Pith. sign in

REVIEW 5 major objections 4 minor 36 references

Conformal Predictive Monitoring for Multi-Modal Scenarios

T0 review · 5 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read GenQPM gives each dynamical mode its own statistically calibrated STL-robustness prediction interval, obtained by sampling future trajectories with a diffusion model and recalibrating per mode with conformalized quantile regression.

desk verdict A solid, useful mode-specific conformal monitoring paper; the theory is right for the mode predictor's own labels, but the text overclaims coverage for learned predictors and contradicts its own Table 1. read the letter →

arxiv 2509.01338 v1 pith:JZU72JC2 submitted 2025-09-01 cs.AI

Francesca Cairoli , Luca Bortolussi , Jyotirmoy V. Deshmukh , Lars Lindemann , Nicola Paoletti This is my paper
classification cs.AI
keywords quantitativepredictivemonitoringsignaltemporallogicconformalpredictiondiffusionmodelsmulti-modaldynamicsruntimeverificationintervalsSTLrobustness
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

Quantitative predictive monitoring for stochastic systems usually returns one wide, mode-agnostic prediction interval for STL robustness, which is uninformative when the system can evolve along qualitatively different modes. GenQPM instead learns a score-based diffusion surrogate of the dynamics, classifies generated trajectories into modes, and applies conformalized quantile regression per mode to produce mode-specific intervals with a per-mode coverage guarantee. Across four case studies the mode-specific intervals are roughly half as wide as the mode-agnostic baseline while still meeting the target coverage, so the same statistical guarantee is paired with much more actionable information about which modes are safe. The paper also shows that the guarantee is conditional on the mode predictor used to partition the calibration set, and explicitly notes that a learned approximate mode predictor that labels calibration and test data differently breaks exchangeability.

What carries the argument

The load-bearing object is the calibrated mode-specific prediction interval CPI^{φ,mode}(s(0)), assembled from three components: a score-based diffusion generative model that approximates the conditional trajectory distribution p(s|s(0)) and thereby yields empirical robustness quantiles; a mode predictor M (known or learned) that partitions trajectories into modes; and conformalized quantile regression (CQR), which turns the raw quantile intervals into calibrated ones by adding a per-mode, per-property score τ^{φ,mode} computed from the calibration nonconformity scores (Eq. 3). The CQR step is what converts approximate surrogate-based quantiles into statistically valid coverage, and the mode

What would settle it

Take a multi-modal scenario with ground-truth mode labels, train a mode classifier under a labeling rule that differs from the test labeling rule, and measure per-ground-truth-mode coverage of the reported CPIs on a held-out test set; if any ground-truth mode has empirical coverage substantially below 1−α, Eq. (5) does not hold for that mode. A simpler falsifier is to use a mode whose calibration set is empty or extremely small: then the empirical quantiles are undefined or the interval is infinite, as already observed for the underrepresented modes in the Navigation case study.

Watch

Extended reading notes

Core claim

The central claim is that mode-conditional coverage can be added to conformal predictive monitoring without sacrificing the 1−α guarantee. GenQPM trains a conditional diffusion model pθ(·|s(0)) once, samples K future trajectories from any current state, filters them through a mode predictor M to isolate trajectories of each mode, and computes empirical α/2 and 1−α/2 quantiles of their STL robustness. Calibration nonconformity scores are then derived per mode, and the (1−α)-th empirical quantile of those scores yields a mode-specific correction τ^{φ,mode}. The resulting calibrated interval CPI^{φ,mode}(s(0)) = [q_lo^{φ,mode}(s(0)) − τ^{φ,mode}, q_hi^{φ,mode}(s(0)) + τ^{φ,mode}] satisfies the

Load-bearing premise

The per-mode coverage guarantee (Eq. 5) assumes calibration and test trajectories are exchangeable after conditioning on the mode predictor that labeled the calibration set; the paper itself notes (Remark 1) that a learned approximate mode predictor that labels calibration and test with different rules breaks exchangeability, so the guarantee then does not apply to ground-truth modes.

Editorial extensions

If this is right

  • An operator can see, at runtime, which mode is about to violate an STL property, because each mode has its own interval rather than one blurred aggregate interval.
  • The generative model is trained once and reused across different STL properties; changing the requirement only requires re-evaluating robustness and recomputing the property-specific calibration scores, not retraining the dynamics surrogate.
  • The union of the mode-specific intervals keeps the original marginal 1−α coverage, so adding mode granularity does not weaken the overall guarantee.
  • In the reported multi-modal benchmarks, the mode-specific intervals are 43–53% narrower than the mode-agnostic baseline (and close to the width of the empirical quantile range), so the same statistical confidence is delivered with substantially more precise predictions.
  • Mode-specific intervals expose which modes are safe and which are unsafe, supporting decision-making that can preemptively avoid a violation by steering toward a safer mode.

Reading between the lines

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

  • If a learned mode classifier drifts between calibration and deployment, the Eq. (5) guarantee will apply to the classifier's own partition, not to the true dynamics modes; a deployed system would need a separate way to detect or correct such mode-label shift, e.g. by using a small ground-truth-labeled set to monitor classifier agreement.
  • The diffusion surrogate enables counterfactual monitoring: because trajectories can be sampled conditioned on a mode, one could ask what the robustness interval would look like for 'what if the agent turned right' without retraining, as long as the sampler can be conditioned or directed toward that mode.
  • The infinite-width intervals observed for underrepresented modes in the Navigation case study suggest a natural stopping rule: when the per-mode calibration set is too small to support reliable quantiles, return 'unknown' or a flag requiring more data rather than a spuriously wide interval.
  • The method's reliance on exchangeability after conditioning on M implies that the guarantees are covariate-independent only along the trajectory distribution; bootstrapped calibration, as used in the experiments, is a practical way to tighten the gap, but the formal guarantee remains tied to the single calibration set used.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 4 minor

Summary. The paper proposes GenQPM, a conformal predictive monitoring method for stochastic systems with multi-modal dynamics. A score-based diffusion model is trained to sample future trajectories conditioned on the current state; a mode predictor M partitions trajectories into G modes; per-mode empirical quantiles of STL robustness are computed; and CQR-style calibration yields mode-specific prediction intervals. The paper claims the intervals satisfy mode-conditional coverage (Eq. 5) and are substantially tighter than the mode-agnostic QPM baseline, with experiments on Signal, Crossroad, Multi-Agent Crossroad, and Navigation case studies.

Significance. If the claims hold, GenQPM would be a useful advance: it combines recent score-based generative models with conformal inference to obtain mode-aware, calibrated intervals for STL robustness, and it releases code. The core conformal argument is standard for a fixed, exact mode predictor, and the experiments show large width reductions in several well-separated multi-modal cases. However, the paper's own Remark 1 limits the theoretical guarantee to the mode predictor used for calibration, which is not the guarantee advertised for learned predictors. Several empirical claims in Section 5.4 are also contradicted by Table 1. The contribution is promising but currently overclaimed.

major comments (5)
  1. [§4.1, Eq. (5), and Remark 1] Equation (5) asserts coverage conditional on M(s)=mode. Remark 1 correctly states that the guarantee applies only to the mode predictor used to partition the calibration set. When M is learned, calibration is partitioned by M's labels while Table 1 evaluates coverage against ground-truth labels, and Remark 1 admits exchangeability fails. Thus the contribution claim that GenQPM 'supports both known and learned mode predictors' (page 3) and the formal statement of Problem 1, which allows M to be unknown, are not supported for learned predictors. The theoretical scope must be narrowed or a separate argument must be provided.
  2. [§5.4, Table 1] The sentence 'CPI always meets the mode-wise desired coverage' is directly contradicted by Table 1. For Xroad-φcar, the Exact column reports mode-wise coverage (93.9, 86.6, 89.4) and the Approximate column reports (94.2, 87.4, 88.6), both below the 90% target in two modes. The Navigation approximate case reports 100% for modes 1 and 4 only because the CPI width is infinite. The empirical validation does not support the blanket coverage claim; the table and text need per-mode coverage, error bars, and an explanation for the violations.
  3. [Abstract and §5.4, Table 1] The abstract claims GenQPM produces prediction intervals that are 'significantly more informative (less conservative)' than the mode-agnostic baseline. Table 1 shows the opposite in two rows: Xroad-φcar widths are 11.86/11.92 vs baseline 10.02/10.03 (+18%/+17%), and Navigation width is infinite vs baseline 1.09. The efficiency claim should be qualified to the cases with well-separated modes and adequate per-mode calibration data; as stated, it overgeneralizes.
  4. [§5.3, Experimental settings] The protocol resamples 500 calibration initial states for each test point ('allowing calibration sets to vary at each test point'). Standard split-conformal/CQR guarantees, including Eq. (5), assume a fixed calibration set exchangeable with the test point. The paper provides no argument that the guarantee survives this bootstrapped, test-dependent calibration procedure. This is load-bearing for the empirical coverage results; the authors should either use a fixed calibration set or analyze the effect of their resampling protocol on the guarantee.
  5. [§5.4, Navigation results] Infinite-width intervals are used to report 100% coverage for underrepresented modes. An infinite interval is trivially valid but carries no predictive information and should not be counted as a success. This conflates coverage with efficiency. The text at the end of Section 5.4 acknowledges the generative model is not capturing the dynamics in such modes, but Table 1 still lists these modes as achieving coverage. The frequency of infinite-width intervals should be reported and excluded or flagged when presenting mode-wise coverage.
minor comments (4)
  1. [Algorithm 1, step 2(d)] The notation uses both q^{φ,mode}_{lo/hi} and q^φ_{lo/hi}; unify it. Step 2(c) assumes Kmode < K and does not specify behavior when Kmode = 0; this can occur for rare modes and should be handled explicitly.
  2. [Table 1] The table formatting is broken in several places: '93.610.80', '90.011.86', '100.' and '(100.,' are missing separators and parentheses. This makes the results hard to read and should be corrected.
  3. [§4.2 and Fig. 7] The section title 'Dynamic multi-agent dynamic environments' is redundant. Also, Fig. 7 caption uses 'Navigator' while the case study is called 'Navigation'; use a consistent name.
  4. [§5.4] The explanation that the approximate mode predictor yields infinite-width intervals 'because the classifier struggles' is an interpretation without quantitative support. Report the classifier's per-mode accuracy or confusion matrix so the reader can judge the approximation quality.

Circularity Check

1 steps flagged · score 3.0 of 10

The conformal calibration is sound, but Eq. (5)'s mode-specific guarantee is defined relative to the same mode predictor M that partitions the calibration set; for learned/approximate M this is a self-definitional scope reduction, not a ground-truth mode guarantee.

  1. self definitional [Remark 1, Section 4.1, and Eq. (5)]
    "Regardless of how M is obtained, the mode-specific validity condition in Equation (5) applies to the mode predictor used to partition the calibration set. However, experiments may reveal deviations from the target coverage level of 1 − α when evaluating the approximate mode predictor against a labeled test set. This occurs because the calibration data (classified by the approximate predictor) and the test data (classified by ground truth labels) lack exchangeability."

    Eq. (5) states P(Rob_phi(s) in CPI^{phi,mode}(s(0)) | M(s)=mode) ≥ 1−α. Algorithm 1 constructs the calibration set as D_c^(mode) = {s in D_c | M(s)=mode} and computes tau^{phi,mode} from exactly those points. Thus the guarantee is, by construction, a class-conditional conformal guarantee for the same classifier M that defines the classes. When M is learned/approximate, this does not deliver Problem 1's guarantee for any independent ground-truth dynamical mode; Remark 1 concedes that coverage evaluated against ground-truth labels can deviate because exchangeability fails. The paper's contribution claim of supporting learned mode predictors therefore reduces to coverage for the predictor's own labels, with ground-truth mode coverage only empirically asserted, not derived.

full rationale

The core conformal calibration in GenQPM is not circular: the diffusion surrogate is trained on D_t, mode-specific quantiles are computed from generated trajectories, and CQR recalibrates those intervals against true robustness values on D_c^(mode). The coverage statement follows from the standard CQR theorem under exchangeability, rather than from fitting the desired coverage result. No fitted parameter is renamed as a prediction, and no external uniqueness theorem is imported through self-citations; the mode-specific extension is applied directly rather than borrowed as a black-box forbidden-alternative argument. The main circularity-adjacent issue is definitional: Eq. (5)'s conditioning event is the same mode predictor M used to partition the calibration set, so the formal guarantee is guaranteed only for M's own labels. Remark 1 states this explicitly, and the approximate-mode experiments are evaluated against ground-truth labels, with Table 1 reporting mode-wise coverages such as 87.4%, 88.6%, and 86.6% for Xroad-phi_car, despite the text claiming 'CPI always meets the mode-wise desired coverage.' This is a scope mismatch and an overclaim, but it is not a case where the derived interval is equivalent to the input by construction. The experimental protocol also resamples calibration sets per test point, which departs from the fixed-exchangeability assumption of split conformal inference, but that is a validity concern rather than a circularity. Overall, the mathematical derivation is self-contained and honest about its conditioning; the score reflects the self-definitional reduction of the learned-mode guarantee, not a wholesale collapse of the method into its inputs.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

No new physical entities are postulated. The free parameters are standard hyperparameters and the significance level; the conformal guarantee does not depend on their exact values, but interval tightness does. The key axiomatic input is exchangeability and correct mode labels.

free parameters (4)
  • Number of generated trajectories per calibration state K = not stated explicitly; 300 trajectories per state used in datasets
    Chosen by hand; small K makes empirical quantiles noisy, affecting interval tightness but not conformal coverage.
  • Significance level alpha = 0.1
    Chosen for all experiments; sets target coverage 90% and quantile levels 0.05/0.95.
  • Diffusion model hyperparameters (epochs, batch size, learning rate) = 200, 512, 0.0005
    Fixed for all case studies; affect generative fidelity and hence interval tightness, not coverage.
  • Bootstrapped calibration size = 500 initial states
    Number of calibration states sampled per test point; affects calibration score distribution and coverage stability.
assumptions (5)
  • domain assumption Calibration and test trajectories are exchangeable (possibly conditionally on mode).
    Stated in Section 4.1: 'guarantees hold under the assumption of exchangeability between calibration and test samples.' This is the backbone of conformal coverage.
  • domain assumption Calibration trajectories are labeled with the correct dynamical mode.
    Section 4.1: 'we assume that the calibration trajectories are labelled with the dynamical mode they belong to.' Needed for class-conditional calibration; violated for approximate predictors with ground-truth labels.
  • domain assumption The diffusion surrogate p_theta approximates the conditional trajectory distribution well enough that empirical quantiles are informative.
    Section 4.1 uses generated trajectories to form quantile intervals; conformal calibration fixes coverage, but tightness and mode separation rely on p_theta quality (see 'Unbalanced generation' and infinite-width cases).
  • domain assumption Observed trajectories in Dt, Dc, and test are drawn from the same stochastic process S.
    Inherited from QPM problem statement (Section 2); required for exchangeability and transfer of the trained surrogate.
  • standard math Standard STL quantitative semantics (Rob) are the ground truth for robustness.
    Section 2 defines Rob; the monitor's target values are computed with these semantics.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Conformal Predictive Monitoring for Multi-Modal Scenarios." pith.science (2026). https://pith.science/paper/JZU72JC2

@misc{pith2026250901338,
  author       = {Pith},
  title        = {Pith review of: Conformal Predictive Monitoring for Multi-Modal Scenarios},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JZU72JC2}},
  note         = {Machine review of arXiv:2509.01338}
}
read the original abstract

We consider the problem of quantitative predictive monitoring (QPM) of stochastic systems, i.e., predicting at runtime the degree of satisfaction of a desired temporal logic property from the current state of the system. Since computational efficiency is key to enable timely intervention against predicted violations, several state-of-the-art QPM approaches rely on fast machine-learning surrogates to provide prediction intervals for the satisfaction values, using conformal inference to offer statistical guarantees. However, these QPM methods suffer when the monitored agent exhibits multi-modal dynamics, whereby certain modes may yield high satisfaction values while others critically violate the property. Existing QPM methods are mode-agnostic and so would yield overly conservative and uninformative intervals that lack meaningful mode-specific satisfaction information. To address this problem, we present GenQPM, a method that leverages deep generative models, specifically score-based diffusion models, to reliably approximate the probabilistic and multi-modal system dynamics without requiring explicit model access. GenQPM employs a mode classifier to partition the predicted trajectories by dynamical mode. For each mode, we then apply conformal inference to produce statistically valid, mode-specific prediction intervals. We demonstrate the effectiveness of GenQPM on a benchmark of agent navigation and autonomous driving tasks, resulting in prediction intervals that are significantly more informative (less conservative) than mode-agnostic baselines.

Figures

Figures reproduced from arXiv: 2509.01338 by the authors.

Figure 1
Figure 1. Left: Vehicle facing multiple possible choices. Right: Distribution of the robust￾ness values of these choices/modalities w.r.t. an STL safety requirement. Bottom-right: uninformative mode-agnostic prediction interval vs informative mode-specific intervals. poral logic specifications becomes a stochastic quantity (some evolutions can be safe, others unsafe). Although probabilistic model checking (PMC) [2] can provid… view at source ↗
Figure 2
Figure 2. GenQPM: dynamic-aware quantitative predictive monitor with mode￾conditional guarantees. for all mode ∈ {1, . . . , G}; above, M maps every realization of length H of the stochastic process into one of the G modes. To solve Problem 1, we will use the class-conditional version of conformal in￾ference to obtain conditional guarantees determined by the mode-conditional calibration sets D (mode) c = {⃗s ∈ Dc|M(⃗s) = mode… view at source ↗
Figure 3
Figure 3. Example of an un￾certain multi-agent environ￾ment: pedestrian following a stochastic behavior with an obstacle (the tree) par￾tially hiding its position to the ego vehicle. However, in multi-agent environments, where inter￾actions are highly unpredictable, the distribution of robustness values becomes significantly more com￾plex. In GenQPM we face such dynamic environ￾ments as follows. For each agent that enters the… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Left: distribution of STL robustness values for Crossroad-ϕright, ϕcar, ϕmulti, Signal and Navigation respectively. Right: Simulated trajectories with colour denoting the STL robustness for the same case studies. 5.1 Case Studies Signal. One-dimensional signal evolving…
Figure 5
Figure 5. Figure 5: Mode predictor M applied to trajectories from Signal, Crossroad and Navigator scenarios [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]
Figure 6
Figure 6. Figure 6: Mode-specific calibrated prediction interval with mode-specific guarantees for Signal (top) and Crossroad-ϕright(bottom) [PITH_FULL_IMAGE:figures/full_fig_p015_6.png]
Figure 7
Figure 7. Figure 7: Mode-specific calibrated prediction interval with mode-specific guarantees for Crossroad-ϕcar, ϕmulti and Navigator. 5.4 Results Results are summarized in [PITH_FULL_IMAGE:figures/full_fig_p016_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

36 extracted references · 26 canonical work pages

  1. [1]

    arXiv preprint arXiv:2107.07511 19 (2021)

    Angelopoulos, A.N., Bates, S.: A gentle introduction to conformal prediction and distribution-free uncertainty quantification. arXiv preprint arXiv:2107.07511 19 (2021)

    arXiv 2021

  2. [2]

    MIT press (2008)

    Baier, C., Katoen, J.P.: Principles of model checking. MIT press (2008)

    2008

  3. [3]

    Newnes (2014)

    Balasubramanian,V.,Ho,S.S.,Vovk,V.:Conformalpredictionforreliablemachine learning: theory, adaptations and applications. Newnes (2014)

    work page 2014

  4. [4]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Bansal, A., Chu, H.M., Schwarzschild, A., Sengupta, S., Goldblum, M., Geiping, J., Goldstein, T.: Universal guidance for diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 843–852 (2023)

    work page 2023

  5. [5]

    Scalable Stochastic Parametric Verification with Stochastic Variational Smoothed Model Checking

    Bortolussi,L.,Cairoli,F.,Carbone,G.,Pulcini,P.:Stochasticvariationalsmoothed model checking. arXiv preprint arXiv:2205.05398 (2022)

    work page Pith review arXiv 2022

  6. [6]

    Springer (2019)

    Bortolussi, L., Cairoli, F., Paoletti, N., Smolka, S.A., Stoller, S.D.: Neural predic- tivemonitoring.In:InternationalConferenceonRuntimeVerification.pp.129–147. Springer (2019)

    work page 2019

  7. [7]

    Interna- tional Journal on Software Tools for Technology Transfer23(4), 615–640 (2021)

    Bortolussi, L., Cairoli, F., Paoletti, N., Smolka, S.A., Stoller, S.D.: Neural predic- tive monitoring and a comparison of frequentist and bayesian approaches. Interna- tional Journal on Software Tools for Technology Transfer23(4), 615–640 (2021)

    work page 2021

  8. [8]

    Information and Computation247, 235–253 (2016)

    Bortolussi, L., Milios, D., Sanguinetti, G.: Smoothed model checking for uncer- tain continuous-time Markov chains. Information and Computation247, 235–253 (2016)

    work page 2016

Show all 36 references
  1. [9]

    Cairoli, F., Bortolussi, L., Paoletti, N.: Neural predictive monitoring under partial observability (2021)

    2021

  2. [10]

    Nonlinear Analysis: Hybrid Systems 57, 101606 (2025)

    Cairoli, F., Kuipers, T., Bortolussi, L., Paoletti, N.: Conformal quantitative predic- tive monitoring of stochastic systems with conditional validity. Nonlinear Analysis: Hybrid Systems 57, 101606 (2025)

    2025

  3. [11]

    In: ISoLA 2022 Symposium, to appear (2022)

    Cairoli, F., Paoletti, N., Bortolussi, L.: Neural predictive monitoring for collective adaptive systems. In: ISoLA 2022 Symposium, to appear (2022)

    2022

  4. [12]

    In: Proceedings of the 26th ACM International Conference on Hybrid Systems: Computation and Control

    Cairoli, F., Paoletti, N., Bortolussi, L.: Conformal quantitative predictive monitor- ing of stl requirements for stochastic processes. In: Proceedings of the 26th ACM International Conference on Hybrid Systems: Computation and Control. pp. 1–11 (2023)

    2023

  5. [13]

    In: Proceedings of the 45th IEEE Conference on Decision and Control

    Djeridane, B., Lygeros, J.: Neural approximation of PDE solutions: An application to reachability computations. In: Proceedings of the 45th IEEE Conference on Decision and Control. pp. 3034–3039. IEEE (2006)

    2006

  6. [14]

    In: Proceedings of International Conference on Formal Modeling and Analy- sis of Timed Systems

    Donzé, A., Maler, O.: Robust satisfaction of temporal logic over real-valued sig- nals. In: Proceedings of International Conference on Formal Modeling and Analy- sis of Timed Systems. pp. 92–106. Springer, Klosterneuburg, Austria (Sep 2010). https://doi.org/10.1007/978-3-642-15297-9_9

    2010 doi

  7. [15]

    In: International Conference on Formal Modeling and Analysis of Timed Systems

    Granig, W., Jakšić, S., Lewitschnig, H., Mateis, C., Ničković, D.: Weakness moni- tors for fail-aware systems. In: International Conference on Formal Modeling and Analysis of Timed Systems. pp. 283–299. Springer (2020)

    2020

  8. [16]

    Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models (2020)

    2020

  9. [17]

    IEEE transactions on Robotics and Automation12(4), 566–580 (1996)

    Kavraki, L.E., Svestka, P., Latombe, J.C., Overmars, M.H.: Probabilistic roadmaps for path planning in high-dimensional configuration spaces. IEEE transactions on Robotics and Automation12(4), 566–580 (1996)

    1996

  10. [18]

    arXiv preprint arXiv:2404.17487 (2024)

    Kiyani, S., Pappas, G., Hassani, H.: Conformal prediction with learned features. arXiv preprint arXiv:2404.17487 (2024)

    2024 arXiv

  11. [19]

    The journal of logic and algebraic programming78(5), 293–303 (2009)

    Leucker, M., Schallhart, C.: A brief account of runtime verification. The journal of logic and algebraic programming78(5), 293–303 (2009)

    2009

  12. [20]

    IEEE Robotics and Automation Letters (2023) 20

    Lindemann, L., Cleaveland, M., Shim, G., Pappas, G.J.: Safe planning in dynamic environments using conformal prediction. IEEE Robotics and Automation Letters (2023) 20

    2023

  13. [21]

    In: Proceedings of the ACM/IEEE 14th International Con- ference on Cyber-Physical Systems (with CPS-IoT Week 2023)

    Lindemann, L., Qin, X., Deshmukh, J.V., Pappas, G.J.: Conformal prediction for stl runtime verification. In: Proceedings of the ACM/IEEE 14th International Con- ference on Cyber-Physical Systems (with CPS-IoT Week 2023). pp. 142–153 (2023)

    2023

  14. [22]

    arXiv preprint arXiv:2409.00536 (2024)

    Lindemann, L., Zhao, Y., Yu, X., Pappas, G.J., Deshmukh, J.V.: Formal veri- fication and control with conformal prediction. arXiv preprint arXiv:2409.00536 (2024)

    2024 arXiv

  15. [23]

    In: Formal Techniques, Modelling and Analysis of Timed and Fault-Tolerant Systems, pp

    Maler, O., Nickovic, D.: Monitoring temporal properties of continuous signals. In: Formal Techniques, Modelling and Analysis of Timed and Fault-Tolerant Systems, pp. 152–166. Springer (2004)

    2004

  16. [24]

    In: Advances in Neural Information Processing Systems

    Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high- performance deep learning library. In: Advances in Neural Information Processing Systems. pp. 8024–8035 (2019)

    2019

  17. [25]

    In: 2021 60th IEEE Conference on Decision and Control (CDC)

    Rodionova, A., Lindemann, L., Morari, M., Pappas, G.J.: Time-robust control for stl specifications. In: 2021 60th IEEE Conference on Decision and Control (CDC). pp. 572–579. IEEE (2021)

    2021

  18. [26]

    arXiv preprint arXiv:1905.03222 (2019)

    Romano, Y., Patterson, E., Candès, E.J.: Conformalized quantile regression. arXiv preprint arXiv:1905.03222 (2019)

    1905 arXiv

  19. [27]

    arXiv preprint arXiv:1803.03237 (2018)

    Royo, V.R., Fridovich-Keil, D., Herbert, S., Tomlin, C.J.: Classification-based ap- proximate reachability with guarantees applied to safe trajectory tracking. arXiv preprint arXiv:1803.03237 (2018)

    2018 arXiv

  20. [28]

    arXiv preprint arXiv:2308.16534 (2023)

    Scassola, D., Saccani, S., Carbone, G., Bortolussi, L.: Conditioning score-based generative models by neuro-symbolic constraints. arXiv preprint arXiv:2308.16534 (2023)

    2023 arXiv

  21. [29]

    Sohl-Dickstein, J., Weiss, E.A., Maheswaranathan, N., Ganguli, S.: Deep unsuper- vised learning using nonequilibrium thermodynamics (2015)

    2015

  22. [30]

    Tashiro, Y., Song, J., Song, Y., Ermon, S.: Csdi: Conditional score-based diffusion models for probabilistic time series imputation (2021)

    2021

  23. [31]

    In: 6th Annual Learning for Dynamics & Control Conference

    Tumu, R., Cleaveland, M., Mangharam, R., Pappas, G., Lindemann, L.: Multi- modal conformal prediction regions by optimizing convex shape templates. In: 6th Annual Learning for Dynamics & Control Conference. pp. 1343–1356. PMLR (2024)

    2024

  24. [32]

    Springer Nature (2022)

    Vovk, V., Gammerman, A., Shafer, G.: Algorithmic Learning in a Random World. Springer Nature (2022)

    2022

  25. [33]

    IEEE Robotics & Automation Magazine 27(2), 102–116 (2020)

    Yel, E., Carpenter, T.J., Di Franco, C., Ivanov, R., Kantaros, Y., Lee, I., Weimer, J., Bezzo, N.: Assured runtime monitoring and planning: Toward verification of neural networks for safe autonomous operations. IEEE Robotics & Automation Magazine 27(2), 102–116 (2020)

    2020

  26. [34]

    Information and Computation204(9), 1368–1409 (2006)

    Younes, H.L., Simmons, R.G.: Statistical probabilistic model checking with a fo- cus on time-bounded properties. Information and Computation204(9), 1368–1409 (2006)

    2006

  27. [35]

    IEEE Journal on Selected Areas in Information Theory (2024)

    Zecchin, M., Park, S., Simeone, O.: Forking uncertainties: Reliable prediction and model predictive control with sequence models via conformal risk control. IEEE Journal on Selected Areas in Information Theory (2024)

    2024

  28. [36]

    In: 2024 ACM/IEEE 15th International Conference on Cyber-Physical Systems (ICCPS)

    Zhao, Y., Hoxha, B., Fainekos, G., Deshmukh, J.V., Lindemann, L.: Robust con- formal prediction for stl runtime verification under distribution shift. In: 2024 ACM/IEEE 15th International Conference on Cyber-Physical Systems (ICCPS). pp. 169–179. IEEE (2024)

    2024

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.