Ranked Activation Shift for Post-Hoc Out-of-Distribution Detection

Gianluca Guglielmo; Marc Masana

arxiv: 2604.08572 · v2 · pith:UPVS5H6Cnew · submitted 2026-03-27 · 💻 cs.LG · cs.CV

Ranked Activation Shift for Post-Hoc Out-of-Distribution Detection

Gianluca Guglielmo , Marc Masana This is my paper

Pith reviewed 2026-05-14 23:50 UTC · model grok-4.3

classification 💻 cs.LG cs.CV

keywords out-of-distribution detectionpost-hoc methodsactivation editingneural network reliabilitymachine learning safetypenultimate layer

0 comments

The pith

Replacing sorted penultimate activations with a fixed in-distribution profile produces consistent out-of-distribution detection without tuning or accuracy loss.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Existing post-hoc methods for detecting out-of-distribution inputs by editing layer activations show unstable results because activation distributions vary across datasets and models. The paper traces one clear failure to scaling operations when activations lack rectification and introduces a method that ranks activation magnitudes then substitutes them with values drawn from a single fixed reference profile computed on in-distribution data. This produces reliable discrimination while leaving the original classifier untouched, so in-distribution accuracy stays exactly the same. The approach needs no hyperparameters and imposes no restrictions on the penultimate activation function. Readers should care because dependable detection of unfamiliar inputs is a basic requirement for deploying neural networks safely in open environments.

Core claim

The paper claims that Ranked Activation Shift achieves strong and consistent out-of-distribution detection by replacing the sorted magnitudes of penultimate-layer activations with a fixed reference profile taken from in-distribution samples. The substitution is performed once, requires no tuning, works for any activation function, and leaves the network's in-distribution predictions unchanged by construction. Separate analysis shows that the resulting inhibition and excitation of activations each contribute independently to improved separation between in-distribution and out-of-distribution inputs.

What carries the argument

Ranked Activation Shift, which sorts the magnitudes of penultimate activations and replaces them with corresponding entries from a fixed in-distribution reference profile to produce a shifted activation vector used for OOD scoring.

If this is right

The method yields stable out-of-distribution performance across multiple datasets and network architectures without per-dataset adjustments.
No hyperparameter search or assumptions about the penultimate activation function are required.
Both the inhibitory and excitatory effects of the shift contribute separately to better discrimination.
In-distribution classification accuracy remains exactly the same as the unmodified model.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Similar ranking-based replacement steps could be tested in other post-hoc correction techniques that currently rely on raw activation statistics.
The observation that activation-distribution mismatch drives instability suggests that training procedures could be modified to encourage more uniform activation profiles.
Reference profiles might be made slightly adaptive per class or per domain to handle structured distribution shifts while retaining the hyperparameter-free property.

Load-bearing premise

A single fixed reference profile computed from in-distribution activations will generalize to produce reliable shifts that discriminate out-of-distribution inputs across different datasets and models.

What would settle it

A controlled experiment on a new model or dataset in which the method's AUROC for out-of-distribution detection falls below that of simple baselines when the penultimate activation distribution deviates strongly from the training reference profile.

Figures

Figures reproduced from arXiv: 2604.08572 by Gianluca Guglielmo, Marc Masana.

**Figure 2.** Figure 2: Overview of RAS. At setup time, RAS computes a reference vector µ by sorting each ID train sample by intensity and then averaging across all. At inference time, each sample’s activations are shifted so that their ranked activation intensities match µ. Subsequently, they are passed to the classifier and the chosen OoD scorer is used. Algorithm 1 RAS: Ranked Activation Shift Require: ID training set Din, fea… view at source ↗

**Figure 3.** Figure 3: RAS (shifted) vs ℓ2- normalization, shown as residuals with respect to the average ranked ID activations [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗

**Figure 4.** Figure 4: Mean AUROC gain of RAS (solid line) when applied to each ResNet layer, [PITH_FULL_IMAGE:figures/full_fig_p011_4.png] view at source ↗

**Figure 5.** Figure 5: Comparison of activation profiles for ID (dashed line) and OoD datasets [PITH_FULL_IMAGE:figures/full_fig_p014_5.png] view at source ↗

read the original abstract

State-of-the-art post-hoc out-of-distribution detection methods rely on intermediate layer activation editing. However, they exhibit inconsistent performance across datasets and models. We show that this instability is driven by differences in the activation distributions, and identify a failure mode of scaling-based methods that arises when penultimate layer activations are not rectified. Motivated by this analysis, we propose RAS, a hyperparameter-free post-hoc method that replaces sorted activation magnitudes with a fixed in-distribution reference profile. Our simple plug-and-play method shows strong and consistent performance across datasets and architectures without assumptions on the penultimate layer activation function, and without requiring any hyperparameter tuning, while empirically preserving in-distribution classification accuracy. We further analyze what drives the improvement, showing that both inhibiting and exciting activation shifts independently contribute to better out-of-distribution discrimination.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper spots a concrete failure mode in scaling-based OOD detectors when activations are unrectified and fixes it with a fixed ID reference profile for sorted magnitudes, giving a simple no-tune method that holds ID accuracy by construction.

read the letter

The main thing here is that they isolate why scaling on penultimate activations breaks down without rectification and replace the sorted values with one fixed reference profile drawn from ID data. That change is the actual contribution, and it stays inside the activation-editing family rather than claiming a new paradigm. They also separate the effect into inhibiting and exciting shifts and show both help discrimination. The method needs no hyperparameters and keeps in-distribution accuracy untouched, which is a clean practical property. They test across datasets and architectures without assuming a particular activation function, and the abstract reports consistent gains. That combination of simplicity and reported stability is what makes the work worth looking at for applied OOD work. The soft spot is the reliance on a single fixed profile. If activation statistics shift between the set used to build the reference and a later ID set or different model, the ranked replacement can lose its edge even while nominal ID accuracy stays the same. The stress-test note on varying means, tails, or rectification status is the right thing to check in the experiments. Without the full tables and ablations it is hard to judge how sensitive the profile choice is or how broadly the consistency claim holds. The paper does not appear circular or to invent new entities, and the math stays straightforward. This is for practitioners who already run post-hoc OOD checks and want something lightweight to drop in without tuning or retraining. It will not interest readers looking for theoretical depth or large paradigm shifts. I would send it to peer review. The motivation is direct, the fix is testable, and the practical angle is clear enough that referees can evaluate the numbers and the generalization limits.

Referee Report

2 major / 0 minor

Summary. The paper analyzes failure modes in state-of-the-art post-hoc OOD detection methods that rely on intermediate-layer activation editing, attributing instability to differences in activation distributions and identifying a specific failure mode in scaling-based methods when penultimate-layer activations are not rectified. It proposes Ranked Activation Shift (RAS), a hyperparameter-free method that replaces sorted activation magnitudes with a fixed in-distribution reference profile. The central claims are strong and consistent performance across datasets and architectures without assumptions on the activation function or hyperparameter tuning, preservation of ID classification accuracy by construction, and that both inhibiting and exciting shifts contribute to improved OOD discrimination.

Significance. If the empirical claims hold, the work provides a simple, plug-and-play post-hoc OOD detector that directly addresses documented inconsistencies in prior activation-editing methods. The parameter-free nature and lack of activation-function assumptions would make it immediately usable across models, while the analysis of shift directions offers mechanistic insight into what drives effective discrimination.

major comments (2)

[Abstract] Abstract and method description: the central claim of 'strong and consistent performance across datasets and architectures without ... hyperparameter tuning' rests on the fixed reference profile generalizing to new ID data and architectures. If penultimate-layer statistics (means, tails, rectification status) differ between the profile-construction set and a later ID test set, the ranked shift can misalign and weaken OOD discrimination, even while preserving nominal ID accuracy. No explicit test of this generalization risk is described.
[Abstract] The manuscript provides no access to full experimental results, tables, or implementation details, so the quantitative support for the performance claims cannot be verified (soundness rated 4.0).

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their detailed and constructive feedback on our manuscript. We address each major comment below and outline the revisions we will make to strengthen the paper.

read point-by-point responses

Referee: [Abstract] Abstract and method description: the central claim of 'strong and consistent performance across datasets and architectures without ... hyperparameter tuning' rests on the fixed reference profile generalizing to new ID data and architectures. If penultimate-layer statistics (means, tails, rectification status) differ between the profile-construction set and a later ID test set, the ranked shift can misalign and weaken OOD discrimination, even while preserving nominal ID accuracy. No explicit test of this generalization risk is described.

Authors: We appreciate the referee's point regarding potential misalignment if the reference profile is applied to ID data with differing activation statistics. In the current experiments, the reference profile is always constructed from the training split of the same ID dataset used for evaluation, ensuring distributional alignment by design. To directly address the generalization concern, we will add a new subsection in the revised manuscript with experiments that vary the reference set construction (e.g., using random subsets or cross-validation folds of the ID training data) and report the resulting OOD detection performance. These results will quantify sensitivity to within-ID variations and support the claim of robustness without hyperparameter tuning. revision: partial
Referee: [Abstract] The manuscript provides no access to full experimental results, tables, or implementation details, so the quantitative support for the performance claims cannot be verified (soundness rated 4.0).

Authors: We agree that the current manuscript version lacks complete tables and implementation details in the main body, which limits immediate verification. In the revised submission, we will include expanded experimental tables reporting all metrics across every dataset-architecture pair, along with a dedicated reproducibility section that provides a link to the full open-source code repository containing the implementation, data preprocessing scripts, and exact experimental configurations. revision: yes

Circularity Check

0 steps flagged

Derivation chain is self-contained without reduction to inputs

full rationale

The proposed Ranked Activation Shift method defines a fixed in-distribution reference profile from training data and uses it to adjust activation ranks for OOD scoring. This is a direct algorithmic construction, not a fitted model whose outputs are then presented as predictions. The paper's analysis of activation distributions and failure modes provides independent motivation, and performance is validated empirically rather than following tautologically from the definition. No load-bearing self-citations or ansatzes smuggled in are present in the provided text.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The approach rests on the premise that a fixed reference profile extracted from in-distribution activations can serve as a universal comparator, with no free parameters introduced beyond that profile.

axioms (1)

domain assumption Penultimate layer activations can be sorted by magnitude and replaced by a fixed reference without changing in-distribution classification accuracy
Stated as holding by construction in the abstract

pith-pipeline@v0.9.0 · 5431 in / 1205 out tokens · 42418 ms · 2026-05-14T23:50:11.733738+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel (J-cost uniqueness) unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

RAS … replaces sorted activation magnitudes with a fixed in-distribution reference profile … ¯a_π(j) ← μ_j … μ = 1/N Σ r(a_i)
IndisputableMonolith/Foundation/AlphaCoordinateFixation.lean alpha_pin_under_high_calibration unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

hyperparameter-free … without assumptions on the penultimate layer activation function
IndisputableMonolith/Foundation/BranchSelection.lean branch_selection (coupling-combiner forces bilinear branch) unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

both inhibiting and exciting activation shifts independently contribute

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.