acceptodds
Desk rejected at ICLR 2027

Learning to Route Supervision Along the Path: Route Optimization for Gradual Domain Adaptation

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Gradual domain adaptation progressively updates a source-trained classifier through unlabeled intermediate domains before reaching the target. However, existing methods often prescribe the update rule or weight schedule in advance, while the reliability of different supervision sources can change along the adaptation path. Source predictions can anchor early updates but become less informative than the adapting model as the domain shifts. Therefore, choosing appropriate supervision at each stage without target labels remains challenging. Based on the above analysis, we introduce PrismBridge, which selects bridge rules that combine frozen source predictions, class-prototype probabilities, and current-model predictions. Specifically, we formulate this choice as route-relaxed path optimization, where fixed-route policies are treated as constrained special cases. Candidates starting from the same state are evaluated using bridge losses, source supervision, and view agreement. A tolerance-aware selector then chooses among near-best candidates according to a predetermined preference order, and source-supervised smoothing stabilizes the selected update. Furthermore, we derive cumulative risk and selection-regret bounds along each policy's route-dependent model trajectory. Experiments across geometric, appearance, temporal, tabular, and corruption shifts demonstrate the effectiveness of PrismBridge, achieving severity-5 accuracy improvements of 8.1 and 15.9 percentage points over the strongest compared baseline on CIFAR-10-C and CIFAR-100-C.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Gradual Domain Adaptation, Adaptive Supervision Routing, Route Optimization, Self-Training, Unsupervised Domain Adaptation

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_DyBMOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a framework for stage-wise supervision selection in gradual domain adaptation. The method integrates frozen source-model predictions, source-prototype probabilities, and current-model predictions into a unified candidate bridging mechanism, selecting supervision routes using candidate scores, a tolerance criterion, and a predefined priority order. The paper establishes bounds on selection regret and cumulative target risk that account for model trajectories induced by routing decisions. Extensive experimental results demonstrate the effectiveness of the proposed method.

Critical strengths

The paper provides a unified framework for stage-wise selection among multiple supervision signals in gradual domain adaptation. The theoretical analysis accounts for model trajectories induced by routing decisions and establishes bounds on selection regret and cumulative target risk, providing theoretical support for the method. Experiments cover various domain shifts, with reported results outperforming the compared methods on multiple benchmarks.

Critical weaknesses
  1. The novelty of the method is relatively limited. The framework integrates several existing supervision mechanisms, with its main addition being stage-wise supervision routing based on local scores, a tolerance criterion, and a predefined priority order. While the overall design is reasonable, its novelty lies primarily in the integration and selection of existing mechanisms.
  2. The analysis of supervision reliability and route-selection effectiveness is insufficient. The central motivation is that the reliability of different supervision sources changes along the adaptation path, necessitating adaptive supervision selection. However, the experiments primarily report final accuracy and component ablations, without directly showing how the quality of frozen-source predictions, prototype-based probabilities, and current-model predictions evolves across adaptation stages.
  3. Key implementation details remain unclear. The paper does not fully specify the candidate routes’ mixture weights, the predetermined preference order, the rule for computing stage-specific tolerances, or the per-stage update budgets.
  4. The correspondence between the framework diagram and method description is not apparent. Figure 2 presents fixed-route mismatch, a route-relaxed bridge-and-smooth stage, and ordered path optimization. However, the mapping between these elements and individual subsections of method is not explicitly indicated.
Questions for authors

See Weaknesses.

7 comments
AuthorsOpenReview ↗

Response to Reviewer DyBM (1)

Thank you for your kindly comments. The following are our response to your concerns.

1. What is the contribution beyond the supervision components? (W1)

PrismBridge introduces a route-relaxed formulation of gradual adaptation: the supervision rule can change across stages, with globally fixed routes retained as constrained special cases.

This targets a different decision from recent methods. STDW (NeurIPS 2025) schedules neighboring-domain loss weights; ProDe (ICLR 2025) corrects vision-language teacher predictions; E-SUOT (ICML 2026) generates intermediate domains through optimal transport; and TCA (CVPR 2025) regularizes class-centroid topology. PrismBridge selects among supervision routes constructed from frozen-source, prototype, and current-model signals. Its routing also differs from BECoTTA (ICML 2024), whose routers activate low-rank network experts: our decision concerns which supervision update to apply at the current stage.

Each selected update changes the classifier and prototype state used by later domains, and our trajectory-dependent analysis accounts for this coupling. The practical selector forms a near-best candidate set using a score tolerance and applies the specified route preference within that set.

In the three-corruption CIFAR comparison, adaptive selection reaches 86.41%, exceeding the fixed plain route within the same control framework (80.69%) by 5.72 percentage points.

2. How does the reliability of each supervision source evolve? (W2)

We directly measure the prediction accuracy of all three supervision sources across the five adaptation stages on Portraits ResNet, averaging three adaptation seeds. Frozen-source, prototype, and current-model predictions are evaluated before each stage update; the final column reports accuracy after the selected update.

StageFrozen source (%)Prototype (%)Current model before update (%)PrismBridge after update (%)
197.8096.4597.8098.02
298.5099.2299.2599.15
393.9094.7894.7097.40
493.2594.9795.3096.40
583.3585.0085.5091.15

Their relative reliability changes along the path: the source leads prototypes by 1.35 points at stage 1, whereas prototypes lead the source by 1.65 points at stage 5. PrismBridge reaches 91.15% at the final stage, exceeding the strongest pre-update supervision signal (85.50%) by 5.65 percentage points.

AuthorsOpenReview ↗

Response to Reviewer DyBM (2)

3. What are the route weights, preference, tolerance, and update budgets? (W3)

Route weights. For the Office-Home Clipart-target branch, the probability-fusion weights are fixed within each configuration:

Office-Home transferFrozen sourcePrototypeCurrent model
Art/Real-World → Clipart0.350.350.30
Product → Clipart0.200.450.35

These weights apply before class-balance correction. The supplementary Portraits/CIFAR route execution settings are:

ConfigurationCandidate routesTeacher-KL coefficientPrototype-loss coefficient
Portraits CNN32smooth, relaxed, lowpass_plain, lowpass_relaxed, sharpen0 for all routes0
Portraits ResNet96anchor, plain, relaxed, lowpass_plain, lowpass_relaxed, sharpen_anchor, blend_plain0.08 for anchor/sharpen_anchor; 0 otherwise0.03
CIFAR base libraryanchor, plain, relaxed0.12 for anchor; 0 otherwise0

The relaxed and lowpass_relaxed routes disable the extra prototype-agreement filter; the other listed routes enable it. The lowpass, sharpen, and blend names specify their target-view transforms; the remaining routes use the identity view.

Preference and tolerance. In the Portraits/CIFAR controls, candidate-name ranks are relaxed: 0, plain/smooth: 1, anchor/prototype: 2, with unlisted names assigned rank 1. Candidates within the configured tolerance of the best score are eligible; lower rank is preferred, and larger score breaks equal ranks. The order and tolerance are fixed within each run.

Clipart uses base tolerance 0.006, non-relaxed tolerance 0.015, and style-candidate tolerance 0.05. Its candidate-name ranks are grayscale_anchor: 0; sharpen_anchor: 1; blend_plain: 2; anchor/smooth: 3; grayscale_plain: 4; plain: 5; prototype: 6; relaxed: 8; lowpass_relaxed: 9; unlisted names: 1.

The Clipart eligibility reference starts at the maximum score. A best non-relaxed route within 0.015 of that maximum can lower the reference to its score; a best style route within 0.05 can lower it further. Style routes contain anchor, blend, grayscale, or sharpen in their names. Candidates within 0.006 of the resulting reference are eligible. When the style condition is met, the selector first considers eligible style routes within 0.006 of the best style score, then applies the configured rank and score tie-break.

Update budgets. The supplementary controls use the following settings:

Supplementary configurationCandidatesBatch sizeUpdates per candidate per stageTolerance
Portraits CNN3251,024300.0015
Portraits ResNet9673843600.004
CIFAR base library31281,1850.0015

The update counts are 3 × 5 × 2, 5 × 12 × 6, and 3 × 5 × 79, respectively: interpolation paces × epochs per pace × minibatches per epoch. Candidates within a setting receive the same number of optimizer updates. The Office-Home smoothing configuration uses batch size 64, three paces, and six epochs per pace, giving 18M updates per candidate, where M is the larger of the source and target minibatch counts.

4. How do the elements of Figure 2 map to the method? (W4)

Figure 2 connects three levels of the method: why route selection is needed, how a stage selects an update, and how selected updates form the adaptation trajectory.

Element in Figure 2Method correspondenceInterpretation
Fixed-route mismatch§3.1, motivation; §3.2, route-relaxed formulationMotivates allowing the supervision rule to change across stages.
Route-relaxed bridge-and-smooth stage§3.3, stage optimizationConstructs candidate supervision, performs candidate updates, and selects a route using the score, tolerance, and preference.
Ordered path optimization§3.4, overall workflow; §3.5, theoretical analysisPropagates the selected state through subsequent domains and analyzes the resulting adaptation trajectory.

Within a stage, candidate routes start from the same incoming classifier and prototype state. After selection, the chosen update determines the state passed to the next domain. Routes specify supervision rules and the updates they induce; prototypes provide class-level semantic references.

Thank you for your detailed responses and additional experiments. I still have two concerns:

  1. (W2) The new results show changes in supervision reliability across stages, but it remains unclear whether the selected routes actually reflect these changes and lead to better adaptation.

  2. (W1/W3) The improvements may partly come from manually defined preferences and other route-specific settings rather than adaptive routing itself. Notably, Table S8 shows that adaptive routing performs similarly to greedy selection and strong fixed routes. Also, the Office-Home settings in the response and the Office-31 experiments in the paper should be clearly distinguished in the revised manuscript.

Overall, the main weakness remains the lack of clear evidence that adaptive routing itself is responsible for the reported gains.

Reviewer_pNqtOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper studies gradual domain adaptation and introduces PrismBridge. At each stage, it selects a supervision rule that combines predictions from the frozen source model, source prototypes, and the current model. A tolerance-aware selector chooses among near-best candidates, and source-supervised smoothing stabilizes the update. The paper also provides theoretical analysis and experiments on several types of domain shift.

Critical strengths
  1. The motivation is clear and the method is easy to follow. Different supervision sources may be useful at different stages, so adapting their combination is a reasonable approach, and the paper illustrates this problem well in Figure 1.

  2. The experiments cover several types of domain shift and report strong results. The ablation studies also help examine the roles of different components.

Critical weaknesses
  1. The connection between the theory and the actual selector is not clear to me. Theorem 3.1 assumes that the selection score is close to the stage cost, but the paper does not show whether this assumption holds for the proposed score or how large the error may be. Theorem 3.2 gives a risk bound, but the experiments do not show how its terms change during adaptation or relate to accuracy. This makes it hard to see what the theory explains about the method's performance.

  2. The contribution of adaptive route selection is also unclear. Ablations in Table 4 show a drop of several percentage points when a fixed route is used, but the result may depend on which route is fixed. The additional fixed-route comparison is useful, although it only covers one benchmark. It is therefore unclear whether adaptive selection consistently improves over a well-chosen fixed rule across the main tasks.

Questions for authors
  1. Could you report some of these quantities on one benchmark, such as bridge displacement or retained-sample coverage, and examine how they relate to accuracy? An analysis using evaluation labels after adaptation would also be useful if needed.

  2. Could you compare adaptive selection with the best fixed route on at least one more main benchmark, such as Portraits or Covertype?

  3. How sensitive is PrismBridge to the order and availability of intermediate domains? For example, what would you expect if two nearby domains were swapped or one intermediate domain were missing?

3 comments
AuthorsOpenReview ↗

Response to Reviewer pNqt

Thank you for asking for direct mechanism measurements and a complete fixed-route comparison.

1. What do the stagewise quantities show? (W1, Q1)

The five-stage measurements on Portraits ResNet, averaged over three adaptation seeds, are:

StageBridge displacementRetained pseudo-label precision (%)Probe coverage (%)Post-update accuracy (%)
10.201498.7896.8598.02
20.238199.7196.8599.15
30.265398.6696.8597.40
40.300897.5096.8596.40
50.297491.3196.8591.15

From stage 2 to stage 5, bridge displacement changes from 0.2381 to 0.2974, retained pseudo-label precision from 99.71% to 91.31%, and post-stage accuracy from 99.15% to 91.15%. These measurements connect changes in bridge geometry and supervision quality to the observed accuracy variation.

Probe coverage remains 96.85% throughout, while actual candidate training coverage is 96.53–96.85%. The fixed probe coverage makes changes in the quality of retained supervision comparable at a common retention level.

At the final stage, PrismBridge's 91.15% accuracy is 5.65 points above the pre-update model (85.50%) and 7.80 points above the frozen source (83.35%), showing effective adaptation as the supervision signals change.

2. How does adaptive selection compare with the best fixed route on Portraits? (W2, Q2)

The supplementary Portraits comparison evaluates every fixed candidate over the ordered domain sequence, using three paired adaptation seeds. CNN32 uses the main Portraits model with a five-route controlled extension; ResNet96 provides a separate seven-route diagnostic. Fixed and adaptive controls share the same acceptance safeguard.

Portraits configurationAdaptive (%)Best fixed routeBest fixed (%)
CNN32, five routes81.00±0.10smooth / relaxed81.13±0.31
ResNet96, seven routes91.15±0.98plain91.20±0.18

By selecting routes during adaptation, PrismBridge avoids a separate search for a globally fixed supervision rule while achieving performance comparable to the best fixed route (91.15% versus 91.20% on Portraits ResNet).

3. What happens when an intermediate domain is swapped or omitted? (Q3)

GDA aims to bridge a large source–target distribution gap through gradual adaptation while managing stagewise error accumulation. Here, we retain the multi-stage adaptation structure and perturb the intermediate-domain path locally, keeping the source domain, final target, source model, and adaptation seeds fixed:

ConfigurationOrderedAdjacent swapOne middle domain omitted
CNN3281.00±0.1080.92±0.0380.12±1.58
ResNet9691.15±0.9890.70±0.3590.25±0.75

Across the two configurations, the largest observed mean accuracy changes are 0.45 points for swapping adjacent domains 2 and 3 and 0.90 points for omitting domain 3. The omission control uses four adaptation stages instead of five. The ordered, swapped, and omission paths are [0,1,2,3,4,5], [0,1,3,2,4,5], and [0,1,2,4,5], respectively.

This limited sensitivity is consistent with PrismBridge's stagewise supervision selection: it re-evaluates the available routes at each stage and selects a supervision rule according to the current domain and model state, providing flexibility to accommodate the tested local path changes.

Thank you for your detailed response, which has addressed most of my concerns.

I would suggest extend the comparison with fixed routes to probably 1-2 more benchmarks, which would help better clarify when adaptive routing is most useful.

I will increase my rating to 3.

We sincerely thank the reviewer for the positive feedback and generous decision to increase the rating. Your careful evaluation, valuable insights, and recognition of our efforts to address the concerns are deeply appreciated. Such encouragement is truly meaningful.

We are also grateful for the constructive suggestion to include additional fixed-route benchmarks. These comparisons would indeed help clarify when adaptive routing is most beneficial, and their inclusion will be carefully considered in the revision. Once again, thank you for your kind support and valuable guidance.

Reviewer_fTsGOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses Gradual Domain Adaptation (GDA), highlighting the limitations of existing methods that rely on fixed adaptation rules or predefined loss schedules despite the changing reliability of supervision sources across intermediate domains. To address this limitation, the authors propose PrismBridge, a route-relaxed framework that adaptively selects supervision strategies by combining predictions from a frozen source model, class prototypes, and the current adapted model. Specifically, the framework evaluates candidate routes using multiple learning objectives and employs tolerance-aware route selection with source-supervised smoothing to stabilize adaptation. Furthermore, the authors provide theoretical analyses of trajectory-dependent regret and target risk bounds, and empirically evaluate the effectiveness of the proposed framework across diverse domain shifts, including geometric, appearance, temporal, tabular, and corruption shifts.

Critical strengths
  1. The authors highlight the limitations of existing GDA methods that rely on fixed adaptation rules or predefined loss schedules. Formulating supervision selection as a route optimization problem provides a well-motivated perspective on GDA.

  2. The proposed framework effectively combines frozen source predictions, class prototypes, and current-model predictions. Tolerance-aware route selection and source-supervised smoothing further support stable adaptation across stages.

  3. The authors provide theoretical analyses of selection regret and cumulative target-risk bounds, accounting for the effects of route decisions on future model trajectories.

Critical weaknesses
  1. Weak connection between theory and practical route selection. Although the theoretical formulation considers future adaptation costs, the actual route selection relies on a local loss-based score (Eq. 10). The assumption that this score approximates the theoretical transition cost lacks sufficient justification.

  2. Limited reliability of route scoring. The proposed score compares losses computed using different pseudo-labels and confidence-based sample masks. Since sample coverage is not explicitly considered, routes excluding difficult samples may receive favorable scores without necessarily achieving better adaptation.

  3. Insufficient clarity in experimental comparisons. Despite substantial improvements on CIFAR-C, it remains unclear whether all baselines are evaluated under comparable conditions regarding source-data access, intermediate-domain information, and computational budgets.

Questions for authors
  1. The theoretical analysis assumes that the practical route score approximates the transition cost. Could the authors justify this assumption and provide empirical evidence supporting the reliability of the proposed score?

  2. How are the predetermined route preference order and stage-dependent tolerance determined in practice?

  3. Could the authors clarify the source-model configurations, data splits, and adaptation/evaluation protocols, particularly for CIFAR-C, where PrismBridge substantially outperforms existing methods such as STDW? Additionally, a more detailed explanation of the factors contributing to these large performance improvements would be helpful.

  4. Given the multiple candidate evaluations and labeled source replay, are the baselines evaluated under comparable data-access and computational budgets? Could the authors provide compute-matched comparisons and analyze the overhead as the number of routes increases?

  5. Figure 1 does not clearly illustrate the distinction between fixed-route adaptation and the proposed adaptive route selection. In Fig. 1(b), prototype anchors (black stars) appear both near feature distributions and along route trajectories, making their interpretation unclear. The authors should clarify what these positions represent and how the illustrated routes correspond to the actual supervision-selection mechanism.

7 comments
AuthorsOpenReview ↗

Response to Reviewer fTsG (1)

Thank you for the detailed questions. We explain the existing design and provide the supplementary controls below.

1. What connects the score to the theory, and can masking favor a route? (W1, W2, Q1)

Eq. (10) provides a tractable stagewise surrogate for comparing supervision routes. Its bridge, smoothing, and consensus terms address supervision fit, the model update, and supervision consistency, respectively. Earlier route decisions determine the classifier and prototype state at which later candidates are evaluated, connecting each local comparison to the induced adaptation trajectory.

The theoretical analysis makes the score-to-cost approximation error explicit in the selection-regret guarantee. The score components motivate the practical criterion; the approximation condition specifies the link required by the guarantee. This explains the respective roles of the practical score and the trajectory-dependent analysis.

The supplementary coverage analysis reports actual training-mask fractions separately from the identity-view diagnostic probe. In the ResNet diagnostic, candidate-by-stage mean training coverage ranges from 96.53% to 96.85%, with every candidate and stage included. Evaluation labels are used only to measure accuracy and pseudo-label precision after updates. These measurements quantify supervision quality and retained-sample coverage.

2. How are preference and tolerance determined? (Q2)

Preference and tolerance are predeclared configuration choices. Candidates within the tolerance of the maximum score are eligible. The deterministic priority is relaxed, then plain/smooth, then anchor/prototype; other route names have the middle priority. Larger score breaks equal priorities. The preference favors the relaxed update when utilities are close.

The supplementary CNN/CIFAR controls use tolerance 0.0015, and the ResNet diagnostic uses 0.004, held constant within each run. The stage-indexed notation permits this constant setting. These values were retained from the existing configurations without tuning on target evaluation labels. The combined tolerance/preference rule attains 91.15±0.98% on the seven-route Portraits diagnostic, while the additional greedy-selection experiment reports 79.13±0.12%. The reported 12.02-point accuracy advantage highlights the value of tolerance-aware preference over greedy selection.

3. What are the CIFAR protocol and the factors behind the STDW comparison? (W3, Q3)

The supplementary controlled comparison uses the same Standard WRN-28-10 source model across methods. Source prototypes use CIFAR-10 TRAIN[0:10000]. Adaptation traverses severities 1–5 of the 10,000 corrupted TEST images, with intermediate/target training labels replaced by −1; final evaluation is at severity 5. Source-label CE during adaptation is disabled for both PrismBridge and native STDW in this control.

MethodGaussian noiseDefocus blurContrastMean accuracy (%)
PrismBridge, adaptive control79.5289.9789.7586.41
Fixed plain, same control framework78.7683.1480.1780.69
STDW, native budget78.9282.8174.8278.85

Across the three corruptions, PrismBridge achieves 86.41% mean accuracy, exceeding the same-framework fixed-plain control (80.69%) by 5.72 percentage points and native-budget STDW (78.85%) by 7.56 points. The consistent gains across all three corruptions, including 9.58 points over fixed plain on Contrast, highlight the value of stagewise route selection.

AuthorsOpenReview ↗

Response to Reviewer fTsG (2)

4. Are data access and compute comparable, and how does cost scale with routes? (W3, Q4)

We compare the Portraits ResNet controls on the same RTX 4090. PrismBridge uses the initial labeled source for supervised smoothing. Alongside native-budget STDW, we evaluate a stage-time-budget variant and a stage-time-budget variant augmented with initial-source CE weight 0.2. The time-budget controls allocate time across all native interpolation values at the stated learning rate.

MethodAccuracy (%)Adaptation time (s)
PrismBridge91.15±0.98958.10
STDW, native budget59.75±0.33116.30
STDW, stage-time budget59.53±15.28951.42
STDW + source CE, stage-time budget88.28±1.02951.54

PrismBridge retains a 2.87-point mean accuracy advantage over the time-budget-matched STDW control with labeled source replay.

The separate Portraits ResNet route-count profile is:

Number of routesAdaptation time (s)Optimizer updatesPeak allocated memory (MiB)
1158.51,8003,931.9
3389.75,4003,930.9
5648.89,0005,187.3
7947.312,6005,279.4

The profiles use declared candidate prefixes; different views have different costs. Peak memory includes the experiment runner's retained frozen source model. After adaptation, inference uses one retained backbone, so candidate-search overhead is paid during adaptation.

5. What do the stars and routes in Figure 1 represent? (Q5)

The intended distinction is between prototype references in feature geometry and choices of supervision rule. The prototype anchors referred to in the question are class-level semantic references for bridge construction. A route specifies how the available supervision is used to produce a candidate update; it is not a sequence of prototype locations.

A fixed route repeats one supervision family across stages. Adaptive selection can choose a different family as the domain and current model change. The route trajectories illustrate this adaptation process schematically. Prototype symbols serve the geometric explanation, while route labels and arrows describe the update choices. The operational correspondence is candidate construction, scoring, tolerance/preference selection, and propagation of the retained model to the next stage.

Thank you for your detailed response and for providing additional experimental results. Your explanations have helped me better understand the proposed Route Selection mechanism and the experimental setup. However, I would like to request further clarification regarding the composition of the Route Library, the rationale behind the Preference Order, and the actual Route Selection behavior during adaptation.

1. Detailed Composition of the Route Library

The specific Mixture Weights, View Sets, and Consensus Rules for each Route Candidate used in the experiments do not appear to be explicitly provided. Could you clarify the exact configurations of the Route Candidates used in the experiments and explain the criteria used to construct the candidate library?

2. Rationale Behind the Preference Order

In the rebuttal, you explained that the Preference Order is predefined as Relaxed → Plain/Smooth → Other Routes → Anchor/Prototype, with Relaxed Updates being preferred when the Scores are similar. However, the specific rationale for assigning the highest priority to Relaxed Routes and lower priority to Anchor/Prototype Routes remains unclear. Could you clarify whether this ordering is motivated by theoretical analysis or empirical observations, or whether it is a predefined heuristic?

3. Actual Route Selection Behavior

The paper emphasizes that the reliability of different Supervision Sources changes as Domain Shift progresses, making it important to select appropriate Routes accordingly. However, the paper does not appear to provide a detailed analysis of how frequently each Route is selected during adaptation or how the selection distribution changes across Stages.

Could you provide the overall selection frequency of each Route and the Stage-wise distribution of selected Routes? In particular, given that Relaxed Routes are assigned the highest Preference, it would be helpful to understand whether the method actually selects diverse Routes or whether a particular Route dominates across most Stages.

Such an analysis would provide important evidence for understanding whether the proposed method genuinely adapts its Supervision Strategy in response to Domain Shift and how these adaptive changes contribute to performance improvements.

Overall, I believe that clarifying the detailed composition of the Route Library, the rationale behind the Preference Order, and the actual Route Selection behavior would further strengthen the validity and reproducibility of the proposed method, while providing more convincing evidence for the central contribution of Adaptive Route Selection.

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

Related papers

Discussion (0)

Sign in to comment.