Uncertainty-Aware Sequential Decision Rules for Event-Triggered LLM Invocation in Streaming Systems

Zhaohui Wang

REVIEW 4 major objections 5 minor 24 references

Uncertainty-Aware Sequential Decision Rules for Event-Triggered LLM Invocation in Streaming Systems

T0 review · 4 major / 5 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read A threshold on a risk functional — not a learned router or periodic sampling — determines when to invoke a large language model in a streaming pipeline, with proven bounds on inter-event time, regret, and calibration.

desk verdict Useful empirical study wrapped in an overclaimed theory: three of six proofs are absent and the keystone submartingale assumption fails, but the experiments and the calibration bound are worth engaging with. read the letter →

arxiv 2607.13048 v1 pith:DPGVRP7Z submitted 2026-06-27 cs.LG cs.AIstat.ML

Zhaohui Wang This is my paper · ORCID

classification cs.LGcs.AIstat.ML MSC 62L1062L1568T05

keywords event-triggeredcontrolLLMinvocationsequentialdecision-makinguncertaintycalibrationstreaminginferenceregretboundsriskfunctionaloptimalstopping

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

The paper tries to establish that the decision of when to call an LLM in a streaming inference pipeline can be reduced to a threshold policy on a risk functional: a scalar number computed from the recent history of a cheap model's predictions and uncertainties. It proves six results under this reduction: a minimum spacing between invocations, optimality of threshold policies, approximate SPRT guarantees, sublinear regret in stationary and changing streams, convergence of online gradient descent for adaptive thresholds, and a bound linking miscalibrated uncertainty to missed critical events. If these results hold, practitioners can budget LLM calls by thresholding accumulated risk, with formal guarantees on cost, missed-event rate, and adaptation to drift. Empirically, the paper reports that anomaly-score risk functions dominate alternatives by roughly an order of magnitude on the invocation–miss Pareto curve, and that all principled triggers show sublinear regret.

What carries the argument

The risk functional R(H_t) — a scalar aggregation of the fast model's anomaly score, predictive uncertainty, and temporal context — together with the threshold policy pi_theta(H_t) = 1[R(H_t) >= theta]. This mechanism unifies diverse trigger families and carries the theoretical results: the threshold is the decision variable learned online, the risk signal is the object with bounded increments and submartingale behavior, and sweeping the threshold traces the invocation–miss Pareto frontier.

What would settle it

Compute the mean of E[R_{t+1}|H_t] - R_t over all degraded timesteps on a large, diverse set of streaming datasets; if the mean is negative or significantly non-positive, the submartingale assumption is violated and the claimed threshold-policy optimality collapses. The paper's own CMAPSS test gives mean +0.006 with p = 0.20, so a larger, more sensitive test is the natural arbiter.

Watch

Extended reading notes

Core claim

The central claim is that every common LLM invocation rule — event-triggered, optimal stopping, SPRT, CUSUM, Bayesian — is a special case of the threshold policy pi_theta(H_t) = 1[R(H_t) >= theta], where R aggregates anomaly score, uncertainty, and temporal context. Under conditions including bounded increments of the risk signal and a submartingale property in degraded states, the paper proves that such policies have a minimum inter-invocation time, are optimal via smooth pasting, achieve O(sqrt(T log T)) regret in stationary streams (extended to O(sqrt((C_T+1) T log T)) with changepoints), and transfer calibration error linearly into miss probability. The empirical evaluation on turbofan d

Load-bearing premise

The submartingale condition — that in degraded states the expected risk tomorrow is at least today's risk — is the load-bearing premise for the optimality of threshold policies; if it fails, the smooth-pasting result does not apply.

Editorial extensions

If this is right

A streaming system can deploy a single threshold policy and automatically inherit a minimum inter-event time, preventing trigger chattering, while the regret bound guarantees the adaptive threshold approaches the best fixed threshold at rate O(sqrt(T log T)).
Calibrating the fast model's uncertainty directly tightens the upper bound on missed critical events, giving a deployment-ready safety budget.
The optimal threshold is stable across two orders of magnitude of LLM cost, so moderate price changes do not require re-tuning.
Anomaly-score-driven risk functionals are a strong default: they require far fewer invocations for the same miss rate than uncertainty-only or linear-combination alternatives.
The unified formulation lets a single online adaptation mechanism (OGD or LinUCB) tune any trigger family, including classical SPRT and CUSUM, in the presence of drift or shift.
A minimum inter-event time bound prevents trigger chattering in LLM-invocation systems, a practical concern that classical event-triggered control did not address in this setting.

Reading between the lines

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

The same risk-threshold abstraction could extend to other expensive-oracle settings, such as human expert review or costly physics simulations, where the oracle is invoked rarely but with high value per call.
The paper's own verification of the submartingale assumption is weak (p = 0.20), so the optimality claim rests on a premise that real degraded streams may only approximately satisfy; a fully distribution-free analysis would remove this fragility.
The calibration-to-miss-rate bound holds only for the uncertainty-threshold component, not for composite risk functionals; a composite safety budget would need a separate calibration treatment.
The strong dominance of anomaly score over uncertainty on CMAPSS may be domain-dependent; late-degradation or concept-drift regimes could reverse it, and the paper leaves that test to future work.

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.

Desk Editor's Note

read the letter

Quick take: this is a solid empirical paper wearing a theorem coat that doesn't quite fit. The six theorems are mostly classical—the paper admits in §8 that it offers no new stopping rules—and three of them (Prop. 1, Thm. 2, Thm. 4) have no proofs in the visible appendix. The keystone Prop. 1 rests on a submartingale assumption that the authors' own test fails to support (p=0.20, rated “Weak”). So don't cite this as new theory. But the empirical core is genuine: real LLM calls, a grounding rubric, failure-mode analysis, cost sensitivity, assumption checks, and a clean ablation of risk function designs. The result that anomaly-score-only risk (R2) dominates alternatives by an order of magnitude in Pareto AUC is practically useful, if not shocking. The calibration-to-miss-rate inequality (Thm. 5) is simple, proved, and correct.

The soft spots, in proportion: (1) Missing proofs for three of six theorems is not minor when the abstract says “we prove six results.” Either the proofs appear or the claim gets scaled back. (2) The regret “confirmation” is an in-sample fit of α in R(T)~cT^α; that estimates sublinearity, it doesn't confirm the bound. (3) The “critical event”/miss definition is never operationalized in the visible text, and every miss-rate, Pareto-AUC, and c_miss result depends on it. (4) The baseline comparison uses default hyperparameters for classical triggers, which the paper itself says understates them; the tuned-threshold results should be headline or the caveat should be stronger. (5) The submartingale premise is not significant, so Prop. 1 is conditional at best.

None of this sinks the paper. The framework works as design guidance, the limitations section is candid (the weak submartingale is acknowledged), and the reproducibility artifacts—code, seeds, cached results—are a real plus. The citation pattern is clean; nothing looks like self-promotion. The right outcome is major revision, not rejection.

I'd send this to a competent referee, mainly to chase the missing proofs and the empirical protocol. Not a desk reject. For a reading group it's a good case study in honest limitation-writing.

Referee Report

4 major / 5 minor

Summary. The paper formalizes the decision of when to invoke an LLM in a streaming inference pipeline as a threshold policy on a risk functional R(H_t), with the trigger firing when R exceeds θ. It claims six theoretical results: a minimum inter-event-time bound; optimality of threshold policies via smooth pasting; approximate SPRT guarantees with estimated parameters; O(√(T log T)) regret under stationarity with an extension to non-stationary streams; O(1/√T) convergence of OGD for adaptive thresholds; and a calibration-to-miss-rate transfer inequality. The empirical study on CMAPSS and CIC-IDS2017 verifies assumptions, ablates eight risk-function designs, compares against six baselines including RouteLLM and a contextual bandit, evaluates real LLM diagnoses, and analyzes cost sensitivity. The central empirical claim is that anomaly-score-driven risk functions dominate alternatives by roughly an order of magnitude in Pareto AUC, and that adaptive triggers achieve sublinear regret.

Significance. If the theoretical results were fully established and the empirical verification were stronger, the paper would make a useful contribution: it connects classical sequential decision theory to LLM invocation, provides a unified trigger framework, and ships a reproducible artifact with per-seed JSON, prompts, and scripts. The honest reporting of assumption checks — including a weak submartingale result and approximate density fits — is commendable, as is the inclusion of a limitations section. The practical finding that anomaly-score-only risk (R2) vastly outperforms uncertainty-only risk (R3) on CMAPSS is a useful caution for practitioners. However, the paper's headline claim of 'proving six results' is not supported by the visible manuscript: only three proofs appear in Appendix A, and the keystone threshold-optimality result remains unproved and rests on a premise that the paper's own data do not significantly support.

major comments (4)

[§3.2, §3.5, Appendix A] The manuscript states in §3 that proofs are deferred to Appendix A, but Appendix A contains proofs only for Theorem 1, Theorem 3, and Theorem 5. Proposition 1 (threshold-policy optimality), Theorem 2 (approximate SPRT), and Theorem 4 (OGD convergence) are stated in the main text without proofs. Proposition 1 is load-bearing: it justifies the entire restriction to policies π_θ(H_t)=1[R(H_t)≥θ]. As written, the claim 'we prove six results' overstates what is established. The authors must either supply the missing proofs or revise the contribution statement to 'state six results and prove three'.
[§5, Table 2(a), Proposition 1] The key premise of Proposition 1 is the submartingale condition E[R_{t+1}|H_t] ≥ R_t in degraded states. The paper's own verification on CMAPSS FD001 reports a mean increment of +0.006, only 52.4% positive increments, and a one-sided t-test p=0.20, graded 'Weak' in Table 2(a) and 'not significant' in the text. This is not enough to establish the premise. Since Proposition 1 is the theoretical justification for threshold policies, the central theoretical claim is unsupported unless a proof is supplied and the premise is either strengthened empirically or replaced by a weaker condition that is actually verified.
[§6.1, Table 4] The claim that the results 'confirm sublinear regret' rests on fitting a scaling exponent α in R(T) ~ cT^α to the empirical cumulative regret curves. The regret benchmark is never defined: Theorem 3 defines regret against the best fixed threshold in hindsight, but Table 4 reports 'Regret' and α without stating the baseline or how the cumulative regret is computed. As reported, α is a free parameter of a curve fit, so α<1 is a fitted descriptive statistic, not a confirmation of the theorem. The paper should report the actual benchmark and the construction of the regret curve, and should distinguish a fitted exponent from a verified bound.
[§3.2, §3.7] The 'unified view' in §3.7 states that all trigger families are instances of threshold policies because each is defined by thresholding its own statistic. This is true by construction but does not by itself provide optimality or regret guarantees. The paper is transparent about this, but the abstract's claim that the framework is 'risk-based sequential stopping' should be read with the caveat that the optimality of the threshold class is unproved (see previous comments).

minor comments (5)

[Abstract and §3] The phrase 'we prove six results' should be changed to match the actual proof content of the manuscript.
[§6.2, Fig. 2] The caption says 'AUC (lower is better)' but the text explains it is the area under the Pareto curve of invocation rate vs. miss rate. Define the AUC explicitly to avoid confusion with classifier AUC.
[§6.4, Fig. 3] The 'Threshold' baseline is the default event-triggered (hidden-state) trigger at fixed hyperparameters, not the swept anomaly-score threshold R2. This distinction is made in the text, but the figure and table would benefit from a separate label to prevent misreading.
[§5, Table 2(a)] The stationarity check uses a two-sample KS test between the first and second halves of the risk sequence in the normal regime. This is a weak test of within-regime stationarity; the paper should acknowledge that it does not rule out more fine-grained non-stationarity.
[Throughout] There are minor typographical issues (e.g., 'T able 1', 'Appendinx') and formatting inconsistencies in the references. A proofreading pass is needed.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: theoretical results are independent mathematical statements; empirical checks are in-sample but not definitionally forced.

full rationale

The paper's six theoretical results are stated as mathematical consequences of explicit assumptions (bounded increments, submartingale, stationarity, epsilon-calibration, convexity). None of these results is defined in terms of the quantity it is said to predict; for example, Theorem 5 is a direct bound derived from the definition of epsilon-calibration, not a restatement of that definition as a prediction. The 'unified view' that SPRT, CUSUM, and other triggers are threshold policies is a transparent reformulation (any trigger comparing a scalar statistic to a threshold is a threshold policy), and the paper does not use this reformulation to derive its theorems; it is organizational. The empirical verification of assumptions on CMAPSS FD001 and the subsequent regret/calibration experiments use the same dataset, but the fitted scaling exponent alpha is a descriptive summary of measured regret compared with a theoretical rate, not a parameter fitted to data and then renamed as a prediction. The weak submartingale evidence (p=0.20) and the absence of proofs for Prop. 1, Thm. 2, and Thm. 4 in the visible appendix are correctness/completeness concerns, not circularity. No load-bearing self-citation chain or imported uniqueness theorem appears; references are to standard external results. The paper itself flags the submartingale as 'not significant' and the framework as 'design guidance rather than an exact description,' which further shows the claims are not presented as forced by the data.

Assumptions & free parameters 5 free parameters · 7 assumptions · 0 invented entities

The theorems themselves contain no fitted constants — L_s in Thm 1 is measured, ε in Thm 5 is a bound input, and the regret exponents are empirical rather than part of any derivation. The framework's cost is instead in domain assumptions (submartingale, bounded increments, stationarity, monotonicity, continuity) that the paper itself verifies only partially; one ad-hoc assumption (convexity of the threshold loss for OGD) is asserted without proof; and the empirical evaluation quietly assumes a critical-event definition the visible text never states. No new ontological entities are introduced — the risk functional and grounding rubric are design artifacts, not postulates needing independent evidence.

free parameters (5)

Regret-curve fit (c, α) in R(T) ~ cT^α = α = 0.47–0.74 depending on trigger
Fitted from empirical cumulative regret on the same CMAPSS runs; used to claim “confirmation” of the O(√T) theory and the abstract's “sublinear regret” headline.
Measured increment bound L_s = L99 = 4.50; mean 0.267
Measured from the FD001 risk sequence; instantiates Theorem 1's bound rather than being fitted to force agreement (empirical inter-event ratios 1.6–2.2× safely exceed the bound).
Risk-variant design weights (R1: α=β=1; R6 EWMA λ=0.9; R4 product; R5 max) = α=β=1; λ=0.9
Arbitrary chosen constructions among the eight ablated risk functions; the ablation exists precisely to compare them.
Optimal stopping discount γ = 0.99
Hyperparameter for one trigger family, disclosed in App. B.
Baseline trigger hyperparameters = CUSUM k=0.5,h=5.0; SPRT α=0.05,β=0.10; cooldown 5; evidence window 3; OGD lr=0.01
Fixed defaults per App. B; the paper notes that untuned classical triggers underperform, so these settings shape the headline baseline comparison.

assumptions (7)

domain assumption Submartingale risk under anomaly: E[R_{t+1}|H_t] ≥ R_t in degraded states
Required for Prop. 1 threshold optimality; verified with mean increment +0.006 and p=0.20 (not significant) — paper's own verdict “Weak” (§5, Table 2a).
domain assumption Bounded increments |s_{t+1} − s_t| ≤ L_s of the trigger signal
Required for Theorem 1; 97.3% of increments within 2·L95 but heavy-tailed (KS rejects exponential, p<10^-20), rated “Empir.” (§5, Table 2a).
domain assumption Within-regime stationarity of the risk process
Backs Theorem 3's regret bound; KS p=0.283 between halves of the risk sequence for RUL≥50, while regret is evaluated over runs that include degradation (§5).
domain assumption Continuous density and eventual monotonicity of risk toward failure
Required for Prop. 1; log-normal fit only “Partial”, monotonicity supported (Spearman ρ=−0.356, 82.2% of bins) (§5, Table 2a).
ad hoc to paper Convexity / bounded subgradients of the threshold-selection loss for OGD (Thm 4)
The standard OGD theorem is invoked, but the threshold loss πθ=1[R≥θ] is a step function in θ; its expected version is not shown to be convex (§3.5).
domain assumption An operational definition of “critical event”/miss exists and is independent of the trigger signal
Every miss-rate, Pareto-AUC, and c_miss result depends on it, yet the visible text never states the definition (RUL<50 is used for degraded-regime analysis in §5).
domain assumption Uniform (worst-case over u) ε-calibration for Theorem 5
The bound requires uniform calibration error; only ECE (average) is measured, and the paper explicitly concedes ECE is not a substitute (§3.6).

how reviews work

0 comments

Cite this review

Pith. "Pith review of Uncertainty-Aware Sequential Decision Rules for Event-Triggered LLM Invocation in Streaming Systems." pith.science (2026). https://pith.science/paper/DPGVRP7Z

@misc{pith2026260713048,
  author       = {Pith},
  title        = {Pith review of: Uncertainty-Aware Sequential Decision Rules for Event-Triggered LLM Invocation in Streaming Systems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DPGVRP7Z}},
  note         = {Machine review of arXiv:2607.13048}
}

read the original abstract

Streaming inference pipelines increasingly pair lightweight fast models with Large Language Models (LLMs) that provide rich semantic understanding at substantial cost. The central question of when to invoke the LLM has received limited formal treatment. We cast this as a risk-based sequential stopping problem, where a trigger policy fires when a risk functional over the observation history exceeds a threshold. Within this framework, we prove six results: a minimum inter-event time bound excluding trigger chattering; optimality of threshold policies via smooth pasting; approximate SPRT guarantees under estimated parameters; O(sqrt(T log T)) regret for stationary streams, extending to O(sqrt((C_T + 1) T log T)) under C_T changepoints; O(1/sqrt(T)) convergence of online gradient descent for adaptive thresholds; and a calibration-to-miss-rate transfer inequality. Several classical trigger families, including event-triggered, optimal stopping, SPRT, CUSUM, and Bayesian triggers, can be expressed as special cases of this framework. On turbofan degradation data (CMAPSS) with real LLM calls, we empirically verify the theoretical assumptions, ablate the risk function design, compare against six baselines including a RouteLLM-style router and contextual bandits, and analyze cost sensitivity and LLM failure modes. The results confirm sublinear regret, with alpha < 1 for all principled triggers; high diagnostic quality, with 92.9 percent of 1600 LLM diagnoses reaching grounding score >= 0.75 under our rubric; and that anomaly-score-driven risk functions dominate alternatives by roughly an order of magnitude on the Pareto AUC.

Figures

Figures reproduced from arXiv: 2607.13048 by the authors.

**Figure 1.** System overview. A fast model processes every streaming observation, producing predictions and calibrated uncertainty. The trigger policy fires when R(Ht) ≥ θt, invoking the LLM on a small subset of timesteps, with the rate controlled by the trigger threshold. An adaptive feedback loop updates θt via OGD or LinUCB. The design objective balances invocation cost against the probability of missing a critical event: min… view at source ↗

**Figure 2.** Risk function ablation on CMAPSS FD001 (5 seeds). AUC: area under the Pareto curve (invocation rate vs. miss rate; lower is better). R2 (anomaly score alone) dominates all alternatives by an order of magnitude. exploited by LinUCB (uncertainty level and recent anomaly statistics) explains its advantage over OGD’s scalar updates. 6.2 Risk Function Ablation The choice of risk functional R(Ht) is a design decision that… view at source ↗

**Figure 3.** Baseline comparison on CMAPSS FD001 (5 seeds, cLLM=1, cmiss=10). Marker color and shape encode method family; total cost is annotated next to each point. LinUCB (bandit) achieves the best invocation–miss-rate tradeoff, closest to the ideal region (shaded). Periodic and random sampling are ineffective despite similar invocation budgets. the evaluation cost model. Unlike the proposed threshold policies, it imposes no … view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: Real LLM grounding scores on CMAPSS subsets (MiniMax-M2.5) with eventtriggered invocation. All configurations exceed the 0.75 good-grounding threshold (dashed line). Threshold triggers achieve the highest and most consistent grounding across subsets; CUSUM shows more …

**Figure 5.** Figure 5: (a) Pareto curves of invocation rate vs. miss rate for four risk function variants (lines, sweeping θ) and four baselines (scatter points). R2 (anomaly score) dominates all alternatives. (b) Optimal threshold θ ∗ as a function of LLM cost cLLM; the shaded region marks …

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

24 extracted references · 1 canonical work pages

[1]

In: Advances in Neural Information Processing Systems (NeurIPS)

Chen, L., Zaharia, M., Zou, J.: FrugalML: How to use ML prediction APIs more accurately and cheaply. In: Advances in Neural Information Processing Systems (NeurIPS). vol. 33, pp. 10685–10696 (2020)

2020
[2]

In: Proceedings of the 33rd International Conference on Machine Learning (ICML)

Gal, Y., Ghahramani, Z.: Dropout as a Bayesian approximation: Representing model uncertainty in deep learning. In: Proceedings of the 33rd International Conference on Machine Learning (ICML). pp. 1050–1059 (2016), arXiv:1506.02142

arXiv 2016
[3]

arXiv preprint arXiv:1603.08983 (2016)

Graves, A.: Adaptive computation time for recurrent neural networks. arXiv preprint arXiv:1603.08983 (2016)

arXiv 2016
[4]

In: International Conference on Machine Learning (ICML)

Guo, C., Pleiss, G., Sun, Y., Weinberger, K.Q.: On calibration of modern neural networks. In: International Conference on Machine Learning (ICML). pp. 1321–1330 (2017)

2017
[5]

Foundations and Trends in Optimization2(3-4), 157–325 (2016).https://doi.org/10.1561/2400000013

Hazan, E.: Introduction to online convex optimization. Foundations and Trends in Optimization2(3-4), 157–325 (2016).https://doi.org/10.1561/2400000013

doi:10.1561/2400000013 2016
[6]

In: IEEE 51st Conference on Decision and Con- trol (CDC)

Heemels, W.P.M.H., Johansson, K.H., Tabuada, P.: An introduction to event- triggered and self-triggered control. In: IEEE 51st Conference on Decision and Con- trol (CDC). pp. 3270–3285 (2012).https://doi.org/10.1109/CDC.2012.6425820 18 Z. Wang

arXiv 2012
[7]

In: Advances in Neural Information Processing Systems (NeurIPS)

Khosla, P., Teterwak, P., Wang, C., Sarna, A., Tian, Y., Isola, P., Maschinot, A., Liu, C., Krishnan, D.: Supervised contrastive learning. In: Advances in Neural Information Processing Systems (NeurIPS). vol. 33, pp. 18661–18673 (2020)

2020
[8]

In: International Conference on Machine Learning (ICML)

Kuleshov, V., Fenner, N., Ermon, S.: Accurate uncertainties for deep learning using calibrated regression. In: International Conference on Machine Learning (ICML). pp. 2796–2804 (2018)

2018

Show all 24 references

[9]

In: Advances in Neural Information Processing Systems (NeurIPS)

Lakshminarayanan, B., Pritzel, A., Blundell, C.: Simple and scalable predictive uncertainty estimation using deep ensembles. In: Advances in Neural Information Processing Systems (NeurIPS). vol. 30 (2017), arXiv:1612.01474

2017 arXiv
[10]

In: International Conference on Machine Learning (ICML)

Leviathan, Y., Kalman, M., Matias, Y.: Fast inference from transformers via speculative decoding. In: International Conference on Machine Learning (ICML). pp. 19274–19286 (2023)

2023
[11]

In: International Conference on World Wide Web (WWW)

Li, L., Chu, W., Langford, J., Schapire, R.E.: A contextual-bandit approach to per- sonalized news article recommendation. In: International Conference on World Wide Web (WWW). pp. 661–670 (2010).https://doi.org/10.1145/1772690.1772758

2010
[12]

In: Proceedings of the AAAI Conference on Artificial Intelligence

Naeini, M.P., Cooper, G.F., Hauskrecht, M.: Obtaining well calibrated probabilities using Bayesian binning into quantiles. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 29 (2015).https://doi.org/10.1609/aaai.v29i1.9602

2015 doi
[13]

arXiv preprint arXiv:2406.18665 (2024)

Ong, I., Almahairi, A., Wu, V., Chiang, W.L., Wu, T., Gonzalez, J.E., Kadous, M.W., Stoica, I.: RouteLLM: Learning to route LLMs with preference data. arXiv preprint arXiv:2406.18665 (2024)

2024 arXiv
[14]

Lec- tures in Mathematics ETH Zürich, Birkhäuser (2006).https://doi.org/10.1007/ 978-3-7643-7390-0

Peskir, G., Shiryaev, A.: Optimal Stopping and Free-Boundary Problems. Lec- tures in Mathematics ETH Zürich, Birkhäuser (2006).https://doi.org/10.1007/ 978-3-7643-7390-0

2006
[15]

In: International Conference on Prognostics and Health Management

Saxena, A., Goebel, K., Simon, D., Eklund, N.: Damage propagation modeling for aircraft engine run-to-failure simulation. In: International Conference on Prognostics and Health Management. pp. 1–9 (2008).https://doi.org/10.1109/PHM.2008. 4711414

2008 doi
[16]

In: Advances in Neural Information Processing Systems (NeurIPS)

Schuster, T., Fisch, A., Gupta, J., Dehghani, M., Bahri, D., Tran, V.Q., Tay, Y., Metzler, D.: Confident adaptive language modeling. In: Advances in Neural Information Processing Systems (NeurIPS). vol. 35, pp. 17456–17472 (2022)

2022
[17]

In: International Conference on Information Systems Security and Privacy (ICISSP)

Sharafaldin, I., Lashkari, A.H., Ghorbani, A.A.: Toward generating a new intrusion detection dataset and intrusion traffic characterization. In: International Conference on Information Systems Security and Privacy (ICISSP). pp. 108–116 (2018).https: //doi.org/10.5220/0006639801080116

2018 doi
[18]

Nature620(7972), 172–180 (2023).https://doi.org/10.1038/ s41586-023-06291-2

Singhal, K., Azizi, S., Tu, T., Mahdavi, S.S., Wei, J., Chung, H.W., Scales, N., Tanwani, A., Cole-Lewis, H., Pfohl, S., et al.: Large language models encode clinical knowledge. Nature620(7972), 172–180 (2023).https://doi.org/10.1038/ s41586-023-06291-2

2023
[19]

IEEE Transactions on Automatic Control52(9), 1680–1685 (2007).https://doi.org/ 10.1109/TAC.2007.904277

Tabuada, P.: Event-triggered real-time scheduling of stabilizing control tasks. IEEE Transactions on Automatic Control52(9), 1680–1685 (2007).https://doi.org/ 10.1109/TAC.2007.904277

2007
[20]

Monographs on Statistics and Applied Probability, Chapman and Hall/CRC (2014).https://doi.org/10.1201/b17279

Tartakovsky, A., Nikiforov, I., Basseville, M.: Sequential Analysis: Hypothesis Test- ing and Changepoint Detection. Monographs on Statistics and Applied Probability, Chapman and Hall/CRC (2014).https://doi.org/10.1201/b17279

2014 doi
[21]

Annals of Mathematical Statis- tics16(2), 117–186 (1945).https://doi.org/10.1214/aoms/1177731118

Wald, A.: Sequential tests of statistical hypotheses. Annals of Mathematical Statis- tics16(2), 117–186 (1945).https://doi.org/10.1214/aoms/1177731118

1945
[22]

In: International Conference on Machine Learning (ICML)

Zinkevich, M.: Online convex programming and generalized infinitesimal gradient ascent. In: International Conference on Machine Learning (ICML). pp. 928–936 (2003) Event-Triggered LLM Invocation in Streaming Systems 19 A Proofs A.1 Proof of Theorem 1 (Minimum Inter-Event Time)...

2003
[23]

By the Azuma–Hoeffding inequality: P[| PT t=1 Zt|> ϵ ] ≤ 2 exp(−ϵ2/(2T )). Setting ϵ = √2TlogT and translating to cumulative cost yieldsRegret(T ) ≤ O(√TlogT).⊓ ⊔ A.3 Proof of Theorem 5 (Calibration-to-Miss-Rate) A missed critical event occurs when the event is critical yetut ...
[24]

Deep Ensemble: K = 5independently trained models

MC Dropout:N = 20forward passes. Deep Ensemble: K = 5independently trained models. Trigger parameters.Threshold: anomaly threshold 1.0, uncertainty threshold 0.5. CUSUM: slack k = 0.5, decision thresholdh = 5.0, warmup 30 samples. SPRT: α = 0.05, β = 0.10, warmup 30 samples. O...

2026

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.

[1] [1]

In: Advances in Neural Information Processing Systems (NeurIPS)

Chen, L., Zaharia, M., Zou, J.: FrugalML: How to use ML prediction APIs more accurately and cheaply. In: Advances in Neural Information Processing Systems (NeurIPS). vol. 33, pp. 10685–10696 (2020)

2020

[2] [2]

In: Proceedings of the 33rd International Conference on Machine Learning (ICML)

Gal, Y., Ghahramani, Z.: Dropout as a Bayesian approximation: Representing model uncertainty in deep learning. In: Proceedings of the 33rd International Conference on Machine Learning (ICML). pp. 1050–1059 (2016), arXiv:1506.02142

arXiv 2016

[3] [3]

arXiv preprint arXiv:1603.08983 (2016)

Graves, A.: Adaptive computation time for recurrent neural networks. arXiv preprint arXiv:1603.08983 (2016)

arXiv 2016

[4] [4]

In: International Conference on Machine Learning (ICML)

Guo, C., Pleiss, G., Sun, Y., Weinberger, K.Q.: On calibration of modern neural networks. In: International Conference on Machine Learning (ICML). pp. 1321–1330 (2017)

2017

[5] [5]

Foundations and Trends in Optimization2(3-4), 157–325 (2016).https://doi.org/10.1561/2400000013

Hazan, E.: Introduction to online convex optimization. Foundations and Trends in Optimization2(3-4), 157–325 (2016).https://doi.org/10.1561/2400000013

doi:10.1561/2400000013 2016

[6] [6]

In: IEEE 51st Conference on Decision and Con- trol (CDC)

Heemels, W.P.M.H., Johansson, K.H., Tabuada, P.: An introduction to event- triggered and self-triggered control. In: IEEE 51st Conference on Decision and Con- trol (CDC). pp. 3270–3285 (2012).https://doi.org/10.1109/CDC.2012.6425820 18 Z. Wang

arXiv 2012

[7] [7]

In: Advances in Neural Information Processing Systems (NeurIPS)

Khosla, P., Teterwak, P., Wang, C., Sarna, A., Tian, Y., Isola, P., Maschinot, A., Liu, C., Krishnan, D.: Supervised contrastive learning. In: Advances in Neural Information Processing Systems (NeurIPS). vol. 33, pp. 18661–18673 (2020)

2020

[8] [8]

In: International Conference on Machine Learning (ICML)

Kuleshov, V., Fenner, N., Ermon, S.: Accurate uncertainties for deep learning using calibrated regression. In: International Conference on Machine Learning (ICML). pp. 2796–2804 (2018)

2018