acceptodds
Under review as a conference paper at ICLR 2027

GLADE: Adaptive Structural Transfer for Powerful Out-of-Distribution Testing

Abstract

Auxiliary labeled data from historical records, related domains, or synthetic sources can be leveraged to improve the power of out-of-distribution testing. However, differences between auxiliary and current anomaly distributions can make directly transferred detection scores ineffective or even harmful. The challenge is to turn auxiliary category information into detection power while controlling false discoveries and guarding against negative transfer. We propose GLADE, an adaptive structural-transfer procedure that uses auxiliary category labels to guide local score learning from current unlabeled data and reference inliers. GLADE further combines these local scores with a global anchor and selects their contribution according to estimated discovery utility. A symmetric learning-and-selection construction ensures finite-sample false discovery rate (FDR) control under exchangeability conditions. Our theory characterizes the potential power gains from structural transfer, the costs of learning and selection, and conditions for protection against negative transfer. Experiments on synthetic and real datasets demonstrate power improvements over matched global baselines, with empirical FDR below the nominal level. Together, these results establish a principled approach to exploiting auxiliary structure for out-of-distribution detection with rigorous control of false discoveries.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Adaptive model selection, conformal inference, false discovery rate control, negative transfer, out-of-distribution testing, structural transfer

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_LYfuOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposes GLADE, a framework for improving existing outlier detectors using auxiliary outlier category labels. It uses these labels to guide local learning on reference inliers and unlabeled target data, then combines the local scores with a global detector through adaptive shrinkage. The paper establishes finite-sample FDR control under exchangeability and symmetry conditions and gives conditional results on power gains and negative transfer. Simulations and four real data benchmarks show improvements over the corresponding global baselines.

Critical strengths

GLADE separates learning category structure from auxiliary data and learning detection scores from current target data. This is a reasonable approach when auxiliary and target anomaly distributions differ, since it allows the detector to adapt without directly transferring source trained detection scores.

Critical weaknesses
  1. GLADE's specific methodological novelty is not sufficiently clear. The main difference appears to lie in the structural routing and local target learning in Section 2.2. However, the paper provides limited explanation and evidence of what distinct benefits the proposed weighted local fitting and aggregation (Eq. (2)) offer over existing ways of using auxiliary information.

  2. The simplicity of individual local models does not establish the complexity or efficiency of GLADE as a whole. The paper emphasizes using simple local models to handle heterogeneous anomalies, but GLADE includes a router A(x)A(x), K0K_0 local models, and repeated training and selection. Using simple logistic models alone is therefore insufficient to assess the overall method’s efficiency. An explicit complexity comparison with global models using the same information is needed.

  3. The connection between the theoretical results and the actual GLADE implementation is limited. Theorem 1 establishes finite-sample FDR control, but its main argument relies on existing adaptive score conformal results (Marandon et al., 2024). It does not establish a distinct statistical advantage of GLADE’s proposed structural learning. Theorems 3 and 4 provide conditions for power gains and protection against negative transfer, but there is insufficient evidence that the implemented GLADE procedure satisfies these conditions and achieves these benefits.

  4. The experiments do not identify which components of GLADE drive the performance gains. They show improvements from applying GLADE as a whole, but structural routing, local target learning, score combination, and selection are added together. Although the comparisons include baselines that already use auxiliary information, comparisons with alternative methods using the same information and component level ablations are limited. It is therefore difficult to determine whether the observed gains come from GLADE’s newly proposed components.

  5. The main text does not bring together the explanations needed to understand the key contributions. It provides an overview of the method, but detailed comparisons with existing methods appear in Appendix A.3, and the full algorithm appears in Appendix A.1. The main text should more clearly distinguish existing components from new contributions and explain the full training procedure.

Questions for authors
  1. Given existing methods that use auxiliary information and adaptive selection, what exactly is GLADE’s main novelty? In particular, what additional benefits do category specific weighted fitting and aggregation provide?

  2. Does “simple model” refer to the form of each individual learner, or to the complexity or efficiency of the overall method? Can the authors compare GLADE with a global learner of similar overall model complexity? How do GLADE’s complexity and performance change as the number of categories increases?

  3. What evidence shows that the implemented GLADE procedure satisfies the conditions in Theorems 3 and 4? Is there evidence that the minimum-gain threshold used in the implementation limits negative transfer?

  4. Can the authors compare GLADE with an alternative that replaces the proposed category-specific weighted local fitting and aggregation with another learning approach, while keeping the same DauxD_{aux}, DtrD_{tr}, unlabeled pool, global-local shrinkage, λ\lambda selection, and conformal calibration? Can such comparisons identify which components of GLADE contribute to the performance gains?

6 comments
AuthorsOpenReview ↗

Auxiliary categories and anomaly directions

We thank the reviewer for asking us to clarify GLADE's contribution and connect the theory more closely to its implementation. Our additional comparisons examine how auxiliary categories help target learning, which fitting choices matter, and what computational cost and adaptive safeguards accompany the gains.

Q1.

Our goal is to improve the power of OOD testing under FDR control by exploiting two complementary sources of structure: (a) the current test data contain information about departures from the inlier distribution in the target population, and (b) labeled auxiliary outliers reveal distinctions among anomaly types. AdaDetect (Marandon et al., Annals of Statistics, 2024) learns from the first source by comparing the current unlabeled pool with reference inliers. Integrative conformal pp-values (ICP) use labeled outliers to reweight inlier-based evidence and then recalibrate the resulting score (Liang et al., JRSSB, 2024). The challenge is to bring these two sources of structure together in a single detection rule.

GLADE introduces a framework for learning a single detection score from structure in the test data and outlier category labels in auxiliary training data. GLADE brings together ICP’s strength of using labeled outliers and AdaDetect’s ability to adapt detection to the current test population. The auxiliary outlier categories indicate where to look for distinct departures. Within each such region, GLADE compares the current unlabeled pool with reference inliers to learn which features distinguish current anomalies. GLADE thus transfers what the auxiliary data reveal about anomaly types while learning from the current data how to detect those anomalies. To guard against negative transfer, GLADE uses the estimated gain in discoveries to determine how much weight to give this category-guided refinement and retains the global detector when the estimated gain is too small.

To examine when category-guided learning helps, we vary anomaly-direction separation in a fixed-norm Gaussian family, under balanced and imbalanced category proportions. GLADE and the parameter-matched interaction logistic learner receive the same routing information, without target category labels.

Table R8. Detection power as anomaly directions separate. Four Gaussian anomaly types in 20 dimensions have fixed target and auxiliary mean norms 44 and 4.64.6; reference, calibration, and test sizes are 2,000, 1,000, and 1,000. GLADE and interaction LR use the same auxiliary-category router; no method receives target category labels. Auxiliary size is 1,000, test anomaly prevalence is 20%20\%, and final conformal BH uses α=0.10\alpha=0.10. Interaction LR uses routing-by-feature interactions with the same coefficient count as GLADE; both share the AdaDetect anchor. Cells in (a–c) are mean power (%) [95% Monte Carlo tt CI]; bold means the largest power within each row. Part (d) reports a paired difference in percentage points. The direction sweep changes both routing distinguishability and scoring geometry.

(a) Balanced category proportions

θ (degrees)GLADEInteraction LRAdaDetect
091.42 [90.81,92.02]91.10 [90.52,91.67]91.28 [90.66,91.90]
3083.02 [82.08,83.96]82.28 [81.28,83.29]82.62 [81.67,83.58]
6071.37 [69.95,72.79]61.73 [59.88,63.59]42.99 [41.14,44.84]
9073.14 [71.95,74.32]56.64 [54.60,58.68]12.97 [11.60,14.34]

(b) Category proportions (0.7,0.1,0.1,0.1)(0.7,0.1,0.1,0.1)

θ (degrees)GLADEInteraction LRAdaDetect
091.42 [90.81,92.02]91.10 [90.52,91.67]91.28 [90.66,91.90]
3085.59 [84.68,86.50]84.88 [83.98,85.78]85.74 [84.85,86.62]
6074.77 [73.90,75.65]70.25 [69.41,71.08]65.49 [64.33,66.64]
9075.13 [74.29,75.96]68.27 [67.55,68.98]56.59 [55.33,57.86]

(c) Minority-type power at 90∘90^\circ under imbalance

Minority typeGLADEInteraction LRAdaDetect
Type 235.15 [31.89,38.42]15.57 [13.33,17.81]6.93 [5.82,8.05]
Type 334.26 [30.80,37.73]15.26 [12.87,17.64]6.69 [5.45,7.94]
Type 432.89 [29.69,36.09]14.24 [12.18,16.30]6.55 [5.41,7.70]

(d) Paired gain at orthogonal directions

Comparison / settingGain (pp)Paired 95% CI
GLADE minus interaction LR, balanced 90°+16.50[15.06,17.94]

Table R8 supports two benefits of auxiliary-guided target learning:

  • Greater power for heterogeneous departures. When anomaly directions are well separated, GLADE retains more power than both its anchor and the interaction learner, showing the value of fitting category-specific contrasts.
  • Improved detection of minority categories. Under imbalance, the gains extend to each minority anomaly type when directions are orthogonal, supporting the use of auxiliary structure to learn departures that are sparsely represented in the current pool.
AuthorsOpenReview ↗

Model complexity and computational cost

Q2.

GLADE exploits the category structure of labeled auxiliary outliers to learn a flexible detector from limited current data. A global learner must discover where anomaly patterns differ and which feature combinations reveal each pattern. Learning both is often impractical when the unlabeled pool contains few examples of each type. GLADE uses auxiliary categories to guide the first task, allowing each local learner to estimate how the current pool differs from reference inliers in its region. Here, “simple” refers to the functional form of each local learner, such as a weighted logistic regression. Combining these fits yields a global score that can detect departures along region-specific directions. Without this structural guidance, fitting a single model flexible enough to capture the same heterogeneity can require substantially more current data than are available in practice.

Theorem 2 gives an exact example of this principle. In its Gaussian model with known groups, the group-specific linear scores together recover the likelihood-ratio ranking. When equally strong shifts point in different directions, this rule has strictly higher power than the best common linear direction. Under the conditions of Theorem 3, the fitted GLADE score retains a power advantage over its anchor when the population gain exceeds the errors from learning and selection.

To compare learners of similar size, we gave a single logistic regression the same features, global anchor, and routing weights, with routing-by-feature interactions and as many regression coefficients as the local models combined. We kept the same data and eight underlying anomaly types and merged auxiliary labels to obtain K0=2,4,8K_0=2,4,8. We also measured full-pipeline CPU time, including routing, cross-fitting, selection, and final fitting.

Table R9. Power and cost as auxiliary labels distinguish more anomaly types. A fixed eight-type, 20-dimensional Gaussian task merges auxiliary labels to obtain K0=2,4,8K_0=2,4,8. GLADE and a single interaction logistic learner use the same routing information and K0(d+2)K_0(d+2) refinement coefficients, excluding the shared router and global anchor; they fit separate weighted local objectives and one routing-interaction objective, respectively. Part (a) reports mean powers and paired gains, with 95% Monte Carlo tt intervals; intervals are not truncated. Part (b) profiles the same task separately with one BLAS thread per process, covering routing, cross-fitting/selection, and final fitting. The share column is median cross-fit/selection CPU divided by median full-pipeline CPU. Part (c) reports mean FDP for the power comparison.

(a) Power (%) and paired gain [95% CI]

K0K_0Matched refinement coefficientsGLADE powerInteraction powerGLADE gain (pp)
2440.89 [0.29,1.50]0.27 [-0.12,0.66]+0.63 [-0.04,+1.30]
4889.34 [6.89,11.78]2.09 [0.94,3.25]+7.24 [+5.10,+9.39]
817657.04 [54.56,59.52]33.36 [29.66,37.05]+23.68 [+20.24,+27.12]

(b) Median full-pipeline CPU time (seconds)

K0K_0GLADEInteraction LRAnchorGLADE cross-fit/ selection share (%)
20.0940.0770.01177.4
40.1530.0930.01178.4
80.2500.1100.01179.5

(c) Mean FDP (%) in the power comparison

K0K_0GLADEInteraction LR
21.290.38
43.381.55
87.396.42

Table R9 shows how power and computation change with category resolution:

  • Finer categories strengthen the power advantage. As auxiliary labels distinguish more anomaly types, GLADE gains more power over the interaction learner despite matching its refinement coefficient count. This supports using category structure through separate local fits.
  • The measured cost remains modest. Full-pipeline times remain short over the category counts studied. The comparison includes the repeated fitting and selection steps, providing an assessment of the complete procedure.

The full procedure fits the router once and K0K_0 local models in each round of cross-fitting and in the final refit. With BB repetitions of VV-fold cross-fitting, and writing CAC_A, CGC_G, and ClocalC_{\mathrm{local}} for the router, anchor, and single-local-model fitting costs, the total fitting cost is approximately CA+(BV+1){CG+K0Clocal}C_A+(BV+1)\{C_G+K_0C_{\mathrm{local}}\}. All mixing weights reuse the same fitted scores.

AuthorsOpenReview ↗

Theoretical safeguards and empirical performance

Q3.

Thank you for asking us to connect the theoretical safeguards to their implementation.

(i) Theory and implementation. GLADE includes the exact anchor, a finite mixing grid, and the prescribed selection and tie rules. Theorem 3 shows when a structural gain survives learning and selection error. Theorem 4 gives utility protection on its uniform-accuracy event when dmin⁡=Δmin⁡/N≥2ηseld_{\min}=\Delta_{\min}/N\geq2\eta_{\mathrm{sel}}, where N=ncal+mN=n_{\mathrm{cal}}+m and ηsel\eta_{\mathrm{sel}} bounds candidate-evaluation error. These results remain conditional on the stated statistical assumptions, including control of the pooled-to-target discrepancy; the implementation does not certify this error bound. FDR control follows separately from Theorem 1 under its exchangeability and symmetry conditions.

(ii) Empirical support. To assess how selection balances power and protection, we compare adaptive and fixed weights across five dataset–anchor settings, holding the fitted candidates, data, and final conformal calibration fixed.

Table R10. Fixed versus adaptive mixing on the same real-data candidates. Terrain and terrain + soil are Covertype representations; CSE is CSE-CIC-IDS2018; Gas is Gas Sensor Drift. Each column fixes the dataset and global anchor. λ=0\lambda=0 uses only the anchor, and λ=1\lambda=1 uses only the local score. Underlined entries indicate the highest mean power among fixed-weight policies in each column, including ties at the displayed precision. Part (b) reports the percentage of all evaluated outcomes with fewer true discoveries than the matched anchor; it is distinct from FDP. Final conformal BH uses α=0.10\alpha=0.10.

(a) Mean final-test power (%)

Mixing policyTerrain MCCTerrain + soil MCCUNSW-NB15 MCCCSE MCCGas BIC
Fixed λ=0 (anchor)16.0930.1357.1447.5388.82
Fixed λ=0.2524.9732.5061.1647.9490.91
Fixed λ=0.528.0233.1662.5347.96‾\underline{47.96}92.51‾\underline{92.51}
Fixed λ=0.7530.2333.26‾\underline{33.26}63.06‾\underline{63.06}47.9091.91
Fixed λ=1 (local)32.59‾\underline{32.59}33.26‾\underline{33.26}60.7647.0686.77
Adaptive (GLADE)32.5933.1562.4947.9891.69

(b) Harmful-outcome rate relative to the anchor (%)

Mixing policyTerrain MCCTerrain + soil MCCUNSW-NB15 MCCCSE MCCGas BIC
Pure local (λ=1)04107651
Adaptive (GLADE)010234

Table R11. Adaptive shrinkage on Gas Sensor Drift/BIC and CSE-CIC-IDS2018/MCC. Local harms means pure-local scoring has fewer true discoveries than the matched anchor; other is its complement. Part (a) reports selected-weight frequencies within each group and the mean selected local weight. Part (b) reports mean power differences from the anchor within the harmful-local subset, with paired 95% Monte Carlo tt intervals. These conditional means are distinct from the overall harmful-outcome rates in Table R10(b). Groups are retrospective; test anomaly labels are used only for evaluation.

(a) Selected-weight frequencies (%) and mean

Setting / anchorGroup0.25.5.751Mean λ̂
Gas / BICLocal harms27.513.745.113.70.00.363
Gas / BICOther22.412.244.920.40.00.408
CSE / MCCLocal harms25.040.814.513.26.60.339
CSE / MCCOther20.825.016.720.816.70.469

(b) Mean power difference from the anchor (pp) [95% CI]

Setting / anchorPure localAdaptive
Gas / BIC-7.88 [-10.09,-5.66]+1.58 [+0.99,+2.18]
CSE / MCC-1.45 [-1.64,-1.26]+0.19 [-0.15,+0.52]
  • Reduced negative transfer. Adaptive mixing substantially reduces harmful outcomes on Gas/BIC and CSE/MCC (Table R10(b)). Within the retrospective harmful-local subsets, smaller selected local weights accompany recovery of mean power above the anchor on Gas and approximately to its level on CSE (Table R11), supporting shrinkage as a way to limit unreliable local influence.
  • Preserved detection gains. Adaptive mixing improves mean power over the anchor in every setting and stays close to the best fixed-weight mean, despite the preferred weight varying across datasets (Table R10(a)). Selection uses no test anomaly labels.

The matched routing and learning ablations in our response to Q4 assess these contributions while holding the remaining pipeline fixed (Tables R13 and R14).

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

The paper introduces a method called GLADE to detect out-of-distribution (OOD) samples. This method is based on using auxiliary labeled outliers whose distribution could be different from that of the current anomalies, a.k.a. OOD samples. The method first obtains a multiclass classifier by training on the auxiliary labeled outliers. Since the auxiliary labeled outliers contain different categories of outliers, the classifier assigns weights according to these categories, which are called local scores. Then, the method combines these scores with those of an existing OOD detector. The main objective is to maximize the average detection power subject to controlling the false discovery rate. The paper provides some theoretical guarantees for the proposed method and shows that the proposed method outperforms baseline detectors empirically across different datasets.

Critical strengths

Interesting methodological contribution: The paper proposes a method to exploit historical auxiliary data whose distribution can be different from that of the current dataset. In this method, instead of directly transferring scores learned from auxiliary data, it uses the auxiliary dataset to guide the learning of local detection scores over the current data. This adaptive combination of scores results in a flexibility to limit negative transfer.

Theoretical and empirical validation: In Theorem 1, the paper establishes finite-sample FDR control under exchangeability. Theorem 2 shows that under a Gaussian model category-specific scoring directions can outperform a common linear score in the sense of achieving a higher power. Experiments on synthetic and four real-world datasets show the proposed method outperforms five baseline detectors. It achieves higher number of true discoveries, while keeping FDR below the target level.

Critical weaknesses

Limited protection against negative transfer: The proposed method chooses the global–local weight using discovery counts over the aggregation of test and calibration data points. The best score for the test data is not necessarily the best score for this aggregated data, since adding calibration inliers changes the outlier proportion. However, in Theorem 4, the authors assume sufficiently accurate candidate evaluation. Even in Example B.2, the authors show that this mismatch can exist even with unlimited data. This limits the claimed resilience against negative transfer.

Questions for authors

Under what practical conditions does using discovery counts on the pooled calibration satisfy the accuracy assumption in Theorem 4?

2 comments
AuthorsOpenReview ↗

When pooled discoveries can guide selection

We thank the reviewer for highlighting the distinction between pooled selection and final testing. Our response has three parts: (i) a population condition under which pooled selection preserves the preferred candidate; (ii) matched evidence that adaptive shrinkage mitigates negative transfer; and (iii) a comparison showing that adaptive mean power remains close to the best fixed-weight mean across datasets.

1. When pooled discoveries can guide selection.

We appreciate the request to clarify the practical conditions behind Theorem 4. A useful sufficient case is a calibrated candidate whose target p-value CDF dominates those of the other candidates over the relevant testing range. Pooling with calibration inliers preserves this ordering, providing a population regime in which pooled discoveries can guide selection.

At the population level, consider fixed candidate scores with exact continuous null calibration. Let Fλ(t)F_\lambda(t) be the target-population p-value CDF and r=ncal/mr=n_{\mathrm{cal}}/m. Adding calibration inliers gives Fλ,rpool(t)=Fλ(t)+rt1+r.F_{\lambda,r}^{\mathrm{pool}}(t)=\frac{F_\lambda(t)+rt}{1+r}. Pooling therefore scales differences between candidate CDFs by the positive factor 1/(1+r)1/(1+r). If Fλ⋆(t)≥Fλ(t)F_{\lambda^\star}(t)\geq F_\lambda(t) for every candidate and 0≤t≤α0\leq t\leq\alpha, then λ⋆\lambda^\star is a population maximizer of BH discovery utility and power in both tasks. Empirical recovery additionally requires a separated pooled optimum, stable fits and BH crossings, and a gain that clears the selection margin despite estimation error. When candidate CDFs cross, as in Example B.2, pooling can change the BH-optimal choice.

Proposition B.2 gives one sufficient regime: as the samples grow, the calibration/test ratio vanishes, fold-trained and fully refitted scores approach the same limits, and the limiting BH cutoffs are stable under small perturbations. Our benchmarks assess selection at positive calibration/test ratios. In this regime, a separate bound on the pooled-to-target gain discrepancy remains necessary, and we complement the asymptotic analysis with matched selection comparisons.

Theorem 4 accordingly remains conditional: for its bound to certify no utility loss, the minimum-gain margin must be at least twice the uniform candidate-evaluation error. Stable results across repeated splits are useful diagnostics, but neither that stability nor consistency alone justifies the default one-count margin. These requirements concern power; finite-sample FDR control relies on the exchangeability and symmetry conditions of Theorem 1.

AuthorsOpenReview ↗

Adaptive shrinkage and fixed-weight comparisons

2. Adaptive shrinkage limits negative transfer while retaining local gains.

To assess whether pooled selection limits negative transfer in practice, we compare the submitted adaptive selector with fixed weights in {0,0.25,0.5,0.75,1}\{0,0.25,0.5,0.75,1\}, holding the fitted global and local scores, data, and final conformal calibration fixed. Here λ=0\lambda=0 is the global anchor and λ=1\lambda=1 is pure-local scoring. The comparison covers five dataset–anchor settings, including two Covertype representations, with final BH at α=0.10\alpha=0.10 (Table R6). A harmful outcome has fewer true discoveries than the matched anchor; this criterion measures a loss in power and is distinct from FDP.

Table R6. Fixed versus adaptive mixing on the same real-data candidates. Terrain and terrain + soil are Covertype representations; CSE is CSE-CIC-IDS2018; Gas is Gas Sensor Drift. Each column fixes the dataset and global anchor. λ=0\lambda=0 uses only the anchor, and λ=1\lambda=1 uses only the local score. Underlined entries indicate the highest mean power among fixed-weight policies in each column, including ties at the displayed precision. Part (b) reports the percentage of all evaluated outcomes with fewer true discoveries than the matched anchor; it is distinct from FDP. Final conformal BH uses α=0.10\alpha=0.10.

(a) Mean final-test power (%)

Mixing policyTerrain MCCTerrain + soil MCCUNSW-NB15 MCCCSE MCCGas BIC
Fixed λ=0\lambda=0 (anchor)16.0930.1357.1447.5388.82
Fixed λ=0.25\lambda=0.2524.9732.5061.1647.9490.91
Fixed λ=0.5\lambda=0.528.0233.1662.5347.96‾\underline{47.96}92.51‾\underline{92.51}
Fixed λ=0.75\lambda=0.7530.2333.26‾\underline{33.26}63.06‾\underline{63.06}47.9091.91
Fixed λ=1\lambda=1 (local)32.59‾\underline{32.59}33.26‾\underline{33.26}60.7647.0686.77
Adaptive (GLADE)32.5933.1562.4947.9891.69

(b) Harmful-outcome rate relative to the anchor (%)

Mixing policyTerrain MCCTerrain + soil MCCUNSW-NB15 MCCCSE MCCGas BIC
Pure local (λ=1\lambda=1)04107651
Adaptive (GLADE)010234

Table R6(b) shows that adaptive mixing reduces harmful outcomes wherever they occur under pure-local scoring, with the largest reductions on Gas/BIC and CSE/MCC. The gain in average power is therefore accompanied by substantially fewer losses relative to the anchor.

To understand this mitigation, we examine selected weights and power changes on Gas and CSE, grouping outcomes retrospectively according to whether pure-local scoring harms the anchor.

Table R7. Adaptive shrinkage on Gas Sensor Drift/BIC and CSE-CIC-IDS2018/MCC. Local harms means pure-local scoring has fewer true discoveries than the matched anchor; other is its complement. Part (a) reports selected-weight frequencies within each group and the mean selected local weight. Part (b) reports mean power differences from the anchor within the harmful-local subset, with paired 95% Monte Carlo tt intervals. These conditional means are distinct from the overall harmful-outcome rates in Table R6(b). Groups are retrospective; test anomaly labels are used only for evaluation.

(a) Selected-weight frequencies (%) and mean

Setting / anchorGroup00.25.25.5.5.75.7511Mean λ^\widehat\lambda
Gas / BICLocal harms27.513.745.113.70.00.363
Gas / BICOther22.412.244.920.40.00.408
CSE / MCCLocal harms25.040.814.513.26.60.339
CSE / MCCOther20.825.016.720.816.70.469

(b) Mean power difference from the anchor (pp) [95% CI]

Setting / anchorPure localAdaptive
Gas / BIC−7.88  [−10.09,−5.66]-7.88\;[-10.09,-5.66]+1.58  [+0.99,+2.18]+1.58\;[+0.99,+2.18]
CSE / MCC−1.45  [−1.64,−1.26]-1.45\;[-1.64,-1.26]+0.19  [−0.15,+0.52]+0.19\;[-0.15,+0.52]

Two patterns in Table R7 support the role of shrinkage:

  • Reduced reliance on unreliable local scores. Mean selected local weights are smaller in the harmful-local subsets than in the remaining outcomes, while intermediate mixtures remain common.
  • Useful local information is retained. Within these subsets, adaptive mixing recovers mean power above the anchor on Gas and approximately to its level on CSE. This supports partial shrinkage as a way to limit local losses while retaining detection gains.

3. Adaptive weighting retains near-best mean power across datasets.

The preferred fixed weight varies across datasets, reflecting different degrees of benefit from local refinement. GLADE nevertheless stays close to the best fixed-weight mean on the prespecified grid and improves on the anchor in every setting (Table R6(a)). Thus adaptive selection captures the benefit of dataset-specific mixing without using test anomaly labels.

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

GLADE proposes a framework to improve existing OOD detectors using labeled historical anomalies. It learns a router from historical anomaly categories, trains multiple local classifiers on current unlabeled data, and adaptively combines their scores with an existing detector. Conformal calibration and BH are used to control the FDR.

Critical strengths

The problem is meaningful, and the proposed framework is reasonable. Learning from current data while using historical category information may reduce the impact of distribution shifts. The method can also be applied to different existing detectors.

The paper provides theoretical FDR guarantees under stated assumptions. Experiments on four real-world datasets show consistent improvements in detection power while maintaining FDR control.

Critical weaknesses

The contribution of structural transfer is not sufficiently isolated. It is unclear whether the improvements come from informative auxiliary-category routing or from additional target-adaptive learning and model capacity. Moreover, the experiments do not establish the benefit of separately training the router and multiple local classifiers compared with a single capacity-matched target-adaptive model using the same information. Figure 9 further suggests that GLADE's incremental gains can diminish when more informative features strengthen the global detectors.

The adaptive selection mechanism is not sufficiently evaluated. Although the paper provides conditional guarantees against negative transfer, it does not systematically report the selected λ\lambda values or fallback frequencies on real datasets, nor does it compare adaptive selection with fixed mixing weights. The reliability of the selection criterion is also unclear when auxiliary information is misleading or when the selection and testing populations differ. Consequently, the practical contribution of adaptive shrinkage and its protection against negative transfer remain insufficiently demonstrated.

The experimental analysis is also limited. The Gaussian simulations use only one informative feature direction, while Figure 9 suggests that GLADE's additional gains can shrink considerably when the global detector receives more informative features.

Questions for authors
  1. How much of GLADE's improvement comes from the auxiliary-category router? Would simpler alternatives, such as random routing or reusing category predictions from MCC, achieve similar results?

  2. Why are multiple local classifiers needed instead of a single target-adaptive classifier using the same features, global scores, and routing information? How much benefit does the proposed design provide?

  3. Could the authors report the selected λ\lambda values across datasets and compare adaptive selection with fixed weights? How effectively does the fallback mechanism prevent negative transfer when auxiliary information is unreliable?

4 comments
AuthorsOpenReview ↗

Category routing, Figure 9, and local fits

We thank the reviewer for recognizing the relevance of the problem and for asking us to clarify the source of the gains. Our additional experiments examine how auxiliary categories help target learning and how reliably adaptive mixing uses that information.

Responses to the questions

Q1.

Thank you for suggesting these controls. To determine whether the gains depend on meaningful category structure, we replaced informative routing with uniform weights, label-independent random weights, or a router trained on shuffled auxiliary labels. The anchor, data, selection rule, and calibration procedure were held fixed.

Table R1. Contribution of auxiliary-category routing. Part (a) replaces informative routing with uniform or label-independent random weights in both local fitting and aggregation, retaining the nominal number of branches. Part (b) trains the router on shuffled auxiliary category labels. Within each comparison, the global anchor, data, cross-fitting, mixing selection, and final conformal BH at α=0.10\alpha=0.10 are matched. Covertype uses terrain or terrain + soil representations; the other benchmarks are UNSW-NB15, Gas Sensor Drift, and CSE-CIC-IDS2018. Power is in percent; gains are in percentage points. Intervals are 95% Monte Carlo tt intervals, paired for gains. All compared methods have mean FDP below 10%10\%.

(a) Covertype terrain: mean power (%) [95% CI]

RoutingAdaDetect anchorMCC anchor
Informative32.48 [32.18,32.79]32.59 [32.30,32.88]
Uniform19.14 [18.07,20.21]23.90 [23.38,24.42]
Random19.13 [18.07,20.19]23.78 [23.28,24.28]

(b) Informative minus shuffled-category routing

Setting / anchorGain (pp)Paired 95% CI
Terrain / AdaDetect+13.33[12.36,14.31]
Terrain + soil / AdaDetect+2.59[2.36,2.83]
UNSW-NB15 / AdaDetect+19.32[18.70,19.94]
Gas Sensor Drift / BIC+1.94[0.83,3.05]
CSE-CIC-IDS2018 / MCC+0.41[0.15,0.66]
  • Meaningful routing contributes beyond adding local branches. On Covertype terrain, informative routing yields substantially higher power than uniform or random routing with the same number of branches, under both anchors.
  • Auxiliary labels provide useful structure across datasets. Informative routing outperforms shuffled-label routing across all four real datasets, supporting the use of category information to organize target learning.

We clarify that GLADE already uses category predictions from an auxiliary multiclass classifier (MCC) for routing. This classifier is fitted to the labeled auxiliary outliers, and its predicted category probabilities define the soft routing weights Ac(x)A_c(x) (Section 2.2). This classifier supplies category information to the local learners; the global anchor supplies the baseline detection score.

These results also clarify Figure 9. GLADE is designed to complement an anchor that has not fully captured useful anomaly structure. With limited current data, the unlabeled pool may contain too few anomalies of each type for a global learner to estimate their distinct feature patterns reliably. Auxiliary categories guide where to look for these departures, while category-weighted target–reference contrasts learn how to detect them in the current population. When richer features already capture much of this structure, there is correspondingly less scope for refinement.

Q2.

Multiple local fits are useful when anomaly categories require different scoring directions. Each expert learns a category-weighted target–reference contrast before aggregation. To quantify this benefit, we use a balanced four-type Gaussian setting with orthogonal anomaly directions and hold the features XX, global scores GG, routing weights AA, data, cross-fitting, mixing grid, selection, and final calibration fixed. We compare GLADE with the following alternatives:

  • Joint probability mixture: retains GLADE's expert family, coefficient count, and routing-weighted aggregation, but fits all experts jointly through the aggregate binary cross-entropy instead of separate category-weighted losses.
  • Interaction logistic regression: replaces the expert mixture with one logistic learner using routing-by-feature interactions and the same coefficient count.
  • Single MLP: replaces the local module with a nonlinear learner taking (G,X,A)(G,X,A) as inputs, with slightly more parameters than GLADE.
AuthorsOpenReview ↗

Matched learners and adaptive mixing

Thank you for asking us to isolate the benefit of local learning.

Table R2. Same-information comparison: four balanced, orthogonal Gaussian anomaly types (d=20d=20; target/auxiliary mean norms 4/4.64/4.6). Reference/auxiliary sizes are 2,000/1,000; calibration and test sizes are 1,000 each, with 20%20\% anomalies. All refinements share the AdaDetect anchor, router, cross-fitting, mixing selection, and final conformal BH at α=0.10\alpha=0.10. Refinement counts exclude the shared router/anchor; the anchor row reports its own coefficients. Joint mixture jointly fits GLADE's prediction family with the same aggregation; interaction LR uses route–feature products (C=0.1C=0.1); the MLP has one hidden layer. Power and FDP are percentages; paired gains are percentage points. Intervals are 95% Monte Carlo tt intervals.

(a) Mean final-test performance

MethodParametersPower (%)95% CIFDP (%)
GLADE8874.40[72.95,75.85]7.96
Joint probability mixture8858.50[56.18,60.83]7.89
Interaction logistic regression8858.18[55.59,60.77]7.88
Single MLP10922.84[19.70,25.98]7.67
Global anchor2114.01[12.08,15.95]7.60

(b) Paired power gains over alternative refinements

ComparatorGLADE gain (pp)Paired 95% CI
Joint probability mixture+15.90[14.48,17.31]
Interaction logistic regression+16.22[14.53,17.90]
Single MLP+51.56[48.72,54.40]

Table R2 highlights two benefits:

  • Separate weighted fitting improves learning within the same score family. GLADE substantially outperforms the jointly fitted mixture despite identical routing, aggregation, and coefficient counts. This comparison directly supports learning category-specific target–reference contrasts separately.
  • The gain persists against single-model alternatives. GLADE also outperforms the parameter-matched interaction model and the larger MLP. Wider-network and regularization checks preserve this advantage in the same setting, and the fitted controls meet their prescribed stopping criteria.

Q3.

Thank you for requesting a closer evaluation of adaptive mixing. To examine whether it retains local gains while limiting negative transfer, we compare Sλ=(1−λ)G+λSlocalS_\lambda=(1-\lambda)G+\lambda S_{\mathrm{local}} at every fixed grid weight with the submitted selector, using identical fitted candidates, data, and final calibration. Table R3 separates power from negative-transfer frequency.

Table R3. Fixed versus adaptive mixing on the same real-data candidates. Terrain and terrain + soil are Covertype representations; CSE is CSE-CIC-IDS2018; Gas is Gas Sensor Drift. Each column fixes the dataset and global anchor. λ=0\lambda=0 uses only the anchor, and λ=1\lambda=1 uses only the local score. Underlined entries indicate the highest mean power among fixed-weight policies in each column, including ties at the displayed precision. Part (b) reports the percentage of all evaluated outcomes with fewer true discoveries than the matched anchor; it is distinct from FDP. Final conformal BH uses α=0.10\alpha=0.10.

(a) Mean final-test power (%)

Mixing policyTerrain MCCTerrain + soil MCCUNSW-NB15 MCCCSE MCCGas BIC
Fixed λ=0 (anchor)16.0930.1357.1447.5388.82
Fixed λ=0.2524.9732.5061.1647.9490.91
Fixed λ=0.528.0233.1662.5347.96‾\underline{47.96}92.51‾\underline{92.51}
Fixed λ=0.7530.2333.26‾\underline{33.26}63.06‾\underline{63.06}47.9091.91
Fixed λ=1 (local)32.59‾\underline{32.59}33.26‾\underline{33.26}60.7647.0686.77
Adaptive (GLADE)32.5933.1562.4947.9891.69

(b) Harmful-outcome rate relative to the anchor (%)

Mixing policyTerrain MCCTerrain + soil MCCUNSW-NB15 MCCCSE MCCGas BIC
Pure local (λ=1)04107651
Adaptive (GLADE)010234

The fixed-weight comparison supports both aims:

  • Adaptive weighting retains near-best mean power across datasets. Although the best fixed weight varies, adaptive mixing remains close to the best fixed-weight mean and improves on the anchor in every setting. It therefore captures the benefit of dataset-specific mixing without using test anomaly labels.
  • Adaptive mixing reduces harmful outcomes. Relative to pure-local scoring, it substantially lowers the frequency of outcomes with fewer true discoveries than the anchor, particularly on Gas and CSE. This supports adjusting local influence to the current data.

To examine how this protection arises, Table R4 reports selected weights and fallback frequencies on Gas and CSE, separating outcomes according to whether pure-local scoring harms the anchor. These groups are defined retrospectively for evaluation only.

  • Selected weights reflect local reliability. Mean local weights are smaller when pure-local scoring is harmful, consistent with reduced reliance on unreliable local information.
AuthorsOpenReview ↗

Shrinkage and anomaly directions

All final tests continue to use BH at α=0.10\alpha=0.10.

Table R4. Adaptive shrinkage on Gas Sensor Drift/BIC and CSE-CIC-IDS2018/MCC. Local harms means pure-local scoring has fewer true discoveries than the matched anchor; other is its complement. Part (a) reports the selected-weight distribution in each group. Parts (b–c) condition on the harmful-local subset: (b) paired-outcome mean power differences with 95% Monte Carlo tt intervals; (c) exact anchor selection, intermediate weights with no remaining loss, and residual loss. The conditional denominator differs from the overall harmful-outcome rates in Table R3(b). Groups are retrospective; test anomaly labels are used only for evaluation.

(a) Selected-weight frequencies (%) and mean

Setting / anchorGroup0.25.5.751Mean λ̂
Gas / BICLocal harms27.513.745.113.70.00.363
Gas / BICOther22.412.244.920.40.00.408
CSE / MCCLocal harms25.040.814.513.26.60.339
CSE / MCCOther20.825.016.720.816.70.469

(b) Mean power difference from the anchor (pp) [95% CI]

Setting / anchorPure localAdaptive
Gas / BIC-7.88 [-10.09,-5.66]+1.58 [+0.99,+2.18]
CSE / MCC-1.45 [-1.64,-1.26]+0.19 [-0.15,+0.52]

(c) Outcomes within the harmful-local subset (%)

Setting / anchorExact anchorPartial shrinkage no lossRemaining lossTotal loss avoided
Gas / BIC27.564.77.892.2
CSE / MCC25.046.128.971.1
  • Partial shrinkage preserves useful local information. Within these harmful-local subsets, adaptive mixing recovers mean power to approximately the anchor level on CSE and above it on Gas. Intermediate mixtures avoid the local loss more often than exact anchor fallback, supporting shrinkage as an active source of protection.

Additional response to the stated weaknesses

To examine when local directions help, we added a four-category Gaussian family in 20 dimensions, keeping each category's signal norm fixed while increasing the separation between anomaly directions. The auxiliary categories share these directions with a different signal norm. We consider both balanced and imbalanced category proportions, and the learned methods do not receive target category labels.

Table R5. Detection power as anomaly directions separate. Four Gaussian anomaly types in 20 dimensions have fixed target and auxiliary mean norms 44 and 4.64.6; reference, calibration, and test sizes are 2,000, 1,000, and 1,000. GLADE and interaction LR use the same auxiliary-category router; no method receives target category labels. Auxiliary size is 1,000, test anomaly prevalence is 20%20\%, and final conformal BH uses α=0.10\alpha=0.10. Interaction LR uses routing-by-feature interactions with the same coefficient count as GLADE; both share the AdaDetect anchor. Cells in (a–c) are mean power (%) [95% Monte Carlo tt CI]; bold means the largest power within each row. Part (d) reports a paired difference in percentage points. The direction sweep changes both routing distinguishability and scoring geometry.

(a) Balanced category proportions

θ (degrees)GLADEInteraction LRAdaDetect
091.42 [90.81,92.02]91.10 [90.52,91.67]91.28 [90.66,91.90]
3083.02 [82.08,83.96]82.28 [81.28,83.29]82.62 [81.67,83.58]
6071.37 [69.95,72.79]61.73 [59.88,63.59]42.99 [41.14,44.84]
9073.14 [71.95,74.32]56.64 [54.60,58.68]12.97 [11.60,14.34]

(b) Category proportions (0.7,0.1,0.1,0.1)(0.7,0.1,0.1,0.1)

θ (degrees)GLADEInteraction LRAdaDetect
091.42 [90.81,92.02]91.10 [90.52,91.67]91.28 [90.66,91.90]
3085.59 [84.68,86.50]84.88 [83.98,85.78]85.74 [84.85,86.62]
6074.77 [73.90,75.65]70.25 [69.41,71.08]65.49 [64.33,66.64]
9075.13 [74.29,75.96]68.27 [67.55,68.98]56.59 [55.33,57.86]

(c) Minority-type power at 90∘90^\circ under imbalance

Minority typeGLADEInteraction LRAdaDetect
Type 235.15 [31.89,38.42]15.57 [13.33,17.81]6.93 [5.82,8.05]
Type 334.26 [30.80,37.73]15.26 [12.87,17.64]6.69 [5.45,7.94]
Type 432.89 [29.69,36.09]14.24 [12.18,16.30]6.55 [5.41,7.70]

(d) Paired gain at orthogonal directions

Comparison / settingGain (pp)Paired 95% CI
GLADE minus interaction LR, balanced 90°+16.50[15.06,17.94]

Table R5 shows that the benefit extends beyond a single informative direction:

  • Local learning helps when anomaly directions differ. When anomaly directions are well separated, GLADE retains substantially more power than both the anchor and the parameter-matched interaction model, under balanced and imbalanced category proportions.
  • The gains extend to minority anomaly categories. Under imbalance, GLADE improves detection of each minority type when the directions are orthogonal. Mean FDP remains below the nominal level throughout the experiment.

Thank the author for the additional experiments. These solve some of my concerns. The additional Gaussian experiments also help illustrate the benefit of local learning when anomaly categories have different scoring directions.. However, I still have some concerns about the architecture of multiple local classifiers. The Joint Probability Mixture also use the same model with different training objective. The comparisons with single model are limited to synthetic data. It remain unclear whether GLADE can have advantages over stong single models,with or without routing information, on real-world datasets. In this case, i'd like to remain borderline in my assessment.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.