- 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
- 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.
- 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.
- 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.
- 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
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 transfer | Frozen source | Prototype | Current model |
|---|---|---|---|
| Art/Real-World → Clipart | 0.35 | 0.35 | 0.30 |
| Product → Clipart | 0.20 | 0.45 | 0.35 |
These weights apply before class-balance correction. The supplementary Portraits/CIFAR route execution settings are:
| Configuration | Candidate routes | Teacher-KL coefficient | Prototype-loss coefficient |
|---|---|---|---|
| Portraits CNN32 | smooth, relaxed, lowpass_plain, lowpass_relaxed, sharpen | 0 for all routes | 0 |
| Portraits ResNet96 | anchor, plain, relaxed, lowpass_plain, lowpass_relaxed, sharpen_anchor, blend_plain | 0.08 for anchor/sharpen_anchor; 0 otherwise | 0.03 |
| CIFAR base library | anchor, plain, relaxed | 0.12 for anchor; 0 otherwise | 0 |
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 configuration | Candidates | Batch size | Updates per candidate per stage | Tolerance |
|---|---|---|---|---|
| Portraits CNN32 | 5 | 1,024 | 30 | 0.0015 |
| Portraits ResNet96 | 7 | 384 | 360 | 0.004 |
| CIFAR base library | 3 | 128 | 1,185 | 0.0015 |
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 2 | Method correspondence | Interpretation |
|---|---|---|
| Fixed-route mismatch | §3.1, motivation; §3.2, route-relaxed formulation | Motivates allowing the supervision rule to change across stages. |
| Route-relaxed bridge-and-smooth stage | §3.3, stage optimization | Constructs 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 analysis | Propagates 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:
-
(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.
-
(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.
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.
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.