
- [App. S15, S26, S25] Several works cited in the appendices are missing from the reference list: Bruckner (2026), Richter et al. (2025), Shekhar and Ramdas (2023), Dima et al. (2025), Huber (1964), Donoho and Jin (2004), and Ong et al. (2025). Please complete the bibliography.
- [§1, App. S15] The phrase 'first to combine' should be slightly moderated in the main text. The paper itself identifies Bruckner (2026) as concurrent single-token fingerprinting and FLIPS as sharing the random-generation signal; the defensible novelty is the specific combination (reusable probe, budget estimation, dilution ε-estimation, attribution) rather than the underlying signal.
- [App. S5, Table 3] Table 3's 'which statistic backs which claim, and at what guarantee level' is an excellent self-assessment. Consider moving a condensed version of it into the main text so that the abstract's guarantees are immediately qualified.
- [App. S2] The δ-method interval for ε is mildly anti-conservative (empirical coverage 0.86–0.87 at nominal 90%, 0.92–0.93 at 95%). The paper describes this accurately as 'slightly optimistic,' but the main text should at least footnote this so readers do not treat the 90% interval as exact.
Circularity Check
No significant circularity: IRIS's detection, attribution, epsilon-estimation, and budget-size claims are validated on held-out and external data, with the paper's own calibration caveats disclosed as limitations rather than hidden reductions.
full rationale
The paper's central derivation is self-contained. Thresholds and budgets are frozen during Audit-Plan Construction on labeled reference/candidate data, then applied to a disjoint suspect stream; the headline power/FPR numbers are measured on held-out diluted or honest streams, not computed from the same fit that set the thresholds. The epsilon estimator Eq. (14) is a method-of-moments inversion of the mixture model Q_epsilon, and its accuracy is evaluated on controlled mixtures with disjoint enrollment and audit splits. The 1/epsilon dilution budget is derived from a Chernoff bound (Thm. 1) and then empirically measured, with the paper explicitly citing Huber contamination and higher criticism as the classical lineage rather than claiming a new phenomenon. The pilot exponent bIauc is honestly labeled a 'budgeting fit, not an information-theoretic lower bound' (Def. 1), and its predictive accuracy is checked on held-out points. The only possible self-citation (Zhang et al. 2026) is motivational in the introduction and is not load-bearing for any technical claim. The calibration-robustness concerns in App. S4/S5 — power met on 17.8% of feasible pairs at the prescribed budget, FPR tail up to 0.39 post-hardening, and the 1000-sample calibration overshoot — are explicitly disclosed and are correctness/scope limitations, not circularity; Table 3 even labels the prescribed budget a 'feasibility diagnostic, not a coverage guarantee.' No equation in the paper reduces to its own input by construction.
Assumptions & free parameters
free parameters (5)
- pilot rank-error exponent bIauc =
probe-dependent; e.g. near 0 for c2,1 to >1 for c2,16 (Fig. 2a, Fig. 12)
- tail exponent κ =
0.00–0.97 depending on pair/probe (Table 22; distribution in Table 23)
- tell-rate margin threshold =
q − p₀ ≥ 0.15
- honest-tail grid A_resp =
as implemented: a₀ ∈ {0.02, 0.05, 0.1, 0.15} (App. S5)
- null-tell anchor level α₀ =
0.1
assumptions (6)
- domain assumption Random-generation outputs expose backend-specific sampling biases that are stable and discriminative across endpoints.
- domain assumption A trusted reference endpoint P serves the advertised model faithfully and can be sampled under known labels during Audit-Plan Construction.
- domain assumption Independent API calls to the suspect endpoint are i.i.d. draws from Q_ε (no sticky or content-adaptive routing).
- ad hoc to paper The (reference, substitute) pair is measurably separated at a usable threshold (q > α₁ with margin).
- standard math Standard large-deviations machinery: Chernoff information theorem, Cramér, CLT with Lindeberg, Hölder, and the TV ≤ √(1−A²) bound.
- ad hoc to paper A pure decoding-temperature retune is second-order and out of audit scope (Prop. 5).
Cite this review
Pith. "Pith review of Which Model Is Actually Serving You? IRIS: Budgeted Black-Box Auditing of Model Substitution and Routing Dilution in LLM Gateways." pith.science (2026). https://pith.science/paper/L6ICYQUS
@misc{pith2026260720860,
author = {Pith},
title = {Pith review of: Which Model Is Actually Serving You? IRIS: Budgeted Black-Box Auditing of Model Substitution and Routing Dilution in LLM Gateways},
year = {2026},
howpublished = {\url{https://pith.science/paper/L6ICYQUS}},
note = {Machine review of arXiv:2607.20860}
}
abstract
Commercial LLM gateways mediate access to hosted models, but the served backend may not match the advertised one: it may substitute a cheaper model on every request or route only a fraction $\epsilon$ of requests to it. Prior black-box auditors often need a privileged signal (log-probabilities, token ranks, or reference samples) or a target-specific probe, fix the query budget in advance, and return a yes/no verdict. We present $\mathrm{IRIS}$, an audit that needs only the returned text: it asks endpoints to generate random numbers or strings, fingerprints the backend, and is the first to combine, in one text-only audit, detection of whole-stream substitution and fractional dilution, attribution of the served backend, routing-fraction ($\epsilon$) estimation, and a query budget it sizes itself. A cheap pilot fits the exponential query-error decay and freezes that budget before any suspect query is issued. On an intra-family Qwen3 ladder $\mathrm{IRIS}$ verifies the backend at $0.99$ AUROC and sharpens attribution as queries accumulate; across a commercial OpenRouter library it catches $\epsilon{=}0.3$ dilution on margin-qualified pairs at $0.85$ mean power ($0.017$ false-positive rate) and recovers $\epsilon$ to within $0.04$ for enrolled diluents; and a live cross-provider audit flags $14$ of $15$ same-model provider pairs by genuine quantization and kernel deviations, corroborated on third-party MET traces. Against comparable black-box auditors, $\mathrm{IRIS}$ matches or beats detection on shared tasks, and adaptive allocation lifts the matched-budget target-hit rate from $73$% to $87$%. Further experiments cover adversarial gateways, knob identifiability, unseen diluents, and false-positive control.
Figures
Figures from the paper (13 more)
Reference graph
Works this paper leans on
-
[2]
Cai, W.; Shi, T.; Zhao, X.; and Song, D. 2025. Are You Getting What You Pay For? Auditing Model Substitution in LLM API s. arXiv preprint arXiv:2504.04715
arXiv 2025
-
[3]
M.; and Thomas, J
Cover, T. M.; and Thomas, J. A. 2006. Elements of Information Theory. Wiley-Interscience, 2nd edition. ISBN 9780471241959
2006
-
[5]
Donoho, D.; and Jin, J. 2004. Higher Criticism for Detecting Sparse Heterogeneous Mixtures. The Annals of Statistics, 32(3): 962--994
2004
-
[6]
Gao, I.; Liang, P.; and Guestrin, C. 2025. Model Equality Testing: Which Model Is This API Serving? In International Conference on Learning Representations (ICLR)
2025
-
[7]
K.; Renda, A.; and Carbin, M
Hopkins, A. K.; Renda, A.; and Carbin, M. 2023. Can LLM s Generate Random Numbers? Evaluating LLM Sampling in Controlled Domains. In ICML Workshop on Sampling and Optimization in Discrete Space (SODS)
2023
-
[8]
Huber, P. J. 1964. Robust Estimation of a Location Parameter. The Annals of Mathematical Statistics, 35(1): 73--101
1964
-
[9]
Nikolic, I.; Baluta, T.; and Saxena, P. 2025. Model Provenance Testing for Large Language Models. In Advances in Neural Information Processing Systems, volume 38, 34126--34153
2025
-
[10]
M.; Di Ferrante, M.; Pazdera, A.; Garner, R.; Jaghouar, S.; Basra, M.; Ryabinin, M.; and Hagemann, J
Ong, J. M.; Di Ferrante, M.; Pazdera, A.; Garner, R.; Jaghouar, S.; Basra, M.; Ryabinin, M.; and Hagemann, J. 2025. TOPLOC : A Locality Sensitive Hashing Scheme for Trustless Verifiable Inference. arXiv preprint arXiv:2501.16007
arXiv 2025
Show all 59 references
-
[11]
M.; and Ateniese, G
Pasquini, D.; Kornaropoulos, E. M.; and Ateniese, G. 2025. LLMmap : Fingerprinting For Large Language Models. In 34th USENIX Security Symposium (USENIX Security 25), 299--318
2025
-
[12]
Richardeau, G.; Dashyan, G.; Le Merrer, E.; and Tredan, G. 2026. FLIPS : Instance-Fingerprinting for LLM s via Pseudo-random Sequences. In International Conference on Machine Learning (ICML)
2026
-
[13]
Richter, L.; He, X.; Minervini, P.; and Kusner, M. J. 2025. An Auditing Test to Detect Behavioral Shift in Language Models. In International Conference on Learning Representations (ICLR)
2025
-
[14]
Shekhar, S.; and Ramdas, A. 2023. Nonparametric Two-Sample Testing by Betting. IEEE Transactions on Information Theory
2023
-
[15]
Z.; and Liu, Z
Sun, M.; Yin, Y.; Xu, Z.; Kolter, J. Z.; and Liu, Z. 2025. Idiosyncrasies in Large Language Models. In Proceedings of the 42nd International Conference on Machine Learning, volume 267 of Proceedings of Machine Learning Research, 57854--57885. PMLR
2025
-
[16]
Zhao, M.; Du, Y.; and Wang, M. 2026. Large Language Models Are Bad Dice Players: LLM s Struggle to Generate Random Numbers from Statistical Distributions. In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 22942-...
2026
-
[17]
Your ``Pro''
Zhang, Jiahao and Li, Xiuyu and Wang, Suhang , journal=. Your ``Pro''. 2026 , url =
2026
-
[18]
Token-Efficient Change Detection in
Chauvin, Timoth\'ee and Lalanne, Cl\'ement and Le Merrer, Erwan and Loubes, Jean-Michel and Ta\"iani, Fran. Token-Efficient Change Detection in. International Conference on Machine Learning (ICML) , year =
-
[19]
2026 , url =
Richardeau, Gurvan and Dashyan, Gohar and Le Merrer, Erwan and Tredan, Gilles , booktitle =. 2026 , url =
2026
-
[20]
2026 , url =
Fang, Yijia and Feng, Yiqing and Li, Bingyu and Zhou, Mingxun , journal=. 2026 , url =
2026
-
[21]
Auditing Black-Box
Zhu, Xiaoyuan and Ye, Yaowen and Qiu, Tianyi and Zhu, Hanlin and Tan, Sijun and Mannan, Ajraf and Michala, Jonathan and Popa, Raluca Ada and Neiswanger, Willie , booktitle =. Auditing Black-Box. 2026 , url =
2026
-
[22]
Model Equality Testing: Which Model Is This
Gao, Irena and Liang, Percy and Guestrin, Carlos , booktitle =. Model Equality Testing: Which Model Is This. 2025 , url =
2025
-
[23]
and Ateniese, Giuseppe , booktitle =
Pasquini, Dario and Kornaropoulos, Evgenios M. and Ateniese, Giuseppe , booktitle =. 2025 , url =
2025
-
[24]
doi:10.18653/v1/2024.findings-acl.683 , year =
Gubri, Martin and Ulmer, Dennis and Lee, Hwaran and Yun, Sangdoo and Oh, Seong Joon , booktitle =. doi:10.18653/v1/2024.findings-acl.683 , year =
2024 doi
-
[25]
Advances in Neural Information Processing Systems , volume =
Model Provenance Testing for Large Language Models , author =. Advances in Neural Information Processing Systems , volume =. 2025 , url =
2025
-
[26]
and Renda, Alex and Carbin, Michael , booktitle =
Hopkins, Aspen K. and Renda, Alex and Carbin, Michael , booktitle =. Can. 2023 , url =
2023
-
[27]
Are You Getting What You Pay For? Auditing Model Substitution in
Cai, Will and Shi, Tianneng and Zhao, Xuandong and Song, Dawn , journal=. Are You Getting What You Pay For? Auditing Model Substitution in. 2025 , url =
2025
-
[28]
The Annals of Mathematical Statistics , volume =
A Measure of Asymptotic Efficiency for Tests of a Hypothesis Based on the Sum of Observations , author =. The Annals of Mathematical Statistics , volume =. 1952 , doi =
1952
-
[29]
The Annals of Mathematical Statistics , volume =
Asymptotically Optimal Tests for Multinomial Distributions , author =. The Annals of Mathematical Statistics , volume =. 1965 , doi =
1965
-
[30]
2006 , doi =
Elements of Information Theory , author =. 2006 , doi =
2006
-
[31]
Large Language Models Are Bad Dice Players:
Zhao, Minda and Du, Yilun and Wang, Mengyu , booktitle =. Large Language Models Are Bad Dice Players:. 2026 , url =
2026
-
[32]
Behavioral Consistency and Transparency Analysis on Large Language Model
Lin, Guanjie and Wan, Yinxin and Pei, Shichao and Xu, Ting and Xu, Kuai and Xue, Guoliang , booktitle =. Behavioral Consistency and Transparency Analysis on Large Language Model. doi:10.1145/3777912.3809156 , year =
-
[33]
Real Money, Fake Models: Deceptive Model Claims in Shadow
Zhang, Yage and Jiang, Yukun and Chen, Zeyuan and Backes, Michael and Shen, Xinyue and Zhang, Yang , journal=. Real Money, Fake Models: Deceptive Model Claims in Shadow. 2026 , url =
2026
-
[34]
Your Agent Is Mine: Measuring Malicious Intermediary Attacks on the
Liu, Hanzhi and Shou, Chaofan and Wen, Hongbo and Chen, Yanju and Fang, Ryan Jingyang and Feng, Yu , journal=. Your Agent Is Mine: Measuring Malicious Intermediary Attacks on the. 2026 , url =
2026
-
[35]
Incompressible Knowledge Probes: Estimating Black-Box
Li, Bojie , journal=. Incompressible Knowledge Probes: Estimating Black-Box. 2026 , url =
2026
-
[36]
doi:10.18653/v1/2026.findings-eacl.273 , year =
Yan, Yuliang and Tang, Haochun and Yan, Shuo and Dai, Enyan , booktitle =. doi:10.18653/v1/2026.findings-eacl.273 , year =
2026 doi
-
[38]
, journal=
Bradley, William F. , journal=. 2024 , url =
2024
-
[39]
Log Probability Tracking of
Chauvin, Timoth\'ee and Le Merrer, Erwan and Ta\"iani, Fran. Log Probability Tracking of. International Conference on Learning Representations (ICLR) , year =
-
[40]
Proceedings of the 42nd International Conference on Machine Learning , series =
Idiosyncrasies in Large Language Models , author =. Proceedings of the 42nd International Conference on Machine Learning , series =. 2025 , url =
2025
-
[41]
arXiv preprint arXiv:2607.10252 , year =
One Token Is Enough: Fingerprinting and Verifying Large Language Models from Single-Token Output Distributions , author =. arXiv preprint arXiv:2607.10252 , year =
-
[42]
International Conference on Learning Representations (ICLR) , year =
An Auditing Test to Detect Behavioral Shift in Language Models , author =. International Conference on Learning Representations (ICLR) , year =
-
[43]
arXiv preprint arXiv:2504.12335 , year =
You've Changed: Detecting Modification of Black-Box Large Language Models , author =. arXiv preprint arXiv:2504.12335 , year =
-
[44]
2025 , url =
Ong, Jack Min and Di Ferrante, Matthew and Pazdera, Aaron and Garner, Ryan and Jaghouar, Sami and Basra, Manveer and Ryabinin, Max and Hagemann, Johannes , journal=. 2025 , url =
2025
-
[45]
The Annals of Mathematical Statistics , volume =
Robust Estimation of a Location Parameter , author =. The Annals of Mathematical Statistics , volume =. 1964 , doi =
1964
-
[46]
The Annals of Statistics , volume =
Higher Criticism for Detecting Sparse Heterogeneous Mixtures , author =. The Annals of Statistics , volume =. 2004 , doi =
2004
-
[47]
IEEE Transactions on Information Theory , year =
Nonparametric Two-Sample Testing by Betting , author =. IEEE Transactions on Information Theory , year =
-
[48]
Bradley, W. F. 2024. LLM s and the Madness of Crowds. arXiv preprint arXiv:2411.01539
2024 arXiv
-
[49]
Chauvin, T.; Lalanne, C.; Le Merrer, E.; Loubes, J.-M.; Ta\"iani, F.; and Tredan, G. 2026 a . Token-Efficient Change Detection in LLM API s. In International Conference on Machine Learning (ICML)
2026
-
[50]
Chauvin, T.; Le Merrer, E.; Ta\"iani, F.; and Tredan, G. 2026 b . Log Probability Tracking of LLM API s. In International Conference on Learning Representations (ICLR)
2026
-
[51]
Chernoff, H. 1952. A Measure of Asymptotic Efficiency for Tests of a Hypothesis Based on the Sum of Observations. The Annals of Mathematical Statistics, 23(4): 493--507
1952
-
[52]
Fang, Y.; Feng, Y.; Li, B.; and Zhou, M. 2026. KBF : Knowledge Boundary as Fingerprint for Language Model and Black-Box API Auditing. arXiv preprint arXiv:2605.29524
2026 arXiv
-
[53]
Gubri, M.; Ulmer, D.; Lee, H.; Yun, S.; and Oh, S. J. 2024. TRAP : Targeted Random Adversarial Prompt Honeypot for Black-Box Identification. In Findings of the Association for Computational Linguistics: ACL 2024, 11496--11517. Bangkok, Thailand: Association for Computational L...
2024
-
[54]
Hoeffding, W. 1965. Asymptotically Optimal Tests for Multinomial Distributions. The Annals of Mathematical Statistics, 36(2): 369--401
1965
-
[55]
Li, B. 2026. Incompressible Knowledge Probes: Estimating Black-Box LLM Parameter Counts via Factual Capacity. arXiv preprint arXiv:2604.24827
2026 arXiv
-
[56]
Lin, G.; Wan, Y.; Pei, S.; Xu, T.; Xu, K.; and Xue, G. 2026. Behavioral Consistency and Transparency Analysis on Large Language Model API Gateways. In Proceedings of the 2026 ACM Internet Measurement Conference (IMC '26)
2026
-
[57]
J.; and Feng, Y
Liu, H.; Shou, C.; Wen, H.; Chen, Y.; Fang, R. J.; and Feng, Y. 2026. Your Agent Is Mine: Measuring Malicious Intermediary Attacks on the LLM Supply Chain. arXiv preprint arXiv:2604.08407
2026 arXiv
-
[58]
P e zik, P.; Kaczy \'n ski, K.; Szyma \'n ska, M.; \.Z arnecki, F.; Deckert, Z.; Kwiatkowski, J.; and Janowski, W. 2025. LLMLagBench : Identifying Temporal Training Boundaries in Large Language Models. arXiv preprint arXiv:2511.12116
2025
-
[59]
Yan, Y.; Tang, H.; Yan, S.; and Dai, E. 2026. DuFFin : A Dual-Level Fingerprinting Framework for LLM s IP Protection. In Findings of the Association for Computational Linguistics: EACL 2026, 5168--5184. Rabat, Morocco: Association for Computational Linguistics
2026
-
[60]
Zhang, J.; Li, X.; and Wang, S. 2026. Your ``Pro'' LLM Subscription May Actually Be ``Free'': Exposing Fingerprint Spoofing Risks in LLM Inference Services. arXiv preprint arXiv:2606.16100
2026
-
[61]
Zhang, Y.; Jiang, Y.; Chen, Z.; Backes, M.; Shen, X.; and Zhang, Y. 2026. Real Money, Fake Models: Deceptive Model Claims in Shadow API s. arXiv preprint arXiv:2603.01919
2026
-
[62]
A.; and Neiswanger, W
Zhu, X.; Ye, Y.; Qiu, T.; Zhu, H.; Tan, S.; Mannan, A.; Michala, J.; Popa, R. A.; and Neiswanger, W. 2026. Auditing Black-Box LLM API s with a Rank-Based Uniformity Test. In International Conference on Learning Representations (ICLR)
2026
Reviewed August 1, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.