acceptodds
Withdrawn from ICLR 2027

Family-Conditioned Hidden Markov Models for Cross-Architecture Out-of-Distribution Monitoring

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Deep Neural Networks (DNNs) are transforming perception tasks across a wide range of applications, yet their reliability can degrade when inputs at deployment differ from the training distribution. This work presents a family-conditioned runtime monitor for detecting such out-of-distribution (OOD) inputs. The monitor extracts per-layer activation representations, converts them into normalized anomaly scores, and discretizes these scores into ordered severity levels. The resulting sequence of levels across the network is used to train a Hidden Markov Model (HMM) for each structural family, capturing characteristic activation patterns shared across its constituent networks. The central hypothesis is that networks within the same structural family exhibit stronger alignment in their activation behavior than networks from different families. Empirical evaluation across networks from three structural families supports this hypothesis, with within-family alignment better than cross-family alignment. The family-conditioned HMM also generalizes to unseen network architectures: under a leave-one-network-out protocol, the AUROC gap between networks used to train the monitor and held-out networks is at most . Furthermore, the proposed monitor achieves OOD detection performance comparable to per-network baseline methods while eliminating the need to construct and calibrate a separate monitor for every network. These results demonstrate the potential of family-level monitoring as a scalable approach to runtime OOD detection across diverse neural network architectures.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Runtime Monitoring of Neural Networks, Out of Distribution Detection, Neural Networks

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.25 over 4 reviews, above 2% of the others.
Reviewer_LQhoOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes a runtime monitor for out-of-distribution (OOD) detection that is shared across all networks belonging to the same structural architecture family (e.g., ResNet-style networks, Vision Transformers), instead of requiring a separate monitor for each individual network. Per-layer activation statistics are turned into a discretized, depth-normalized sequence of severity levels, over which a single Hidden Markov Model is fitted per family. The paper argues, and empirically supports, that this family-level monitor generalizes to unseen networks of the same family while remaining competitive with existing per-network monitoring baselines.

Critical strengths

S1. The central idea — that a single monitor can be shared across an entire family of structurally related networks instead of requiring individual per-network calibration — is conceptually interesting and, to my knowledge, not explored in this form in prior runtime- monitoring work. The within- vs. cross-family alignment analysis (Table 1) and the leave-one-network-out evaluation (Table 3) provide reasonably convincing evidence that such sharing is possible without a substantial loss in detection performance, which is the paper's most substantive contribution.

Critical weaknesses

W1. The paper does not sufficiently develop or exploit its own central idea. The hypothesis that activation-based severity patterns align within an architecture family is supported only empirically, never explained conceptually. The paper leaves the reader with more open questions than insights, and the reviewer did not gain a deeper understanding of the underlying problem.

W2. The evaluation compares only against other activation-based monitors (Mahalanobis, Outside-the-Box), both of which, like the proposed method, require network-specific calibration. The paper's own related-work discussion identifies a distinct class of output-based OOD detectors (e.g., Hendrycks & Gimpel, 2017) that operate purely on output statistics and may already generalize across architectures without any family-specific or per-network calibration at all. No such baseline is included, so it remains unclear whether the added complexity of the proposed pipeline (per-layer scoring, relative-depth staging with arity correction, family-level HMM fitting) provides any practical advantage over a much simpler alternative. This directly undermines the paper's central motivation of scalable, easy-to-deploy monitoring.

W3. The paper is not self-contained and is written in a vague, unevenly detailed way. The abstract and introduction are generic, and central terminology (what "family" refers to) only becomes concrete well into the paper. The contributions are stated too abstractly to be parsed on first reading. Key concepts are used without explanation: the BUILD/INFER flag parameters, the undefined "combined-layer Mahalanobis" baseline in Table 2, and the Related Work paragraph that lists prior monitors without contrasting them with the proposed approach. Meanwhile, the narrative sections are verbose. A substantial revision of the writing would be needed before the contribution can be assessed fairly by a broad readership.

Questions for authors

Q1. Under what conditions can a monitor be expected to generalize across different network architectures, and why? This is the fundamental question the paper raises but leaves unresolved: the central hypothesis is validated empirically on the specific families tested, but no account is given of the conditions under which such generalization should or should not be expected to hold.

Q2. Is the proposed approach competitive with the state of the art in OOD detection at large? The paper only compares against two activation-based monitoring methods, while the OOD detection literature contains a much broader and more mature body of work. Without such a comparison, it is unclear whether the proposed approach is competitive at all.

Reviewer_Fcb6OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper introduces a novel out-of-distribution detection method that constructs Hidden Markov Models (HMMs) from statistical feature bins produced by layer-wise features of neural networks belonging to similar architectural families. Experimental results show competitive performance against the selected baselines. However, the experimental protocol deviates from existing community standards. At the same time, the necessity of the proposed heuristic design choices is not well supported in the manuscript (e.g., why did the authors choose T=6 for all neural networks?). Given these weaknesses, I do not find this paper to be at the level required for acceptance and clearly recommend rejection.

Critical strengths
  1. The use of a traditional combination of HMMs and statistical features on CPUs is interesting because the computational cost of OOD detection could be low.

  2. The experimental results for RQ1, showing that networks within the same family have lower diff/std scores, support the claims of this paper.

  3. The authors evaluate their proposed framework across three distinct neural network families, including shallow CNNs, ResNets, and Vision Transformers. Furthermore, the leave-one-network-out experiments provide empirical evidence that the family-conditioned HMM can generalize to unseen architectural variants within the same family.

Critical weaknesses
  1. Even though OOD detection is a well-established field, their experimental design, including dataset selection and evaluation metrics, deviates from the protocols followed by recent papers. Specifically, [1] and [2] commonly use ImageNet and evaluate performance using both False Positive Rate (FPR) and AUROC, while the proposed Family HMM method reports only AUROC for the baseline comparisons in RQ2, without including FPR. Moreover, since the goal of the OOD detection task is the same, comparisons with these strong baselines should have been included in their paper. I strongly recommend that the authors explain why these comparisons were not included. Including such comparisons would improve the evaluation, but it would also fundamentally alter the paper's experimental scope, and I am unlikely to raise my score.

  2. Although their design combining statistical features of neural network activations with HMMs is interesting, the ablation studies determining the contribution of each component of the proposed method are limited. For example, they decided to use T=6 stages without explaining why this particular parameter was chosen.

  3. The statistical significance of the experimental results is not established. The authors seem to have reported their results without repeated independent experimental runs or measures of variability. Therefore, I am less confident in the reliability of their results.

References

[1] Ding, Y., Liu, X., Unger, J. and Eilertsen, G., 2026. Enhancing out-of-distribution detection with extended logit normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 24823-24832).

[2] Chen, W., Yeh, R.A., Mou, S. and Gu, Y., 2025, June. Leveraging perturbation robustness to enhance out-of-distribution detection. In 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (pp. 4724-4733). IEEE.

Questions for authors

See weaknesses

Reviewer_4pprOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper asks whether architecturally related networks can share a single OOD monitor. For each network, the method fits a front-end on ID activations: per-layer KMeans clusters, Ledoit–Wolf covariances, and cluster-wise Mahalanobis distances normalized by a 95th-percentile radius. Layer scores are then z-scored, grouped into relative-depth stages, reduced by a stage-wise maximum with an “arity correction”, and quantile-binned into six levels. The resulting discrete sequences from networks in the same family are pooled to fit an HMM, whose negative log-likelihood serves as the OOD score, with a threshold set at the 99th percentile of ID calibration data. The authors evaluate 13 networks (5 shallow CNNs, 5 ResNet-style networks, 3 ViTs) under a leave-one-network-out (LONO) protocol, and report that LONO AUROC stays close to a pooled-inclusive setting, along with a stage-wise alignment analysis (RQ1) and comparisons against Mahalanobis and Outside-the-Box baselines.

Critical strengths
  1. A practically motivated problem and a sensible evaluation unit. Reusing monitors across architecture changes is a real deployment concern, and holding out whole networks (rather than only input samples) matches that concern better than a standard split. The data-split description in §4.1 (pipeline fitting / HMM fitting / threshold calibration / final evaluation) is clear and, as written, appropriate.
  2. Candid reporting. The authors openly acknowledge that combined-layer Mahalanobis is stronger in six of eight comparisons, discuss the gap between AUROC and FPR@95%TPR, and state the limitations on scale and hyperparameter sensitivity. This makes the results easier to interpret and is a good basis for a stronger revision.
Critical weaknesses
  1. The evidence does not yet isolate what is being shared, or why it works. (a) As I read §4.1 and Algorithms 2–3, the held-out network still requires its own ID activations to fit the full front-end (clusters, covariances, radii, layer statistics, stage quantile boundaries). What transfers is the HMM and the final threshold, which is narrower than the abstract’s suggestion that monitors need not be built and calibrated per network. Fitting time, storage, and adaptation cost are not reported, so the practical saving is unquantified; the front-end may well dominate the cost of the HMM. (b) The key controls are missing. With the same front-end, how do a global HMM, per-network HMMs, independent-stage categorical likelihoods, an observed Markov model, and simple aggregation (mean/max/bin-sum) perform? Because quantile binning fixes the ID marginals by construction, the shared performance could come mostly from the per-network normalization rather than from family structure or sequential modeling. Close LONO and pooled-inclusive results cannot rule this out. (c) If I understand Algorithm 3 correctly, the arity correction subtracts a stage-specific constant before the stage’s quantile boundaries are re-estimated from the same shifted ID scores. The constant then cancels, and the bin sequence given to the HMM would be unchanged. If so, the correction has no effect on detection as specified (it may still affect the continuous diagnostics in RQ1). I may be misreading the implementation, and I would be glad to be corrected, but the paper should then clarify the difference between the algorithm and the code.
  2. The evaluation does not yet support the deployment-level claims. (a) Table 2 compares pooled-family AUROC for the HMM with mean per-network AUROC for the baselines. These are different statistics (pooling can raise or lower AUROC relative to per-network averaging), so the table is not a like-for-like comparison. (b) Near-identical mean AUROC does not establish a stable shared threshold. In Table 3, ResNet/CIFAR-100 has an AUROC difference of 0.0001 but an FPR@95%TPR difference of 11.91 points. In addition, FPR@95%TPR is not the false-positive rate at the paper’s 99% ID-calibrated threshold, and the realized ID FPR/OOD TPR at that threshold on each held-out network is never reported. With 512 ID test samples per network, a 1% target is about 5 events, and there are no training seeds, repeated splits, or intervals. (c) For shallow networks with 3–4 monitored layers and T=6, it is unclear how empty stages enter the HMM (dropped, masked, or padded), and raw NLL depends on sequence length, which affects cross-network score comparability under a shared threshold. The depth-index formula also appears inconsistent (ℓ=0 maps to stage −1 under the stated expression; Algorithm 3 uses ℓ=1..L).

Taken together, the claim that family-level sharing offers a practical benefit is plausible but not yet demonstrated.

Questions for authors

Clarifications whose answers could change my assessment:

  1. For a new network, which components must be refit on its own ID data, and what are the fitting time, memory/storage, and online overhead compared with fitting a per-network HMM?
  2. Under Algorithm 3, does removing the arity correction leave the bin sequences exactly unchanged (same front-end, samples, seeds)? If not, what differs between the algorithm and the implementation?
  3. How are stages handled for networks with fewer than six monitored layers (layer-to-stage mapping, masking vs. dropping, length normalization of the score)?
  4. For each LONO network, what are the held-out ID FPR and OOD TPR at the shared 99% threshold? Could you provide the raw values and sign (not only absolute differences) behind the Table 3 gaps, especially the 0.1191 FPR gap?
  5. Could Table 2 be reported with the same aggregation for all methods (per-network macro-average, plus pooled scores separately), with training-seed variation or confidence intervals?

Suggestions (not required for my assessment, but would strengthen the paper): 6. Same-front-end comparisons of family / global / per-network HMMs, independent-stage likelihoods and simple aggregation, plus a within-stage shuffle that preserves each stage’s ID marginal while breaking cross-stage dependence. 7. A cross-family transfer control, and RQ1 repeated with a common ID dataset and comparable hook granularity across families, since the current families differ in training data and monitoring granularity. 8. Stronger baselines such as Energy and ReAct under the same ID-only calibration constraint.

Minor points: the abstract’s “at most 0.009” AUROC gap appears to be 0.0095 for ViT/SVHN in Table 3; Algorithm 1 uses an undefined L_train while §4.1 describes validation-based calibration; the OpenOOD citation (Wu et al., 2024) in the reference list appears to be the BAM paper rather than OpenOOD; and HMM details (number of hidden states, restarts, EM stopping criteria, emission model) are not given.

Reviewer_y5mKOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes a family-conditioned out-of-distribution detection framework that transforms intermediate-layer activations into normalized anomaly scores, aggregates them into relative-depth stages, and models the resulting discrete sequences using a shared Hidden Markov Model. The authors hypothesize that architectures within the same structural family exhibit similar activation-derived anomaly patterns and evaluate this hypothesis using 13 neural networks across shallow CNN, ResNet, and Vision Transformer families. Experiments suggest that the proposed method achieves competitive OOD detection performance against selected per-network baselines.

Critical strengths
  1. The paper explores family-level OOD monitoring, which is a potentially useful direction for reducing the need to maintain separate detection models for structurally related neural networks.

  2. The method combines clustering, Mahalanobis distances, covariance shrinkage, quantile discretization, and HMMs into a unified monitoring pipeline.

Critical weaknesses
  1. The proposed pipeline largely consists of combining established components, including KMeans, Mahalanobis distance, quantile normalization, and HMMs. Although the family-level formulation is potentially useful, the paper does not demonstrate that the proposed combination introduces a fundamentally new modeling capability beyond existing feature-based OOD detection methods. More importantly, the paper lacks comparisons with simpler family-level alternatives that could achieve the same objective without requiring an HMM. The central question is whether the proposed HMM provides any essential benefit over simpler approaches to aggregating and sharing normalized anomaly scores.

  2. The HMM treats stages along network depth as sequential observations and introduces hidden-state transitions to capture their dependencies. However, network depth is not a temporal process, and adjacent layers are related through deterministic neural-network computations rather than an underlying stochastic state-transition mechanism. While HMMs can model ordered non-temporal data, the authors provide neither theoretical justification nor experiments demonstrating that this particular dependency model is suitable or necessary. Simple alternatives such as averaging stage scores, taking their maximum, summing marginal log-likelihoods, or learning a lightweight classifier over stage scores should be considered. Without these comparisons, it remains unclear whether the reported performance gains arise from sequential modeling or merely from combining information across multiple layers.

Questions for authors

Can the authors compare their approach against simple stage aggregation, independent categorical likelihood models, and lightweight discriminative models trained on the same stage sequences? Is there evidence that modeling transitions between adjacent network stages provides benefits beyond modeling their marginal distributions?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.