acceptodds
Under review as a conference paper at ICLR 2027

Energy Ring: Stabilizing Feature Energy for Reliable Test-Time Discovery

Abstract

Test-Time Discovery (TTD) aims to recognize known classes and detect novel ones while the model is already deployed, operating directly on a continual unlabeled test-time stream. Existing TTD methods are overly conservative, making known and unknown features collapse together and difficult to separate during test-time adaptation. In this paper, we revisit TTD from an energy perspective and find that the natural energy gap between known and unknown classes is unstable and easily collapses. We therefore propose Energy Ring to stabilize this gap by anchoring the two types of classes on separate energy rings. A lightweight adapter enlarges the norm-based gap while preserving known-class directions, and a perturbation consistency module forms compact clusters for unknown classes. During testing, energy-guided confidence weighting improves prediction, and a stability-based rectification mechanism corrects unreliable updates. Experimental results show that Energy Ring significantly improves TTD performance against existing TTD methods. Our code will be available.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Test-Time Discovery, Novel Category Discovery, Online Learning, Test-Time Training

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_eu1AOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This work proposes Energy Ring for test-time discovery, using a lightweight adapter to organize known and simulated unknown features. Building on DAA, it combines norm-weighted prediction with perturbation-based stability measures to guide online adaptation and memory renewal, aiming to improve novel-class discovery while preserving known-class recognition.

Critical strengths

The overall idea is appealing, particularly the combination of radial feature structure and stability-guided adaptation. The figures are well designed, and the paper is clearly written and easy to follow. Experiments across four datasets, component ablations, sensitivity analyses, and paired comparisons with DAA provide useful evidence about the method’s benefits and limitations.

Critical weaknesses

1.The contribution appears incremental. Appendix J acknowledges the DAA foundation, while radial constraints and consistency-based selection have precedents in Ring Loss, Objectosphere, and SENTRY. The existing ablations do not establish an advantage over simpler alternatives under matched budgets.

2.The simulated-unknown mechanism lacks targeted validation. Figure 3 shows norm separation without isolating the benefits of mixing and noise. Section 3.2 specifies Gaussian noise, but the submitted train_ERA generates mixed features without adding it.

3.Performance depends on insufficiently documented parameter selection. In Table 14, real-time TA+CA drops from 81.18 with manually selected radii to 71.93 with median calibration. Appendix K.5 mentions grid search without clearly specifying validation data or label access.

4.The submitted defaults diverge from the described method. Prediction weighting uses alpha=0 rather than Table 3’s 0.05; test perturbations imply approximately 84° changes despite the local-stability interpretation; and the default CIFAR warm-up includes held-out-class training images. These findings concern the submitted defaults, not necessarily the reported experiments. The authors need to provide the exact experimental configurations and code revision.

5.TA+CA alone does not establish better discovery quality. Table 11’s 70+30→70+∞ change raises TA+CA from 81.18 to 119.90, driven by CA increasing from 47.66 to 86.55, while TA slightly decreases and post-evaluation TA falls from 37.74 to 23.65. DAA shows a similar trend. Table 10’s 0.74–1.09-point real-time HCA gains provide positive counterevidence, but cluster counts, fragmentation statistics, and unknown-only matching accuracy are still needed to distinguish improved discovery from changes in clustering granularity.

Questions for authors
  1. Which exact code revision and configurations produced the reported results? The submitted defaults disable prediction weighting (alpha=0), apply approximately 84° test perturbations, and include held-out-class training images in CIFAR warm-up. Were these settings overridden, and was Gaussian noise actually added to simulated features?

  2. What validation data, label access, and selection criteria were used to choose the radii, weighting coefficient, and discovery threshold? Were DAA and Energy Ring given comparable tuning budgets, and does the improvement persist under selection using only known-class data?

  3. How do the authors distinguish better discovery from increased fragmentation? In Table 11, unlimited capacity raises TA+CA from 81.18 to 119.90 while TA barely changes and post-evaluation TA declines. Beyond Table 10’s modest overall HCA gains, can the authors report actual cluster counts, per-class fragmentation, and unknown-only one-to-one matching accuracy?

8 comments

W1. Contribution beyond DAA and simpler alternatives.

Relationship to prior work. As acknowledged in Appendix J, DAA provides the foundation for our TTD framework and a directly relevant baseline. We build on its implementation and compare in a common environment to reduce implementation differences. We agree that radial constraints and consistency-based selection have precedents: Ring Loss (Zheng et al., CVPR 2018) softly constrains feature norms toward a common radius; Objectosphere (Dhamija et al., NeurIPS 2018) separates known/unknown magnitudes for open-set rejection; SENTRY (Prabhu et al., ICCV 2021) uses prediction consistency to selectively minimize or maximize entropy during domain adaptation. Our contribution concerns preserving and using radial structure during online novel-class discovery, together with perturbation-based warm-up and stability-guided reassignment of earlier samples (Sec. 3.2–3.4).

Direct comparisons with simpler alternatives. We additionally evaluate DAA with the following norm-margin regularizer during both warm-up and test-time training:

=\mathbb{E}_{(z_k,z_u)} \left[\max\left(0,\;m-\bigl(\lVert z_k\rVert_2-\lVert z_u\rVert_2\bigr)\right)^2\right],\qquad m>0.$$ Norms are computed before feature normalization. During warm-up, $z_k$ and $z_u$ are known and simulated-unknown features; during testing, they follow DAA's known/novel pseudo-label and replay rules. We use $m=20$ in both stages. The loss penalizes gaps below $m$ without fixing absolute radii, and requires no unknown ground-truth labels for training. We also evaluate our method with energy removed: radial training losses and energy-weighted prediction (α=0) are disabled, while perturbation warm-up and stability-guided rectification are retained. | Method | Real Time TA+CA | Post TA+CA | KF | |---|---|---|---| | DAA | 78.42 | 73.52 | 3.54 | | DAA + norm-margin regularization | 76.23 | 73.58 | 3.97 | | Ours w/o energy | 79.82 | 75.63 | 3.50 | | Energy Ring | 81.18 | 76.66 | 3.65 | Even without energy, our method improves real-time and post TA+CA over DAA by **1.40 and 2.11 points**, respectively, supporting the contribution of perturbation warm-up and stability-guided rectification beyond norm preservation. Energy further adds **1.36 and 1.03 points**. Adding a generic norm margin to DAA does not reproduce this benefit: it imposes a relative gap on an angular framework, whereas Energy Ring anchors radial levels according to the observed norm ordering and uses them in both prediction and adaptation. **Why perturbation robustness helps without energy.** During warm-up, perturbed views of the same simulated unknown form positives, while views of different features form negatives (**Sec. 3.2, Eq. (4)**), encouraging consistency and separation. During testing, assignment agreement under perturbations defines stability (**Eq. (9)**). Reliable unknown samples support adaptation, while low-stability queued samples are prioritized for reclassification. They are reinserted according to their revised labels, and the affected prototypes are recomputed as queue means. Thus, earlier assignments can be corrected even when radial information is disabled. The existing **Table 4/Appendix D.4** also shows that adding rectification to T+W raises real-time TA+CA from **76.59 to 81.18** and post TA+CA from **71.04 to 76.66**.

Direct evaluation of reassignment quality. We compare selection policies with the same rechecking budget and a fixed offline cluster-to-label mapping. A correct assignment places a sample in its correct novel-class queue, including moving to another queue or returning to its original correct queue after updating. An incorrect assignment places it in an incorrect queue, including correct-to-incorrect and incorrect-to-incorrect changes. Samples predicted as known are removed from novel queues and counted separately; this includes known samples previously misassigned to novel queues and subsequently recognized as known. All percentages use the total number of rechecking events as their denominator, so the three outcomes sum to 100%. Labels are used only for evaluation. DAA + STMR provides the random-rechecking control.

ModelSelectionCorrect assignment (%)Incorrect assignment (%)Removed as known (%)
DAA + STMRRandom35.0830.2034.72
Ours w/o energyLowest stability43.5925.5130.90
OursRandom38.6228.7432.64
OursLowest stability56.3421.3922.27

Our full method achieves the highest correct-assignment rate and lowest incorrect-assignment rate. Within our method, stability-guided selection improves these rates over random selection by 17.72 and 7.35 percentage points, respectively. Even without energy, our method produces more correct and fewer incorrect assignments than DAA + STMR (43.59% vs. 35.08%; 25.51% vs. 30.20%). These results support the contribution of perturbation robustness and stability-guided correction beyond norm preservation.

We will report warm-up/online update counts, memory/replay limits, and tuning-trial counts for the compared variants with their exact configurations. The common implementation controls implementation differences; the tuning-budget and selection protocol are addressed separately in W3/Q2 below.

The protocol distinction is already stated in Appendix A/Table 5, reproduced below (K: known classes; U: unknown classes). TTD requires both discovery and evaluation in real time, making DAA the direct baseline for the comparisons above.

TaskTrainTestDiscoveryEvaluation
NCDKUPost-hocPost-hoc
GCDKK + UPost-hocPost-hoc
CGCDKK + U₁ + U₂ + …Post-hocPost-hoc
OCDKK + UReal-timePost-hoc
TTAKShifted KN/AReal-time
TTDKK + UReal-timeReal-time

W2/W4/Q1. Exact implementation, synthetic noise, and perturbations.

Thank you for distinguishing the submitted defaults from the configurations used for the reported experiments. We will provide the complete, cleaned code with a pinned experiment-matched revision and per-table configuration files, covering the effective α, known-class filtering, Gaussian-noise scale, and perturbation settings. In particular, the α=0 value is only a placeholder in the code constructor and is definitely not the experimental default; it should not be interpreted as the setting used for the reported results. We confirm that the experimental project uses additive Gaussian noise after MixUp and known-class images only for warm-up, consistent with Sec. 3.2 and Appendix B. The submitted package was a simplified, draft version as a part of Supplementary Material and the complete release will make each reported result traceable to its executed configuration.

For the simulation mechanism, MixUp places features between known representations, while Gaussian noise broadens these neighborhoods. Our simulated-unknown construction follows DAA’s feature-space simulation strategy, using MixUp with Gaussian noise (Sec. 3.2; Appendix I/Fig. 12). Its purpose is to expose the adapter to unknown-like feature patterns and prepare it for subsequent discovery. We additionally evaluated alternative simulation strategies, which also provided effective initialization, indicating that our framework is not restricted to this particular construction. Our methodological focus is the energy structure and stability-guided adaptation. Actual novel-class prototypes are initialized from arriving test samples and refined during discovery.

To clarify the role of the simulated-unknown construction, we add an ablation that isolates the effects of MixUp and Gaussian noise.

Simulate strategySynthetic featureSimulated norm before warm-upTrue K/U norm gap after warm-upReal Time TA+CAPost TA+CAKF
NoneNot usedN/A5.8362.5955.835.37
Noise onlyr + n64.076.3070.3569.394.81
MixUp onlyλr + (1−λ)r′56.518.5578.6772.754.10
MixUp + noiseλr + (1−λ)r′ + n52.6410.1881.1876.663.65

These results show that the DAA-style MixUp + Gaussian-noise construction is effective because MixUp places synthetic features between known representations while Gaussian noise broadens these neighborhoods, exposing the adapter to unknown-like patterns and preparing it for subsequent discovery. At the same time, alternative simulation strategies also provide effective initialization, indicating that our framework is not restricted to this particular construction. Our methodological focus remains the energy structure and stability-guided adaptation and actual novel-class prototypes are initialized from arriving test samples and refined during discovery, rather than being fixed by the simulation strategy.The known-only training protocol is already specified in Appendix B/Table 6. The full split table is reproduced below; test labels are withheld from the online learner and used for evaluation.

DatasetSplitLabelledKnown classesUnknown classesSamples
CIFAR100DTrainYes70035,000
TestNo703010,000
CUB200DTrainYes14004,195
TestNo140605,794
Tiny-ImageNetDTrainYes140070,000
TestNo1406010,000
AircraftDTrainYes7002,331
TestNo70303,333

For example, CIFAR100D warm-up uses 35,000 images from 70 known classes, the 30 held-out classes are excluded from warm-up and first encountered by the model at test time. We will supply the filtering configuration and sample counts. Table 3 reports α=0.05 as the best listed weighting setting; the release will explicitly distinguish the package's α=0 default from the effective value used for each result, rather than treating a default as the experimental configuration.

Reviewer_29WTOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Energy Ring, a TTD framework designed to recognize known classes while continuously discovering novel classes from unlabeled test streams. Unlike prior methods that primarily rely on angular feature discrimination, the proposed approach exploits feature-norm differences between known and unknown classes.

Critical strengths
  1. Interesting motivation from feature geometry. The paper identifies feature magnitude as a potentially useful signal for distinguishing known and unknown classes, providing an alternative perspective to existing direction-based TTD approaches.
  2. A coherent framework for discovery and adaptation. The combination of energy-ring constraints, perturbation consistency, prototype-based prediction, and test-time rectification provides a unified approach to balancing novel-class discovery and known-class retention.
Critical weaknesses
  1. Limited methodological novelty beyond DAA. The proposed framework largely follows DAA, with the principal distinction being the preservation and amplification of feature-norm differences through dual energy-ring constraints. Other components, including perturbation consistency, prototype matching, and memory-based rectification, are closely related to established techniques. The paper should better clarify what fundamentally distinguishes the proposed framework from a relatively straightforward extension of DAA with norm regularization.
  2. Warm-up dependency. The proposed framework relies on simulated unknown features generated from known-class representations to establish the dual energy rings before testing. However, it remains unclear whether these simulated features provide a reliable approximation of genuinely unseen categories. Moreover, the sensitivity to predefined ring radii raises concerns about generalization across different feature distributions. The current experiments do not sufficiently disentangle the benefits of warm-up initialization from those of subsequent test-time adaptation.
  3. Insufficient clarification and validation of online novel-class learning. The method dynamically introduces prototypes for newly discovered classes and updates the adapter using pseudo-labels and memory queues. However, the manuscript does not sufficiently clarify how discovered-class prototypes are incrementally maintained, how unreliable initial assignments are corrected, or whether stability-based rectification can effectively prevent error accumulation. These issues are particularly important for continual discovery in unlabeled streams.
  4. Potentially misleading motivation in Fig. 1. The comparison suggests that DAA suffers from limited directional discrimination, whereas Energy Ring benefits from intensity differences. However, DAA explicitly optimizes feature directions, and a shared feature norm does not necessarily imply insufficient angular separability. The primary distinction appears to be direction-only versus joint radial-angular discrimination, rather than an inherent limitation in DAA's ability to separate feature directions.
Questions for authors
  1. Novelty relative to DAA: Could the authors compare Energy Ring against DAA augmented with a simple feature-norm margin or regularization term? This would help determine whether the performance improvements arise from the proposed dual-ring formulation or from introducing feature-magnitude information more generally.
  2. Warm-up dependency: How much of the improvement comes from warm-up rather than online test-time adaptation? Could the authors provide experiments without warm-up or with different warm-up initialization qualities, and evaluate whether test-time adaptation can recover from suboptimal initial energy rings?
  3. Novel-class learning and rectification: How are prototypes updated after their initialization from the first discovered sample? Can the stability-based rectification mechanism revise incorrect class assignments or merge mistakenly separated novel classes? Further implementation details and analysis would strengthen the claimed reliability of continual discovery.
  4. Validity of the energy assumption: To what extent does the assumption that known classes have higher feature norms than unknown classes hold across different backbones and distribution shifts? How sensitive is the approach to the choice of energy-ring radii, and can these radii be determined without dataset-specific tuning?
4 comments

W1/Q1 Novelty relative to DAA and the interpretation of Fig. 1.

Thank you for identifying an ambiguity in Fig. 1. DAA explicitly optimizes angular discrimination; a shared norm does not prevent angular separation. The intended distinction is that normalization removes the radial coordinate, whereas Energy Ring combines radial and angular information. We will revise the illustration, caption, and related wording to make this distinction explicit.

For the suggested simple baseline, we evaluated DAA with the following additional norm-margin regularizer during both warm-up and test-time training:

=\mathbb{E}_{(z_k,z_u)} \left[\max\left(0,\;m-\bigl(\lVert z_k\rVert_2-\lVert z_u\rVert_2\bigr)\right)^2\right], \quad m>0.$$ Operationally, the norm-margin regularizer is applied to feature norms after backbone and all norms are computed **before feature normalization**. During warm-up, the norm-margin regularizer is applied to pairs of known-class features and simulated-unknown features. During test-time training, it is applied to features assigned to known and discovered novel classes according to DAA's existing pseudo-label and replay rules. The same margin m is used in both stages; unknown ground-truth labels are not needed. The loss penalizes a norm gap smaller than m and becomes zero once the gap reaches m, so it encourages a **minimum gap**, without fixing either absolute radius. In our experiments, we set $m=20$ (the backbone feature norm is around 70, and the loss then scaled), resulting **average feature norm is 1.15 for known classes and 0.87 for unknown classes**, but we believe that **such a modification has already deviated from DAA's original design**. | Method | Real Time TA+CA | Post TA+CA | KF | |---|---|---|---| | DAA | 78.42 | 73.52 | 3.54 | | DAA + norm-margin regularization | 76.23 | 73.58 | 3.97 | | Ours | 81.18 | 76.66 | 3.65 | The experiment shows that **forcing a norm margin onto DAA is not appropriate and can even be counterproductive**. DAA's original normalized-feature formulation focuses on angular discrimination; adding a norm gap introduces an auxiliary radial constraint, while its cosine-based prediction still discards magnitude. Energy Ring instead builds on the observed known/unknown norm ordering, anchors the two groups to distinct radial levels, and uses this structure in both prediction and continued adaptation. This result is consistent with that design difference: imposing a generic margin on an angular framework does not reproduce the benefit of preserving and using radial structure throughout discovery.

W2/Q2 Simulated unknowns and warm-up versus online adaptation.

The simulation rationale is already discussed in Appendix I/Fig. 12, the existing online-component ablation is Table 4, analyzed in Appendix D.4. We reproduce the latter first. T denotes test-time training, W radial prediction weighting, and P perturbation rectification.

TWPRT TART CART TA+CAPost TAPost CAPost TA+CAKF
---30.8142.1372.9434.9138.5373.442.27
✓--24.6853.7478.4224.8942.6867.576.84
-✓-29.6145.5475.1532.4436.4568.893.05
✓✓-31.6044.9976.5934.5236.5271.044.22
✓✓✓33.5247.6681.1837.7438.9276.663.65

Adding P to T+W improves real-time TA+CA by 4.59 points, post TA+CA by 5.62 points, and reduces KF by 0.57. T alone increases CA but raises KF to 6.84, while T+W+P gives the strongest joint real-time/post-evaluation result. These results already quantify the contribution of online rectification. Appendix K.1/Fig. 13 additionally evaluates warm-up perturbation count; Appendix K.2–K.3/Figs. 14–15 evaluate rectification frequency and replay size.

To disentangle initialization from subsequent adaptation, we compared test-time adaptation without warm-up, warm-up without test-time adaptation, and the full method, keeping the prediction framework and other experimental settings fixed. Test-time adaptation includes adapter updates and stability-guided rectification.

Warm-upTest-time adaptationReal Time TA+CAPost TA+CAKF
✗✓62.5955.835.37
✓✗75.1568.893.05
✓✓81.1876.663.65

Removing warm-up leads to poor discovery performance despite test-time adaptation, showing that online updates do not compensate for inadequate initialization in this setting. Conversely, warm-up alone provides an initial representation but limits its adaptation to the actual novel classes encountered during testing. Enabling test-time adaptation increases real-time TA+CA from 75.15 to 81.18 and post TA+CA from 68.89 to 76.66. These results demonstrate complementary roles which warm-up establishes initial separation and perturbation robustness, while test-time learning refines the representation using the observed stream.

For the simulation mechanism, MixUp places features between known representations, while Gaussian noise broadens these neighborhoods. Our simulated-unknown construction follows DAA’s feature-space simulation strategy, using MixUp with Gaussian noise (Sec. 3.2; Appendix I/Fig. 12). Its purpose is to expose the adapter to unknown-like feature patterns and prepare it for subsequent discovery. We additionally evaluated alternative simulation strategies, which also provided effective initialization, indicating that our framework is not restricted to this particular construction. Our methodological focus is the energy structure and stability-guided adaptation. Actual novel-class prototypes are initialized from arriving test samples and refined during discovery.

Simulate strategySynthetic featureSimulated norm before warm-upTrue K/U norm gap after warm-upReal Time TA+CAPost TA+CAKF
NoneNot usedN/A5.8362.5955.835.37
Noise onlyr + n64.076.3070.3569.394.81
MixUp onlyλr + (1−λ)r′56.518.5578.6772.754.10
MixUp + noiseλr + (1−λ)r′ + n52.6410.1881.1876.663.65

W3/Q3 Prototype maintenance, reassignment, and error accumulation.

The maintenance mechanism is described in Sec. 3.3–3.4, with rectification-frequency and replay-size analyses in Appendix K.2–K.3/Figs. 14–15. A newly discovered class is initialized with its first test feature; subsequently, its prototype is the mean of its maintained feature queue, μc=∣Qc∣−1∑z∈Qcz\mu_c=|Q_c|^{-1}\sum_{z\in Q_c}z. Queue renewal keeps the prototype responsive to recent samples. Stability SS in Eq. (9) measures assignment consistency under perturbations. High-stability unknown samples support adaptation, while low-stability queued samples are prioritized for re-evaluation. Rechecked samples are reassigned to the corresponding novel-class queues or removed from novel queues if recognized as known, updating the affected queue means. Rectification therefore revises sample assignments without a separate prototype-merging operation.

We also separately measures correction success under different strategies. Correct assignment means that a sample is assigned to the correct novel-class queue, either by moving to another queue or returning to its original correct queue after updating. Incorrect assignment includes both correct-to-incorrect and incorrect-to-incorrect queue assignments. Other situation occurs when a sample is reclassified as known and removed from the novel-class queue and this includes known-class samples initially misassigned to novel queues and subsequently recognized as known.

Let NcN_c, and NeN_e count correct and incorrect outcomes, respectively, and NN denote the total number of rechecking events. We report:

=100\\%\times\frac{N_c}{N},\qquad \mathrm{Incorrect\ rate} =100\\%\times\frac{N_e}{N},$$ | Model | Selection | Correction (%) ↑ | Damage (%) ↓ | | --- | ---- | ---- | -- | | DAA + STMR | Random | 35.08 | 30.20 | | Ours w/o energy | Lowest stability | 43.59 | 25.51 | | Ours | Random | 38.62 | 28.74 | | Ours | Lowest stability | 56.34 | 21.39 | Our full method achieves the highest correct-assignment rate (56.34%) and the lowest incorrect-assignment rate (21.39%) among the compared variants. Compared with random selection within our method, stability-guided selection increases correct assignments by **17.72 percentage points** and reduces incorrect assignments by **7.35 points**, supporting its effectiveness in directing rechecking toward unreliable samples. Even without energy, our method outperforms DAA + STMR on both measures (43.59% vs. 35.08% correct; 25.51% vs. 30.20% incorrect). These results show that the perturbation and rectification mechanisms remain beneficial without energy, while the full method delivers the best reassignment quality. ## Q4/W4 Validity of the energy assumption. Experiments are already included in **Appendix D.2–D.3, Tables 7–8** and **Appendix L.1, Eq. (23)/Table 14**, together with the radius sweep in **Fig. 7**. Table 7 reports the natural norm differences across ViT-B/16, VGG16, and ResNet18. The norm of known class features is indeed to some extent larger than that of unknown class features. Energy Ring is designed to exploit and strengthen the cue when it is present, not to rely on a fixed frozen-backbone gap. | Setting | Backbone | known-class norm | unknown-class norm | simulated-class norm | | -- | -- | ---: | -: | ---: | | CIFAR100D | ViT-B/16 | 67.00 | 63.25 | 52.64 | | CUB200D | ViT-B/16 | 71.73 | 68.21 | 58.27 | | Tiny-ImageNetD | ViT-B/16 | 72.09 | 66.52 | 56.48 | | AircraftD | ViT-B/16 | 66.13 | 64.01 | 59.02 | | CIFAR100D | VGG16 | 92.17 | 82.02 | 73.91 | | CIFAR100D | ResNet18 | 28.51 | 17.44 | 15.72 | | CUB200D | VGG16 | 85.39 | 78.18 | 74.17 | | CUB200D | ResNet18 | 32.58 | 27.95 | 25.22 |

Thank you again for your time and effort in reviewing our submission. We have submitted our response to your questions and would appreciate it if you could take a moment to review our response. Please feel free to let us know if any further clarification is needed.

Reviewer_9GzcOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies Test-Time Discovery (TTD): after deployment, a model must both recognize known classes and discover and learn novel classes on a continuous, unlabeled test stream. The authors revisit TTD from an "energy" perspective, observing that the natural feature-energy (feature-norm) gap between known and unknown classes is unstable and easily collapses. They propose Energy Ring, which anchors known classes on a high-energy ring and unknown classes on a low-energy ring. It consists of four parts: a lightweight energy ring adapter (ERA) that enlarges the norm gap while preserving known-class feature directions; a perturbation-consistency module that forms compact, stable clusters for unknown classes; energy-weighted confidence prediction at test time; and a stability-based rectification mechanism that corrects unreliable updates. On CIFAR100D, CUB-200D, Tiny-ImageNetD, and AircraftD, it improves TA/CA and clustering metrics over baselines such as HM and DAA.

Critical strengths
  1. Clear perspective and a real problem. The observation that the known/unknown energy gap naturally exists but is unstable speaks directly to TTD's core difficulty (catastrophic forgetting vs. novel-class discovery); anchoring the two onto separated rings is intuitive and practical, a targeted fix for DAA's norm-erasing design.
  2. Complete and reproducible method. Warm-up + test-time stages, energy-weighted prediction, and stability rectification form a coherent pipeline with full ablation and hyperparameter analysis, and code will be released.
Critical weaknesses
  1. Small improvements without significance testing. Compared with the DAA baseline, most metrics improve by only 1–2 points and some are lower. The "3.33%" claimed in the abstract does not appear in the experiments, and significance testing is missing.
  2. Energy-ring radii require manual presetting, suggesting a potential generalization issue with the method.
Questions for authors
  1. Which metric and dataset does the "3.33%" claimed in the introduction refer to? I did not see 3.33% in the experimental section. If the method can genuinely deliver 3.33%, the absence of significance tests might be tolerable. Otherwise, can the authors report significance tests against DAA on the four datasets? As it stands the improvements look small, and significance support is needed to substantiate the effectiveness claim.
  2. The energy-ring radii are currently preset and depend on the backbone's norm scale. How would they be set adaptively for a different backbone or scenario?
  3. Beyond preserving the norm difference, is there any substantive mechanism difference from DAA? Can the authors provide an ablation that directly isolates "removing only the norm-preservation component" to identify the source of the contribution?
4 comments

W1/Q1. Magnitude of improvement, the “3.33%” statement, and significance.

The 3.33% statement is a descriptive summary: the mean relative improvement in TA+CA over DAA, equally averaged across four datasets and real-time/post-evaluation protocols (eight comparisons). With S = TA+CA,

\frac{S^{\mathrm{ER}}_{d,p}-S^{\mathrm{DAA}}_{d,p}}{S^{\mathrm{DAA}}_{d,p}}\times100\%.$$ We used this aggregate to give a concise initial overview in the abstract and introduction, and will state its metric and averaging protocol explicitly. The rounded Table 1 entries yield approximately 3.33%; we will provide the full-precision calculation to resolve the last-digit difference. We also assess statistical robustness in **Appendix M Table 16**, so the aggregate is not the sole basis for our effectiveness claim. The statistical evidence is provided below. Each comparison uses ten paired full-stream runs sharing the class split, pretrained checkpoint, stream order, and batch partition. We pre-specify `TA+CA` as the primary discovery metric and report paired mean differences, 95% confidence intervals, two-sided paired t-tests, Cohen's `d_z`, raw and Holm-corrected p-values, and paired sign-flip p-values. This paired design tests the method difference directly rather than inferring significance from overlapping marginal ranges. | Dataset | Protocol | Runs | DAA `TA+CA` (mean +/- SD) | Energy Ring `TA+CA` (mean +/- SD) | Delta [95% CI] | `t(9)` | raw p | Holm p | `d_z` | sign-flip p | | --- | --- | ---: | --- | --- | --- | ---: | ---: | ---: | ---: | ---: | | CIFAR100D | real-time | 10 | 76.69+/-1.27 | 81.17+/-1.25 | 4.4816 [3.4196, 5.5435] | 9.5468 | 0.000005 | 0.000011 | 3.0190 | 0.001953 | | CIFAR100D | post | 10 | 72.90+/-1.62 | 76.01+/-1.85 | 3.1103 [1.5289, 4.6918] | 4.4491 | 0.001602 | 0.001602 | 1.4069 | 0.003906 | | CUB200D | real-time | 10 | 112.84+/-1.93 | 115.92+/-1.43 | 3.0728 [1.2899, 4.8556] | 3.8989 | 0.003625 | 0.007251 | 1.2330 | 0.005859 | | CUB200D | post | 10 | 108.86+/-2.49 | 110.67+/-1.48 | 1.8055 [-0.1187, 3.7298] | 2.1226 | 0.062775 | 0.062775 | 0.6712 | 0.066406 | | Tiny-ImageNetD | real-time | 10 | 95.36+/-1.85 | 98.16+/-1.31 | 2.7923 [1.2771, 4.3076] | 4.1687 | 0.002416 | 0.002416 | 1.3183 | 0.005859 | | Tiny-ImageNetD | post | 10 | 59.59+/-1.51 | 62.63+/-1.29 | 3.0402 [1.6200, 4.4604] | 4.8426 | 0.000918 | 0.001836 | 1.5314 | 0.001953 | | AircraftD | real-time | 10 | 68.60+/-0.62 | 69.78+/-1.59 | 1.1766 [-0.1337, 2.4870] | 2.0314 | 0.072771 | 0.072771 | 0.6424 | 0.087891 | | AircraftD | post | 10 | 57.34+/-1.69 | 59.87+/-1.87 | 2.5254 [0.6865, 4.3642] | 3.1068 | 0.012584 | 0.025169 | 0.9824 | 0.015625 | We conducted 10 matched full-stream runs per dataset and protocol. All eight mean differences are positive, and **six comparisons remain significant at 0.05 with the reported Holm correction**; the exact sign-flip tests support the same six comparisons. CUB200D post-evaluation and AircraftD real-time show positive but statistically inconclusive gains. These paired tests support effectiveness across the evaluated settings while identifying where the evidence is weaker.

W2/Q2 Radius selection.

These experiments are already included in Appendix D.2–D.3, Tables 7–8 and Appendix L.1, Eq. (23)/Table 14, together with the radius sweep in Fig. 7.

We compare manual and median-calibrated radii in Appendix Table 14.

Radius choiceRT TART CAPost TAPost CAKF
Manual [70,40]33.5247.6637.7438.923.65
Median-calibrated [67,52]30.6541.2833.7934.213.93

rkn=median⁡x∈Dkn∥f(x)∥2,run=median⁡r~∈Dsim∥r~∥2.r_{\mathrm{kn}} = \operatorname{median}_{x \in D_{\mathrm{kn}}}\lVert f(x)\rVert_2, \qquad r_{\mathrm{un}} = \operatorname{median}_{\tilde{\mathbf r} \in D_{\mathrm{sim}}}\lVert \tilde{\mathbf r}\rVert_2.

Eq. (23) in appendix estimates r_kn from the median known-feature norm and r_un from the median simulated-feature norm, using only warm-up data. It produces approximately (67,52), without unknown-class validation labels. The manual pair (70,40) is shared across all four ViT datasets; r_kn = 70 is consistent with their known-feature means of 66.13–72.09 in Table 7. Median calibration preserves the ordering but produces a smaller radial gap: real-time TA+CA is 71.93 versus 81.18, and post TA+CA is 68.00 versus 76.66. This already demonstrates automatic scale calibration, while also showing that it does not yet match the selected pair.

Fig. 7 in main text also shows the grid search of radius. With rknr_{kn} fixed at 70, runr_{un} = 30–50 gives 80.3–81.2, whereas the extreme gap at runr_{un} = 10 gives 65.0 and equal radii give 78.2. The full figure also varies rknr_{kn}, supporting a moderate gap near the natural feature scale. Across backbones the absolute scale changes substantially, so Table 7 supports the ordering prior rather than transfer of the identical (70,40) pair.

r_knr_un = 10203040506070
7065.073.880.681.280.379.278.2

We also compare observed versus reversed radial ordering in Appendix Table 8. The observed ordering outperforms the reversed assignment by 6.92 points in real-time TA+CA and 6.12 points post-evaluation. Reversing it also falls below DAA. Thus, the gain depends on preserving the empirical known/unknown ordering; imposing any two arbitrary radii is insufficient. The mechanism is that the observed-order objective enlarges an existing radial difference while preserving known-class directions; reversing it instead forces initially lower-norm unknown features across the known-feature radial region

Assignmentr_kn / r_unRT TART CART TA+CAPost TA+CAKF
DAA—32.4046.0278.4273.523.54
Observed order70 / 4033.5247.6681.1876.663.65
Reversed order40 / 7030.0944.1774.2670.543.82

Q2 Raddi transfer to another backbone and scenario.

Table 7 in Appendix reports the natural norm differences across ViT-B/16, VGG16, and ResNet18. The following analysis reports average feature norms across datasets and backbones. The same ordering appears in all listed same-domain settings, including ViT-B/16, VGG16, and ResNet18.

SettingBackboneknown-class normunknown-class normsimulated-class norm
CIFAR100DViT-B/1667.0063.2552.64
CUB200DViT-B/1671.7368.2158.27
Tiny-ImageNetDViT-B/1672.0966.5256.48
AircraftDViT-B/1666.1364.0159.02
CIFAR100DVGG1692.1782.0273.91
CIFAR100DResNet1828.5117.4415.72
CUB200DVGG1685.3978.1874.17
CUB200DResNet1832.5827.9525.22
*CIFAR10-CVGG1683.5282.1281.06
*CIFAR10-CResNet1830.1428.8027.37
*CIFAR100-CVGG1684.8584.3184.05
*CIFAR100-CResNet1829.6729.5628.35

The corrupted-domain results make the limitation explicit: under CIFAR10-C and CIFAR100-C, the norm gaps become much smaller. Simultaneous severe domain shift and novel-category discovery is a broader open-world problem than the current TTD benchmark; in that setting, radial evidence can weaken or vanish.

Q3 Beyond preserving the norm difference, is there any substantive mechanism difference from DAA?

Yes. Beyond the radial objective, our method retains perturbation-based warm-up for unknown representations and stability-guided online rectification. These are described in Sec. 3.2, Eq. (4), and Sec. 3.3–3.4, Eqs. (9)–(12), with further discussion in **Appendix D.4 and J **.

During warm-up, perturbations of the same simulated unknown feature form positives, while perturbations from different features form negatives. The objective encourages consistency within each group and separation between groups. During testing, Eq. (9) measures whether a sample's assignment survives repeated perturbations. Reliable unknown samples support adaptation, while queued samples with low stability are prioritized for re-evaluation and label reassignment. Thus, even without energy information, the method still learns perturbation robustness and uses it to direct self-correction, rather than selecting samples at random.

The existing component ablation in Table 4 already shows that adding perturbation rectification to T+W increases real-time TA+CA from 76.59 to 81.18 and post TA+CA from 71.04 to 76.66. We additionally directly compare our method with and without energy, alongside DAA as below. Removing energy disables the radial loss during warm-up/online adaptation and energy-weighted prediction (α=0), while keeping perturbation warm-up, stability-guided correction, and all other settings unchanged.

MethodReal Time TA+CAPost TA+CAKF
DAA78.4273.523.54
Ours w/o energy79.8275.633.50
Ours81.1876.663.65

Even without energy, our method improves real-time TA+CA over DAA from 78.42 to 79.82 and post TA+CA from 73.52 to 75.63, without increasing forgetting. These results support the contribution of perturbation-based warm-up and stability-guided rectification beyond energy modeling: learning perturbation robustness and using it for self-correction remain beneficial when the energy components are disabled. Incorporating energy further improves real-time and post TA+CA by 1.36 and 1.03 points, respectively, showing that radial information complements these mechanisms.

We also separately measures correction success under different strategies. Correct assignment means that a sample is assigned to the correct novel-class queue, either by moving to another queue or returning to its original correct queue after updating. Incorrect assignment includes both correct-to-incorrect and incorrect-to-incorrect queue assignments. Other situation occurs when a sample is reclassified as known and removed from the novel-class queue and this includes known-class samples initially misassigned to novel queues and subsequently recognized as known.

ModelSelectionCorrection (%) ↑Damage (%) ↓
DAA + STMRRandom35.0830.20
Ours w/o energyLowest stability43.5925.51
OursRandom38.6228.74
OursLowest stability56.3421.39

Our full method achieves the highest correct-assignment rate (56.34%) and the lowest incorrect-assignment rate (21.39%) among the compared variants. Compared with random selection within our method, stability-guided selection increases correct assignments by 17.72 percentage points and reduces incorrect assignments by 7.35 points, supporting its effectiveness in directing rechecking toward unreliable samples. Even without energy, our method outperforms DAA + STMR on both measures (43.59% vs. 35.08% correct; 25.51% vs. 30.20% incorrect). These results show that the perturbation and rectification mechanisms remain beneficial without energy, while the full method delivers the best reassignment quality.

Thank you again for your time and effort in reviewing our submission. We have submitted our response to your questions and would appreciate it if you could take a moment to review our response. Please feel free to let us know if any further clarification is needed.

Comments on the paper

1 comment

Dear AC and all reviewers,

We sincerely thank you for the detailed reviews and constructive suggestions. Most empirical questions only concern hyperparameter sensitivity or the validity of reported observations. The submitted appendix already addresses many of these questions, and we supplemented those analyses with targeted experiments in the rebuttal. After careful revisions and clarifications, we believe all concerns have been fully addressed through the rebuttal and updated experiments.

  • Reported gains (9Gzc). We clarified that 3.33% is the average relative TA+CA gain over DAA across four datasets and two evaluation protocols. All eight paired comparisons favor Energy Ring on average, with six statistically significant after Holm correction (Appendix M, Table 16).

  • Norm differences and radius selection (9Gzc, 29WT, eu1A). Appendices D and L document norm differences across backbones and sensitivity to radius choice. Energy Ring's radii were fixed using training statistics, and DAA used its published settings.

  • Contribution beyond DAA (9Gzc, 29WT, eu1A). Energy Ring outperforms DAA with a simple norm margin. The variant without radial components still improves over DAA, supporting the role of perturbation-based warm-up and rectification. Additional ablations separate the effects of simulation, warm-up, and online adaptation.

  • Reassignment and fragmentation (29WT, eu1A). Under the same rechecking budget, stability-guided selection raises correct novel-queue assignments from 38.62% to 56.34% compared with random selection. Under unlimited capacity on CIFAR100D, Energy Ring improves unknown-only matching accuracy while producing fewer occupied clusters than DAA.

  • Experimental configuration (eu1A). We clarified known-class warm-up, Gaussian noise in simulated features, and perturbation settings. We will provide the code revision and configurations used for the reported experiments.

We appreciate your consideration of the submitted and additional evidence in the final decision.

Best regards,

The authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.