acceptodds
Under review as a conference paper at ICLR 2027

When Less Compute Is More: Adaptive Early Exit Improves Pretrained Outlier Detection

Abstract

Pretrained tabular foundation models process every dataset at a fixed depth, with inference costs growing with dataset size. To address this, we present the first study of depth-adaptive early-exit for pretrained outlier detection models. While early-exit is typically motivated by efficiency, we uncover a surprising benefit: exiting at the optimal intermediate layer can also improve detection performance on diverse real-world benchmarks by 4.7–7.3% on average, consistent across three distinct foundation models. First, we investigate the factors driving these gains, and identify a key mechanism: context pollution, i.e., the presence of outliers among in-context samples. Our analysis reveals that nearby in-context samples exert increasing influence on query predictions at greater depths, consistent with a retrieval-based view of these models. In effect, early-exit alleviates the adverse effects of retrieving accurate-yet-polluted neighbors, with gains of 13–21% when context pollution matches the natural outlier rate. Motivated by these findings, we pretrain a plug-in router to select a dataset-specific exit layer, using query outlier labels as privileged information available only during router training. The router operates post hoc, leaving the base model parameters and prediction head unchanged. Experiments on three large real-world benchmarks show that, on clean context, the router recovers up to 45% of the oracle gain with up to 1.8× speedup across three pretrained backbones, with larger gains as context pollution increases.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Tabular Foundation Models, Outlier Detection, Early-Exit Inference, In-Context Learning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_DADcOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies depth-adaptive early exiting for pretrained tabular outlier detection (OD) foundation models. The authors first observe that intermediate-layer representations, when fed into the frozen final prediction head, can outperform full-depth predictions, and that this "oracle early-exit" gain grows substantially when the in-context set is polluted by outliers. They attribute this to context pollution: deeper layers aggregate more heavily from nearby context samples, so outlier neighbors progressively mask query outliers in the representation. Building on this, they propose ROUT, a lightweight post-hoc router (≈4.6M params) that reads frozen backbone representations, layer-wise outlier scores, raw features, and 29 geometric features, and selects a dataset-specific exit layer. ROUT is pretrained on ~211K synthetic datasets, uses query outlier labels as privileged information (LUPI) only during training, and combines a sequential stopping loss with a layer-wise selection loss and a depth penalty. Backbones and prediction heads remain entirely frozen. Experiments span three OD-TFM backbones (OutFormer, TACTIC, ICLAD), three real-world benchmarks (OddBench 690, OvRBench 755, ADBench 47) and a synthetic benchmark (SynBench 800), under clean, held-out pollution, and near-duplicate pollution settings. ROUT reportedly matches or improves full-depth AUROC in 11 of 12 settings while computing 15–63% fewer layers, closing up to 45% of the oracle gap and achieving up to ~1.8× speedup.

Critical strengths

The paper is well organized: motivating observation → mechanistic explanation → method → extensive evaluation.

(1) A genuinely counter-intuitive and well-supported empirical finding.** The observation that early exit can improve outlier-detection accuracy, not just save compute, and that this effect strengthens with context pollution, is the paper's most consequential contribution. It is supported across three architecturally and training-regime-distinct backbones, four benchmarks, and both injected-pollution modes, with consistent directional results (Table 1, Table 4, Figures 9–13, 18). This is the kind of finding that changes how practitioners think about depth in in-context OD models, and it is not an artifact of a single model or dataset.

Critical weaknesses

(1) The method sidesteps rather than solves the train–test distribution shift that defines the problem.** The paper's own framing is that real-world OD datasets exhibit hidden pollution, yet ROUT is never trained on any real data or any human-annotated outlier labels. The authors' justification is that "inference-time datasets lack ground truth labels," but this conflates test-time label availability with training-time data realism: one could label outliers on a held-out pool of real datasets for router training, or at minimum perform a leave-one-benchmark-out study to quantify prior mismatch. As it stands, the training-data ablation (Table 14) shows 2.3–4.4% AUROC swings from removing a single prior source, which is precisely the symptom of an unaddressed shift. The paper acknowledges prior mismatch (" 0.76") but offers no remedy, no diagnostic, and no characterization of when the router should be distrusted. This is consequential because the entire value proposition rests on synthetic priors being adequate proxies for real pollution.

(2) The central causal claim about context pollution on real data is under-evidenced, and the mechanism lacks formal grounding.** The pollution mechanism is convincingly demonstrated only where the authors inject the pollution. On real benchmarks, the existence and magnitude of hidden pollution is inferred from one t-SNE visualization (Figure 2) plus layer-wise AUROC curves — there is no quantitative estimate of pollution rate, no ablation that removes suspected polluted real context rows and measures the effect on oracle exit depth, and no test of whether the real-world oracle gain survives when pollution is controlled for. Since the real-world gain is the headline result (7.3% / 7.0% / 4.7% for the three backbones), the causal story behind it should be nailed down more tightly than a correlation between sibling influence and AUROC loss on synthetic targeted outliers. Relatedly, the paper never formalizes why optimal exit depth should decrease monotonically with pollution; the explanation remains a retrieval/attention analogy plus empirical correlations.

Questions for authors
  1. Train–test shift. Why not train the router on a held-out pool of real datasets with human- or rule-annotated outlier labels (or at least a small real labeled set), instead of relying exclusively on synthetic priors? Could you report a leave-one-benchmark-out experiment (train without ADBench, test on ADBench, etc.) to quantify how much ROUT's exit selection degrades under prior mismatch? What is the failure mode when a test dataset's pollution pattern is outside the synthetic prior support — does ROUT degrade to full depth, or does it confidently pick a wrong early layer?

  2. Direct outlier identification. ROUT outputs an exit layer, not an outlier score. The pollution experiments use author-injected outliers, and Figure 2's hidden-pollution evidence is qualitative. Can you provide a quantitative estimate of the hidden pollution rate on real benchmarks (e.g., by identifying context rows that a strong supervised detector would flag), and show that removing those rows measurably changes the oracle exit layer? Without this, the claim that real-world early-exit gains are driven by hidden pollution remains correlational.

  3. Baseline fairness. Only LA-Entropy is compared, and it is a weak baseline on this task (below Half-depth for two of three backbones). Why were CALM, Router-Tuning, FlexiDepth, or a simple entropy-based exit with the original head not included? Since LA-Entropy retrains per-layer heads while ROUT does not, is the comparison measuring routing quality or head quality? Also, would the authors consider moving Oracle / Best-fixed-layer out of Table 2 into a clearly separated reference table?

5 comments
AuthorsOpenReview ↗

Response to Reviewer DADc

We sincerely thank the reviewer for the thoughtful and encouraging review. We are glad the reviewer found the paper well organized, and we truly appreciate the recognition of our finding that early exit can improve detection, not just save compute, as "the kind of finding that changes how practitioners think about depth in in-context OD models." The reviewer's questions helped us strengthen the real-data evidence, and we address each of them below. Please see the revisions in blue text in the revised copy of our submission, which is now uploaded on OpenReview.

W1 / Q1. Training on synthetic data only, leave-one-benchmark-out, and the failure mode under prior mismatch.

We thank the reviewer for this important question on the train–test shift.

Why not train on real datasets? Real-world benchmarks are too few to train the router from scratch. Note that one training example for rOUT is an entire dataset, and the three real-world benchmarks together offer only 1,492 such examples, compared to over 200,000 synthetic ones. rOUT therefore follows the standard practice of tabular foundation models such as TabPFN (Hollmann et al., 2023; 2025) and TabICL (Qu et al., 2025; 2026), which are pretrained purely on synthetic data and applied zero-shot to real datasets; the OD-TFM backbones that rOUT builds on are trained the same way. No real-world benchmark is used to train rOUT, and its stopping threshold is selected on synthetic validation data. This is in fact more challenging than the leave-one-benchmark-out setting, and rOUT still recovers up to 45% of the Oracle gain (Table 2).

Leave-one-benchmark-out. Since real datasets are too few to train from scratch, we instead fine-tuned the pretrained rOUT on real labels, in response to reviewer DADc’s request. We fine-tuned on two real-world benchmarks and tested on the third, with early stopping on held-out datasets of those two benchmarks and the stopping threshold selected on synthetic validation data, as in the paper. In our experiments, this yielded a slight gain over zero-shot rOUT, but not a significant one (the 95% CI of the difference includes zero on all three benchmarks):

Table R3: Fine-tuning on real labels does not significantly improve rOUT on the held-out benchmark. Leave-one-benchmark-out fine-tuning, added as Table R3 in Appendix C.3 (mean AUROC at matched depth; Difference = fine-tuned − zero-shot rOUT, with its 95% CI over datasets)

Held-out benchmarkMean computed layersFull depthrOUTrOUT, fine-tunedDifference [95% CI]
ADBench5.5.853.860.867+.007 [−.007, +.023]
OvRBench5.4.709.730.732+.002 [−.001, +.004]
OddBench5.5.740.761.764+.003 [−.001, +.007]

As a sanity check, the router does learn from real labels: on the datasets it is trained on, it reaches 69–88% of the Oracle gain, up from 22–38% before fine-tuning. This fit did not carry over much to the held-out benchmark, suggesting that synthetic pretraining already captures most of what transfers across datasets.

Prior mismatch and failure mode. Outside its training priors, the router becomes slightly more conservative but does not fall back to full depth. To test this, we evaluated the routers of Table 14, each of which has never seen the setting that its removed data source resembles:

Table R4: Outside its training data, a router falls behind rOUT and exits slightly deeper, yet still stops well before full depth. Each router of Table 14 lacks one data source and is evaluated in the setting that source resembles; added as Table R4 in Appendix C.3 (AUROC, mean computed layers in parentheses)

RouterOut-of-support settingAblated routerrOUTFull depth
w/o polluted contextsOvRBench, 1:1 held-out.573 (6.1).586 (5.0).556 (10)
w/o TabPFN priorsOvRBench.716 (6.2).730 (5.4).709 (10)
w/o OutFormer priorsSynBench.947 (5.8).967 (5.7).975 (10)

In all three cases, the mismatched router performs below rOUT and exits slightly deeper, while still stopping well before full depth. Prior mismatch thus shows up as weaker layer choices at a slightly deeper average exit, rather than a collapse to very shallow layers or to full depth. On real-world data, these choices still improve over Full depth. On SynBench Full depth is already close to the Oracle and leaves little for early exit to gain even for rOUT. Here the mismatched router falls further behind Full depth than rOUT does. Our remedy is broad coverage: as Table 14 shows, each data source matters most in the setting it resembles, which is why rOUT trains on all of them.

Revision. We added the fine-tuning and out-of-support results (Tables R3 and R4 above) to Appendix C.3, and stated the zero-shot protocol explicitly in Section 5.

W2 / Q2. Evidence for hidden pollution on real data, and why the optimal exit moves earlier under pollution.

We thank the reviewer for this constructive suggestion on strengthening the real-data evidence.

Quantifying and removing hidden pollution. Hidden pollution is present in most real-world datasets. Removing it narrows the oracle gain and shifts the oracle layer toward full depth, which reverses the effect of added pollution in Section 3. As the reviewer suggested, for each dataset we train a supervised detector (a random forest) on its labeled query rows and flag the context rows it confidently predicts as outliers: these appear in 66% of the datasets and make up 1.2% of the context on average. We then remove them, compared against removing the same number of random rows:

Table R5: Removing flagged hidden pollution raises full-depth AUROC, shrinks the oracle gain and moves the oracle layer deeper; removing random rows has no effect. Added as Table R5 in Section 3 (873 datasets, OutFormer; first row: values on the original context; other rows: change from it; Wilcoxon signed-rank test, ∗∗∗p<.001^{\ast\ast\ast}p<.001)

ContextFull-depth AUROCOracle − Full depthMean oracle layer
Original.675.0596.13
Flagged rows removed+.019$^{\ast\ast\ast}$−.008$^{\ast\ast\ast}$+0.20$^{\ast\ast\ast}$
Same number of random rows removed (Control)−.000+.000−0.00

This simple detector flags only a small share of the context, yet removing these rows significantly narrows the oracle gain and shifts the oracle layer toward full depth. Removing the same number of random rows leaves both unchanged. This intervention on real contexts thus confirms that hidden pollution drives early-exit gains and agrees with the controlled pollution experiments in Section 3.

Controlled intervention on real data. We would like to clarify that the planted-sibling analysis is a controlled intervention on real datasets: for each targeted query outlier we turn one context row into its sibling and keep everything else fixed. Turning the same row into a copy of a random inlier leaves the AUROC of the targeted outliers unchanged (Appendix B.1). The masking is therefore specific to the sibling. Together with the removal experiment above, this establishes context pollution as a major mechanism behind the real-world gains.

Why the optimal exit moves earlier. Here we provide an intuitive decomposition of why more pollution favors an earlier exit, writing the AUROC of layer ll at pollution ratio rr as A(l,r)≈A(l,0)−g(r)S(l)A(l,r)\approx A(l,0)-g(r) S(l). The clean-context AUROC A(l,0)A(l,0) grows with depth, and so does the sensitivity to pollution S(l)S(l) (Figures 5, 14 and 15). Because gg increases with rr, more pollution makes the cost of depth outweigh its benefit at an earlier layer. Indeed, from clean context to 1:1 the mean oracle layer moves earlier for every backbone and real-world benchmark.

Revision. We added the hidden-pollution experiment to Section 3, with details in Appendix C.2.

Q3. Choice of baselines, routing quality versus head quality, and the reference rows of Table 2.

We thank the reviewer for these helpful suggestions on the baselines.

Why not CALM, Router-Tuning or FlexiDepth? Our focus on early exit follows directly from the mechanism we identify. Methods that skip layers (Router-Tuning and FlexiDepth) or also repeat them (Dr.LLM by Heakl et al. (2026) and PoLaR by Li et al. (2026)) decide which layers to compute. Under context pollution the harm is not tied to a particular layer. Instead, it mainly accumulates with depth because every additional layer aggregates the polluted context once more. Stopping early therefore becomes the most natural remedy. In addition, the three methods the reviewer names make per-token decisions in LLMs and aim to match the full model at lower cost. An OD-TFM instead needs one exit layer per dataset, and its intermediate layers can outperform full depth. These methods also retrain the backbone (CALM) or train with a language-modeling loss (Router-Tuning and FlexiDepth), which a frozen OD-TFM does not allow.

Confidence-based exits with the frozen final-layer head. Following the reviewer's suggestion, we added a CALM-style exit and LA-Entropy with the frozen final-layer head, and rOUT outperforms both for every backbone on every real-world benchmark. The CALM-style exit uses the hidden-state saturation of CALM (Schuster et al., 2022), the cosine similarity between query representations of consecutive layers. We do not use its softmax response because with a two-class head it depends only on the predicted outlier probability, like entropy. LA-Entropy with the frozen final-layer head is the simple entropy-based exit the reviewer suggests. Each exit averages its confidence over the query rows and stops at the first layer where it passes a threshold selected on synthetic validation data.

Table R6: rOUT is more accurate than every confidence-based exit on every backbone and real-world benchmark. Excerpt of the revised Table 2 (AUROC, mean computed layers in parentheses)

BackboneMethodOddBenchOvRBenchADBench
OutFormerFull depth.740 (10).709 (10).853 (10)
LA-Entropy (retrained heads).709 (2.5).680 (2.9).770 (2.5)
LA-Entropy (frozen head).701 (7.1).689 (7.6).794 (7.4)
CALM-style.732 (6.9).693 (6.4).839 (5.7)
rOUT.761 (5.5).730 (5.4).860 (5.5)
ICLADFull depth.769 (12).745 (12).891 (12)
LA-Entropy (retrained heads).704 (1.6).696 (2.0).786 (2.1)
LA-Entropy (frozen head).744 (8.8).735 (10.0).883 (10.1)
CALM-style.761 (7.3).728 (6.6).843 (5.5)
rOUT.776 (6.8).755 (6.0).891 (6.0)
TACTICFull depth.736 (12).715 (12).832 (12)
LA-Entropy (retrained heads).693 (2.1).682 (2.4).780 (2.2)
LA-Entropy (frozen head).735 (12.0).712 (11.9).832 (12.0)
CALM-style.736 (12.0).715 (12.0).832 (12.0)
rOUT.747 (10.2).729 (10.0).849 (10.1)

Routing or head quality? To stay aligned with the original method (Küken et al., 2025), the LA-Entropy baseline in Table 2 retrains a head for each layer, and the revised Table 2 also reports it with the frozen final-layer head. The two frozen-head exits instead share the frozen backbone and head with rOUT and differ only in how the exit layer is chosen, so this comparison directly measures routing quality. With the frozen head, confidence proves to be a weak stopping signal: the two exits stop later than rOUT in all but one setting yet are less accurate in every setting. We attribute the advantage of rOUT largely to two core designs that confidence-based exits lack. Sample Attention reads the representations of individual context and query rows and so captures information that a single average confidence cannot. Layer Attention compares all computed layers and lets rOUT look back to return an earlier layer, and removing it significantly lowers AUROC (Table 5).

Reference rows in Table 2. We thank the reviewer for this suggestion. In the revision, the caption of Table 2 states explicitly that Oracle and Best fixed layer use test labels and cannot be deployed.

Revision. We added LA-Entropy with the frozen final-layer head and the CALM-style exit to Table 2, explain the choice of baselines in Section 5, and revised the caption of Table 2.

Finally, we noticed that our paper was flagged for ethics review. Our study uses only publicly available benchmarks and synthetic data, and we are happy to clarify further if the reviewer has a specific concern in mind.

If the new real-data evidence and the added baselines address the reviewer's concerns, we hope the reviewer might consider revisiting the score. We also welcome any further questions during the remaining discussion period.

Reviewer_kvTVOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3(+1)
Read review
Paper summary

This paper investigates post-hoc depth-adaptive early exiting for pretrained tabular outlier detection foundation models, studying three backbones: OutFormer, ICLAD, and TACTIC. The authors observe that intermediate-layer representations, when passed through the frozen final-layer prediction head, can outperform full-depth inference, with the best exit layer varying across datasets. They investigate context pollution as a contributing mechanism, hypothesizing that nearby context outliers can mask query anomalies and that their influence can increase with depth.

Motivated by these findings, the paper proposes rOUT, a post-hoc router that decouples computation depth from prediction-layer selection. It sequentially decides when to stop and retrospectively selects among the layers already computed, without modifying the pretrained backbone or prediction head. The router is trained on synthetic datasets using query labels as supervision and privileged inputs that are progressively withdrawn, requiring no target-dataset labels at inference.

Experiments cover three real-world benchmarks (OddBench, OvRBench, and ADBench) and the synthetic SynBench, under clean, held-out-polluted, and near-duplicate-polluted contexts. On clean real-world benchmarks, rOUT improves or matches full-depth AUROC across all nine backbone–benchmark combinations, while providing measured inference speedups that account for router calls and additional feature computation. The paper further examines depth-dependent behavior, pollution mechanisms, and router component and training-data ablations.

Critical strengths

1. The paper demonstrates a practically useful improvement in both detection quality and inference efficiency without modifying pretrained backbones. The finding that intermediate representations can outperform the final layer using the same frozen prediction head is interesting, especially because the optimal layer varies substantially across datasets. rOUT addresses this opportunity through a well-motivated design that separates the depth computed from the layer ultimately selected for prediction. This retrospective selection is particularly appropriate when additional layers may degrade anomaly detection rather than improve it. The use of privileged query labels during synthetic pretraining, followed by gradual withdrawal of privileged inputs, also provides a reasonable approach to learning a routing policy for unlabeled deployment.

Importantly, the empirical gains are not limited to an oracle analysis. Table 2 shows that rOUT improves AUROC in eight of the nine clean real-world backbone–benchmark combinations and matches full-depth performance in the remaining one. For example, OutFormer improves from 0.740 to 0.761 on OddBench and from 0.709 to 0.730 on OvRBench. Table 3 reports corresponding wall-clock speedups of 1.61× and 1.77×, respectively. These measurements include router calls and additional feature computation, and the reported layer budgets count the depth actually computed rather than merely the layer selected for output. This provides meaningful evidence that the method can jointly improve accuracy and efficiency.

2. The investigation of depth-dependent detection failures is informative and supported by several complementary experiments. Rather than treating early-exit improvements purely as an empirical observation, the paper investigates why deeper processing can be detrimental. The study compares three differently pretrained backbones, examines the distribution of optimal exit layers, and introduces both held-out and near-duplicate context pollution while maintaining a fixed context size. The planted-sibling experiments and gradient-based sensitivity analysis provide suggestive evidence that the influence of nearby context samples can increase with depth, offering a plausible explanation for why intermediate predictions may outperform final-layer predictions.

The evaluation also includes useful controls and ablations. The comparison with fixed-depth and learned early-exit baselines, matched-depth component ablations, synthetic training-data ablations, and analysis of the depth penalty help clarify the contributions of the routing design. The reported differences between backbones are especially informative: they suggest that the benefits of post-hoc early exit depend not only on the available oracle improvement but also on the backbone's pretraining regime and depth-dependent behavior. These observations are valuable beyond the particular router proposed.

Critical weaknesses

1. The claim that context pollution amplifies the benefits of learned adaptive exiting requires important qualifications across backbones, particularly for contamination-aware models. The paper convincingly shows that oracle early-exit opportunities generally increase under context pollution. However, the learned router does not consistently capture these additional opportunities. Table 12 shows that rOUT underperforms full-depth TACTIC in five of the nine held-out pollution settings, with smaller relative improvements than in the corresponding clean setting in all nine. At the 1:1 pollution ratio, its relative AUROC gains are -0.3% on OddBench, -0.6% on OvRBench, and -1.3% on ADBench. On ADBench, rOUT underperforms full depth at all three pollution levels, despite the substantial oracle headroom that remains.

This distinction is important because the motivating analysis emphasizes larger early-exit opportunities under pollution, whereas the practical contribution depends on whether the learned router can identify and exploit them. For example, on TACTIC under 1:1 held-out pollution, the Oracle improves over Full depth from 0.636 to 0.754 on OddBench, from 0.626 to 0.720 on OvRBench, and from 0.777 to 0.831 on ADBench, yet rOUT slightly underperforms Full depth in all three cases. TACTIC uses a polluted-context pretrained checkpoint in the main evaluation, unlike OutFormer, which is pretrained with clean contexts. This suggests that backbone pretraining and routing generalization may interact in ways not fully captured by the proposed explanation.

Furthermore, the inference-time benefits are uneven. Under clean contexts, TACTIC achieves only 1.03×–1.22× wall-clock speedups and computes approximately 10 of its 12 layers. Tables 12 and 13 do not report corresponding computed depths or inference times under pollution, making it difficult to determine whether the weaker or negative accuracy gains are offset by meaningful computational savings. Since pollution robustness is central to the paper's motivation, a more explicit distinction between oracle opportunities and realized routing gains, together with pollution-specific quality–cost measurements, would strengthen the empirical conclusions.

2. The mechanistic interpretation and attribution of the router's predictive information remain insufficiently isolated. The planted-sibling analysis provides useful evidence of depth-dependent sensitivity, but its interpretation requires care. Appendix B.1 explains that the targeted-outlier "AUROC" in Figure 5 compares the outlier scores under clean or polluted contexts against a fixed set of inlier scores obtained under clean context. This diagnostic isolates changes in targeted outlier scores, but it differs from ordinary within-condition AUROC. In particular, it may respond to changes in the relative score calibration between context conditions that would not necessarily produce equivalent changes in AUROC when both inlier and outlier scores are evaluated under the same condition. Similarly, the gradient-norm influence statistic measures local sensitivity to a planted sibling, but does not by itself establish the direction of that influence or show that it causally produces the observed masking. The strong across-layer correlations are suggestive, but do not fully establish the proposed mechanism.

The cross-backbone interpretation is further complicated by checkpoint differences. Figure 15 uses the clean-context TACTIC checkpoint, whereas the main TACTIC experiments use the polluted-context checkpoint (Appendix C.1). Consequently, the mechanistic evidence for TACTIC is not directly matched to the backbone used to assess the router's performance and pollution robustness.

A related attribution issue concerns the router's inputs. Section 4 and Appendix B.2 show that rOUT receives not only compressed backbone representations but also raw input features, intermediate outlier scores, and 29 handcrafted geometric features, including nearest-neighbor distances, local outlier factors, and reverse-neighbor counts. Table 5 finds that removing the additional features reduces AUROC by 1.0%–1.7% at matched computation depth, which is substantial relative to the total improvements over full-depth inference. However, this ablation removes the feature groups together, making it unclear how much of the routing improvement comes from backbone representations and intermediate scores versus external geometric statistics. This does not invalidate the practical performance of rOUT, but limits the conclusions that can be drawn about the information responsible for its routing decisions. More targeted feature-group ablations and matched mechanistic controls would strengthen the explanation.

Questions for authors

1. Clarification of routing behavior and efficiency under context pollution. Table 12 shows that the Oracle retains substantial early-exit gains under held-out pollution on TACTIC, while rOUT underperforms Full depth in five of the nine polluted settings. What explains this gap between oracle opportunities and realized routing gains? Does the router face a distinct generalization challenge when applied to a backbone already pretrained with polluted contexts, or are the results primarily attributable to TACTIC's irregular layer-wise performance? Please provide the mean computed depths or measured inference times for Tables 12 and 13, particularly for the negative-gain TACTIC settings, so that the accuracy–efficiency trade-off under pollution can be assessed directly.

2. Clarification and matched validation of the proposed masking mechanism. Does the depth-dependent masking pattern in Figure 5 persist when both inlier and outlier scores are evaluated under the same context condition, rather than comparing polluted-context outlier scores against clean-context inlier scores? Could the authors distinguish changes in relative score calibration from changes in within-condition discrimination? Additionally, why does Figure 15 use the clean-context TACTIC checkpoint, and does the reported influence–performance relationship persist for the polluted-context checkpoint used in the main experiments? For held-out pollution, evaluating the zero-pollution baseline on the same retained query subset used in the polluted conditions would also help separate the effects of context pollution from differences in the evaluated queries.

3. Disentangling the contribution of additional router features. Table 5 removes all additional features together, although these include three distinct groups: raw input features, layer-wise backbone outlier scores, and 29 handcrafted geometric features. Could the authors report separate ablations for these groups at matched computation depth? In particular, how well does rOUT perform using only backbone representations and layer-wise outlier scores, without the external geometric features? This would help establish whether the router primarily exploits depth-dependent backbone information or also relies substantially on conventional neighborhood-based anomaly statistics. A brief clarification of whether the 64-dimensional PCA representations undergo the same quantization at inference as during training, and how well the training-fitted projection transfers to unseen real-world datasets, would also be useful.

4 comments
AuthorsOpenReview ↗

Response to Reviewer kvTV

We sincerely thank the reviewer for the thorough and constructive review, and for recognizing both the practical value of rOUT, which improves detection quality and inference efficiency without modifying the pretrained backbones, and our investigation of depth-dependent detection failures. The reviewer's suggestions helped us sharpen the mechanism and attribution analyses, and we address each of them below. Please see the revisions in blue text in the revised copy of our submission, which is now uploaded on OpenReview.

W1 / Q1. Oracle opportunity versus realized gains under pollution, and the computed depth.

Oracle opportunity versus realized gain on TACTIC. We believe the gap comes mainly from the pollution robustness of TACTIC-N rather than from a separate generalization challenge for the router. TACTIC-N (the backbone in our main experiments) is pretrained with each context contaminated by unlabeled anomalies of the same dataset at 5–30% of the inliers (Marszałek et al., 2026), which closely resembles held-out pollution. It is therefore much less sensitive to context pollution (also see W2 / Q2 below), so its deep layers lose little AUROC and early exit has less to gain. Its AUROC also rises with depth on average, so the Best fixed layer is the last layer in every polluted setting and the remaining gains must be captured dataset by dataset. rOUT adapts to this by staying deep at about 10 of 12 layers. It still improves over Full depth in all nine near-duplicate settings and at the 1:4 and 1:2 held-out ratios on OddBench and OvRBench.

Computed depth under pollution. We thank the reviewer for pointing out this missing measurement and have added rOUT's mean computed layers to Tables 4, 12 and 13. For illustration, we show TACTIC under held-out pollution, the setting the reviewer highlights:

Table R1: On TACTIC under held-out pollution, rOUT computes about 10 of 12 layers and stays close to full-depth AUROC. Excerpt of Table 12 (AUROC; Full depth computes all 12 layers)

TACTICClean1:41:21:1
OddBenchFull depth.736.703.678.636
rOUT.747.712.686.634
rOUT's computed layers10.210.19.99.9
OvRBenchFull depth.715.688.663.626
rOUT.729.698.670.622
rOUT's computed layers10.010.110.110.0
ADBenchFull depth.832.825.807.777
rOUT.849.813.802.767
rOUT's computed layers10.19.99.89.5

In all five settings where rOUT falls slightly below Full-depth TACTIC, the difference is not significant (the 95% CI of the mean difference includes zero) and rOUT computes 17–21% fewer layers. On OutFormer and ICLAD, rOUT computes about half of the layers under pollution, while improving AUROC in all but one of the 36 polluted settings.

Revision. We added the mean computed layers to Tables 4, 12 and 13, and a discussion of oracle opportunity and realized gain per backbone to Appendix C.

W2.1 / Q2. Validation of the masking mechanism.

We thank the reviewer for these constructive suggestions on strengthening the evidence for the masking mechanism.

Within-condition AUROC. As the reviewer notes, Figure 5 compares targeted outliers under polluted context against query inliers under clean context, which may also reflect a calibration shift. To address this, we recomputed AUROC with both groups scored under the same context, and redrew Figures 5, 14 and 15. The depth-dependent masking persists: for OutFormer, the final-layer AUROC loss is 0.119 (vs. 0.129 originally), and its correlation with sibling influence across layers remains 0.99. The same pattern holds for ICLAD and TACTIC, with the random-inlier control at zero for all three backbones.

Direction of the sibling's influence. To examine whether siblings specifically mask their corresponding query outliers, we use a controlled intervention: for each target, we replace one context row with its sibling while keeping everything else fixed, and measure the change in its ranking relative to query inliers. We find that 88% of targets drop in rank after planting, whereas replacing the same row with a random inlier produces no comparable effect. Moreover, closer siblings cause stronger masking, and the AUROC loss closely tracks increasing sibling influence across layers. Together, these observations support the proposed masking mechanism.

The polluted-context TACTIC checkpoint. We used TACTIC-C for Figure 15 because its inlier-only pretraining helps isolate the effect of a planted sibling. Following the reviewer's suggestion, we repeated the analysis on TACTIC-N, which is pretrained with polluted contexts. The same pattern persists, but at a smaller magnitude: sibling influence grows with depth, targeted AUROC decreases, and the random-inlier control remains at zero. The final-layer loss is 0.020 vs. 0.091 for TACTIC-C, consistent with greater robustness to pollution (see W1 / Q1).

Clean baseline on the retained queries. Under held-out pollution, the outliers moved into the context leave the query set, so the polluted conditions use a subset of the original queries. Following the reviewer's suggestion, we evaluated the clean baseline on this retained subset. Its AUROC differs from that on the full query set by at most 0.004 across all backbones and benchmarks, far below the effect of pollution. The observed changes therefore come from the polluted context rather than from the change in the evaluated queries.

Revision. We redrew Figures 5, 14 and 15 with within-condition AUROC. Appendix B.1 now describes this protocol and the rank change of targeted outliers and covers both TACTIC checkpoints, and Appendix C adds the retained-query check.

W2.2 / Q3. Contribution of the additional router features, and the PCA encoding.

Feature-group ablations. We thank the reviewer for this suggestion, and would like to clarify that the "w/o additional feat." row of Table 5 removes only the 29 geometric features. In the revision we relabel this row "w/o geometric feat." and add the two remaining feature ablations, removing either the raw input features or the layer-wise outlier scores:

Table R2: Geometric features and layer-wise scores both contribute; raw input features add little. Lower block of the revised Table 5 (first row: rOUT's AUROC; other rows: relative change (%) from rOUT at matched depth; one-sided Wilcoxon, ∗p<.05^{\ast}p<.05, ∗∗p<.01^{\ast\ast}p<.01, ∗∗∗p<.001^{\ast\ast\ast}p<.001)

OutFormerOddOvRAD
rOUT (full).761.730.860
w/o geometric feat.−1.7$^{\ast\ast\ast}$−1.0$^{\ast\ast\ast}$−1.5$^{\ast}$
w/o layer-wise scores−0.7$^{\ast\ast\ast}$−0.6$^{\ast\ast}$−0.8
w/o raw input feat.−0.2+0.2+0.4
Full depth−2.7$^{\ast\ast\ast}$−2.9$^{\ast\ast\ast}$−0.9$^{\ast}$

This separation makes the source of the routing signal clearer. The routing signal comes primarily from the backbone: without the geometric features, rOUT still outperforms Full depth on OddBench and OvRBench. Layer-wise outlier scores and geometric features both contribute on top of it, and the geometric features add neighborhood information the representations do not encode. Raw input features add little, since the backbone representations are already computed from them.

PCA encoding. The 64-dimensional projection and its int8 quantization are fit once on synthetic training data and applied unchanged at inference. To assess its transfer to real data, we compare it with a PCA fit on each real benchmark itself. On the backbone representations of real datasets, the synthetic-fit projection keeps about 90% as much variance as the benchmark-fit PCA, averaged over the three benchmarks and consistent across early, middle and final layers. The encoding learned from synthetic data thus transfers well to unseen real-world datasets.

Revision. We add Table R2 as part of Table 5, which now reports all three feature ablations, Section 5 describes them, and Appendix B.2 states that the same encoding is used at inference and reports its transfer.

If the computed-depth results, the within-condition analysis and the feature ablations address the reviewer's concerns, we hope the reviewer might consider revisiting the score. We also welcome any further questions during the remaining discussion period.

We sincerely thank the reviewer for taking the time to read our responses and for updating the assessment of the paper's soundness. We are glad that our clarifications addressed the reviewer's concerns. The reviewer's constructive feedback has also helped us strengthen the mechanism and attribution analyses in the revised paper.

Reviewer_22hbOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work investigates adaptive early exiting for pretrained tabular outlier detection foundation models, and demonstrates that reducing inference depth can improve not only computational efficiency, but also detection accuracy. The authors identify context pollution, where unlabeled outliers appear among in-context samples, as an important factor contributing to performance degradation at deeper layers.

Motivated by these observations, the authors propose rOUT: a post-hoc, dataset-adaptive routing mechanism that jointly learns when to stop computation and which previously computed layer to use for prediction. Experimentally, rOUT matches or improves full-depth performance in the majority of clean-context configurations while reducing the number of computed layers.

Critical strengths
  • The paper does a good job of exploring the problem empirically.
  • There’s no doubt the topic studied is practically useful. The paper proposes a practical post-hoc routing framework.
Critical weaknesses
  • Somewhat unclear what is the motivation behind predicting outliers in tabular data while taking into account pollution. This seems like an overly specific task. Also, the evaluation protocol leaves uncertainty about performance in naturally contaminated and fully unlabeled settings.
  • I found the paper to be somewhat hard to follow. Fig 1 is missing confidence bars, seeds, statistical significance, etc. Also the figures and their captions are unclear and dense. Importance and necessity of Figure 3 is unclear.
  • How is the first part of the paper, talking about pollution and early exists, related to the second part of the paper, proposing a new early exist method? It almost feels like two papers combined into one.
Questions for authors
  • What’s the definition of pollution/outlier/siblings/etc.? paper would be clearer if these had proper mathematical definitions.
  • Isn’t the result of figure 4 expected? In what sense is it surprising or explains the paradox? why is it important to show?
  • How does figure 5 resolve the presented paradox?
5 comments
AuthorsOpenReview ↗

Response to Reviewer 22hb

We sincerely thank the reviewer for the time and the constructive comments, and for recognizing our thorough empirical exploration of the problem and the practical value of both the topic and our post-hoc routing framework. We address each point below. Please see the revisions in blue text in the revised copy of our submission, which is now uploaded on OpenReview.

W1.1. Motivation for accounting for context pollution.

Let us clarify the motivation: Foundation models for tabular outlier detection take two (disjoint) sets as input: context set (akin to train set) and query set (akin to test set), where query points cross-attend to context points in getting labeled by the model.

There are two detection settings: 1) clean (or inlier-only) context and unlabeled query, or 2) fully unlabeled: where now both context and query sets contain unknown/unlabeled outliers. Note that neither uses any labeled outliers.

Both settings are valid and used heavily in practice. In fact, the fully unlabeled setting is more prevalent: when it is not easy to gather clean train data, the fully unlabeled input data is split into two, as context and query, to be fed into the foundation model, which yields a “polluted” context; i.e. context likely contains (unknown/unlabeled) outliers in it.

The important things to note here are: 1) The method does not depend on pollution: rOUT's main results (Table 2) use the benchmarks' designated inlier-only context splits. 2) Pollution is common/default in practice: Without labels to exclude outliers, the context is potentially contaminated. This is the standard evaluation setting for classical unsupervised detectors. 3) Nominally clean splits still carry hidden pollution: Even when a benchmark designates an inlier-only context split, its inliers are clean only by label. Label noise and outlier-like inliers are pervasive in real data (see Fig 2; and the newly added Table R3 on pg. 5).

Revision. We revised our draft as follows:

  • Opening of the introduction: State that real-world contexts are unlabeled and therefore naturally and implicitly polluted, and tie this directly to the early-exit question.
  • In Table 2 and its caption: Describe the "clean" setting as the benchmarks' designated inlier-only context split, and note that hidden pollution may remain.

W1.2. Performance in naturally contaminated and fully unlabeled settings.

Fully unlabeled. rOUT is fully unlabeled at inference in every experiment we report. Context rows carry no labels, and no label of the target dataset is used to choose the exit layer, set the stopping threshold, or tune any hyperparameter. Query labels appear only during the router's pretraining on synthetic datasets, where they serve as privileged information.

Controlled pollution ratios, designed for comparability. We first establish the early-exit effect on benchmarks whose designated context split is inlier-only, where only hidden pollution remains (Table 2), and then raise the pollution ratio step by step. In held-out pollution, the context receives each dataset's own real outliers, held out from the query set, at context-to-query outlier ratios of 1:4, 1:2, and 1:1, while the datasets and the context size stay fixed. This design makes all ratios and backbones directly comparable and isolates pollution as the variable, which is what attributes the growing early-exit gains to pollution itself (Figure 4, Table 4). At 1:1, the context pollution rate mimics the dataset's natural outlier rate (median 0.8–1.2$\times$, Table 4), and rOUT improves relative AUROC over full-depth OutFormer by 5.4–10.1% across the three real-world benchmarks, up from 0.9–3.0% on inlier-only context (Table 4).

Revision. We revised our draft in Sections 2 and 3 to explain the controlled-ratio design and state that at the 1:1 ratio the context pollution rate is close to the natural outlier rate.

W2. Confidence bars, seeds and significance in Figure 1, clarity of the captions, and the role of Figure 3.

Fig 1: The oracle markers in Figure 1 (★ for clean and capital P for polluted context) show 95% confidence intervals across datasets. For significance tests, see Table 1 which reports the oracle gains under clean and polluted context, with p-values for their increase under pollution, for 9 settings: 3 model backbones and 3 benchmarks; in 8 of 9 settings, p ≤ 10⁻⁵, so the increase is significant. No seeds are involved as the backbones are frozen and inference is deterministic.

Figure captions: R-22hb is right that Fig 1, and several other figures, are conveying a lot of information. We revised the captions of Figures 1, 3, 4 and 5 to clearly specify a) what the figure displays, and b) what the key take-aways are. Please see the changes in blue in the revised paper, and let us know if further clarification is needed.

Why Figure 3 matters. Figure 3 is the bridge from the analysis to the method. Its bottom panel shows that the best (oracle) exit layer varies widely across datasets on every real-world benchmark, and its top panel shows that the gains come from datasets whose best layer is shallow or intermediate. No single fixed layer captures these gains. Interestingly, 1) AUROC gains can be as large as 0.8 in AUROC on a single dataset; 2) Gains from shallow/mid layers tend to be larger than late layers on all three real-world benchmarks for all backbones; and 3) About 49% of datasets have shallow- or mid-layer oracle, where the gains tend to be larger. This demonstrates the promise and the impact that strategic early exit can bring, and sets the stage for our rOUT design.

Therefore, these analyses and observations in the first part of the paper exactly set the motivation for our work in the second part of the paper: designing and training a router to effectively predict the oracle layer, to materialize the observed gains in practice.

Revision. We state this link explicitly in Section 2 and at the start of Section 4.

W3. How the pollution analysis relates to the proposed early-exit method.

The two main parts of our paper address successive questions: why exit early (Sections 2-3), and how to choose the exit without labels (Section 4). As Reviewer DADc notes, “The paper is well organized: motivating observation → mechanistic explanation → method → extensive evaluation.”

More specifically, Sections 2–3 establish that intermediate layers can outperform full depth, particularly under context pollution, and investigate a contributing mechanism. The dataset-specific oracle gains then motivate Section 4: a practical router that selects an exit layer without access to ground-truth labels. Section 5 demonstrates that rOUT effectively realizes those gains (Table 2 and others).

Thus, the analysis establishes the opportunity and the method makes it actionable. We truly believe these two parts are fully complementary and the paper presents a unified piece of research.

Q1. Mathematical definitions of pollution, outlier, sibling, etc.

We thank the reviewer for this constructive suggestion: precise definitions indeed can help make the analyses easier to follow.

Revision. We stated the label notation and added formal definitions: the oracle layer and context pollution in Section 2 (Definitions 1–2), and the sibling and the pollution ratio in the Appendix (Definitions 3–4); in brief:

  • Outliers and inliers. A dataset is split into a context set Dctx\mathcal{D}_{\mathrm{ctx}} and a query set Dqry\mathcal{D}_{\mathrm{qry}}; each point x\mathbf{x} has a label y(x)∈{0,1}y(\mathbf{x})\in\lbrace 0,1\rbrace, with 1 for outliers. Labels are never observed at inference.
  • Oracle layer (Definition 1). l⋆(D)=arg⁡max⁡lAUROCl(D)l^\star(\mathcal{D})=\arg\max_l \mathrm{AUROC}_l(\mathcal{D}), the exit layer with the highest AUROC through the frozen final-layer head.
  • Context pollution (Definition 2). The context Dctx\mathcal{D}_{\mathrm{ctx}} is polluted if its pollution rate ε=∣{x∈Dctx:y(x)=1}∣/∣Dctx∣\varepsilon=|\lbrace \mathbf{x}\in\mathcal{D}_{\mathrm{ctx}}:y(\mathbf{x})=1\rbrace|/|\mathcal{D}_{\mathrm{ctx}}| is positive. Even an inlier-only split with ε=0\varepsilon=0 by label may hold mislabeled outliers, i.e., hidden pollution.
  • Sibling (Definition 3). A near-copy of a query outlier xq\mathbf{x}_q: xs=xq+ρrqu\mathbf{x}_s=\mathbf{x}_q+\rho r_q\mathbf{u}, where rqr_q is the distance from xq\mathbf{x}_q to its nearest context point and u\mathbf{u} a random unit direction; ρ=0.5\rho=0.5 in Figure 5 and ρ∼U(0.1,1)\rho\sim U(0.1,1) in near-duplicate pollution.
  • Pollution ratio (Definition 4). Polluting points replace context inliers, so ∣Dctx∣|\mathcal{D}_{\mathrm{ctx}}| is fixed, and the ratio is r=npol/∣{x∈Dqry:y(x)=1}∣∈{1:4,1:2,1:1}r=n_{\mathrm{pol}}/|\lbrace \mathbf{x}\in\mathcal{D}_{\mathrm{qry}}:y(\mathbf{x})=1\rbrace|\in\lbrace 1{:}4, 1{:}2, 1{:}1\rbrace, where npoln_{\mathrm{pol}} is the number of polluting context points (held-out outliers or siblings).

Q2. Whether the result of Figure 4 is expected, how it relates to the paradox, and why it is important.

The expected result of Figure 4 is that context pollution reduces detection performance. The surprising result is that intermediate layers, evaluated through a head trained only on the final layer, can outperform full-depth inference, and that this advantage grows under pollution.

Figure 4 shows this for OutFormer, and Figures 9–13 show the same pattern for the other backbones and for near-duplicate pollution: 1) Pollution hurts deep layers more than shallow ones. 2) A fixed intermediate exit retains much of the full-depth performance at lower cost. 3) The best exit per dataset (the oracle) improves over full depth in both clean and polluted settings, and the more pollution, the more early exit pays off. These are surprising findings that were not reported in the literature before. They also unearth early-exit as an opportunity to not only cut back inference cost but also improve detection.

These results establish that full depth can be suboptimal and motivate dataset-specific exit selection through rOUT. Because potentially polluted contexts are standard in unsupervised OD, this opportunity is directly relevant to practical deployment and is exactly what rOUT delivers.

Revision. We revised the caption and opening sentence of the Figure 4 discussion to highlight the unexpected finding and its practical implications.

Q3. How Figure 5 resolves the paradox.

The paradox is that, surprisingly, intermediate layers can detect outliers better than the final layer, even through a prediction head trained only on the final layer. Figure 5 investigates one possible mechanism: whether context points resembling query outliers, i.e. their “siblings” (Definition 3), mask those outliers, particularly in deeper layers. We isolate and show the impact of this effect by manually planting siblings into otherwise clean contexts.

Figure 5 (left) shows that planting siblings consistently reduces AUROC, with larger reductions in deeper layers on average, consistent with our observations on real datasets. The middle panel shows that siblings’ influence on the corresponding query predictions increases with depth. The right panel shows a strong correlation between this influence and the AUROC reduction. In contrast, planting a copy of a random inlier produces no comparable effect.

Together, these results support context pollution as a contributing mechanism: outlier-like context points can mask similar query outliers, with deeper layers more susceptible to this effect.

Revision. We revised the Figure 5 caption to lead with this takeaway, followed by guidance on reading and interpreting each panel.

If these clarifications address the reviewer's concerns on the motivation, the link between analysis and method, and the definitions, we hope the reviewer might consider revisiting the score. We also welcome any further questions during the remaining discussion period.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Reviewer Discussion and Revisions

We sincerely thank all three reviewers and the AC for their time and constructive discussion. We are grateful for the reviewers' positive assessments and thoughtful consideration of our responses and revisions.

Strengths highlighted by the reviewers.

  • A counterintuitive finding: Early exit can improve detection accuracy, not just efficiency, with greater oracle gains under context pollution. This is "the kind of finding that changes how practitioners think about depth in in-context OD models" (Reviewer DADc).
  • A practical post-hoc router: rOUT improves detection quality and inference efficiency without modifying pretrained backbones (Reviewers 22hb and kvTV). Importantly, "the empirical gains are not limited to an oracle analysis" (Reviewer kvTV).
  • Clear organization: "The paper is well organized: motivating observation → mechanistic explanation → method → extensive evaluation" (Reviewer DADc).
  • Extensive evaluation: A thorough empirical study across three backbones and four benchmarks (Reviewers 22hb and DADc).

Key additions during the discussion.

In response to the reviewers, we conducted additional experiments and analyses, reported in Tables R1–R6 and incorporated into the revised manuscript:

  • Motivation and presentation (Reviewer 22hb): We added formal definitions, clarified the relevance of context pollution to real-world outlier detection, and made the connection between the analysis and the proposed method explicit.
  • Efficiency and mechanism (Reviewer kvTV): Mean computed layers under pollution help clarify the accuracy–efficiency trade-off (Table R1). Separate feature ablations show that geometric features and layer-wise scores both contribute, whereas raw input features have little effect (Table R2). Within-condition AUROC, targeted-outlier rank changes, and the polluted-context TACTIC checkpoint provide further evidence for the masking mechanism (Figures 5, 14, and 15).
  • Real data and baselines (Reviewer DADc): Fine-tuning with real labels yields no significant improvement on held-out benchmarks in our experiments (Table R3), while routers evaluated outside their training distribution perform slightly below rOUT but still exit well before full depth (Table R4). Removing predicted hidden pollution from real contexts reduces the oracle gain and shifts the optimal exit layer deeper (Table R5). rOUT outperforms newly-added baselines using the same frozen backbone and prediction head, highlighting the benefit of its routing decisions (Table R6).

All revisions are highlighted in blue in the uploaded manuscript. We thank the reviewers again for their constructive feedback, which has helped us strengthen the paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.