acceptodds
Under review as a conference paper at ICLR 2027

HeteroWeave: Variable-Cardinality Block Composition in Pretrained Model Reassembly

Abstract

Existing heterogeneous pretrained model reassembly methods usually retain one block at each aligned position, limiting the potential for different pretrained sources to contribute jointly. Allowing multiple blocks to participate at the same position increases reassembly flexibility, but turns block identity and participation cardinality into coupled architectural decisions, substantially expanding the composition space. Meanwhile, additional parameter count and computation do not translate monotonically into performance gains, creating explicit tradeoffs between performance and resource cost. To address these challenges, we propose HeteroWeave to reformulate heterogeneous pretrained model reuse as a variable-cardinality, multi-objective reassembly problem that explicitly accounts for performance–resource trade-offs. It jointly determines block identity and participation cardinality within a unified composition space, while shared interfaces and lightweight adaptation make heterogeneous compositions executable. To avoid training every candidate, we design a low-cost proxy that estimates performance from block-level activation-pattern statistics, using square-root aggregation to reduce score inflation as the activation space expands. HeteroWeave then optimizes the proxy, parameter count, and FLOPs as three objectives via Pareto-guided evolutionary search, retaining diverse nondominated architectures for different training and inference preferences. We validate HeteroWeave across four tasks, covering image classification, vision-language retrieval, semantic segmentation, and object detection. On classification, one nondominated representative improves accuracy by 1.08 percentage points over the single-block reassembly baseline, while using 26.7% fewer parameters and 11.9% fewer FLOPs.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Pretrained Model Reassembly, Heterogeneous Model Reuse, Variable-Cardinality Composition, Multi-Objective Architecture Search, Performance–Resource Trade-off

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_5LorOpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

HeteroWeave constructs models by selecting one or more pretrained blocks at each aligned position. Adapters reconcile incompatible interfaces, and selected block outputs are fused. A training-free activation score, CLAS, ranks candidate architectures, while a multi-objective evolutionary search balances this score against parameter count and FLOPs. The paper evaluates ImageNet classification and additional compositions for retrieval, segmentation, and detection.

Critical strengths
  1. The main ImageNet result improves accuracy and resource use together. The medium-size model has good gains over DeRy while reducing parameters and FLOPs significantly.

  2. The capacity-matched experiment directly examines whether parallel composition helps beyond adding parameters. The parallel extension outperforms the serial extension under nearly identical resource costs. The cross-task results also show that adding more blocks is not uniformly beneficial, which supports treating participation count as a search variable.

Critical weaknesses
  1. CLAS is validated against accuracy after only one or five training epochs. However, the paper itself shows that early performance can differ from the final ranking. The partial correlations and seed-stability checks are useful, but they do not establish that CLAS reliably identifies the best architectures after the full 100-epoch training schedule.

  2. The contribution of variable cardinality is not fully separated from the rest of the search pipeline. Comparing against DeRy changes more than the allowed number of blocks, while the capacity-matched experiment tests particular extensions rather than the searched architecture space. A single-block version using the same CLAS score, adapters, search budget, and selection rule would provide a cleaner test.

Questions for authors

Following form the two weaknesses:

  1. How well does CLAS correlate with fully trained accuracy on a representative candidate subset? Does square-root aggregation improve this correlation relative to the same score without that transformation?

  2. What happens when the complete search is repeated with cardinality restricted to one at every position, keeping the source pool, resource limits, evaluation budget, and selection rule fixed?

3 comments
AuthorsOpenReview ↗

Response to W1–W2 and Q1–Q2: CLAS Ranking and Variable Cardinality

W1 / Q1 — CLAS ranking and square-root ablation

Thank you for these thoughtful questions. We evaluate all proxies against 100-epoch ImageNet-1K results of 20 reassembled architectures, including CLAS without the square root. All proxies use the models’ initial states.

Table C.1: Spearman correlation with fully trained Top-1 accuracy. Values are means ± sample SD over three proxy-scoring seeds.

ProxySpearman ρ ↑
SynFlow−0.11 ± 0.00
Fisher0.33 ± 0.08
SNIP0.40 ± 0.03
GradNorm0.43 ± 0.03
ZiCo0.59 ± 0.01
NASWOT0.68 ± 0.01
Jacov0.75 ± 0.01
CLAS (Ours)0.81 ± 0.02
CLAS (no square root)0.71 ± 0.01

(a) Final-performance ranking. CLAS achieves the highest mean Spearman correlation, supporting its ranking utility beyond short training.

(b) Square-root benefit. Removing only the square root reduces correlation from 0.81 ± 0.02 to 0.71 ± 0.01. The transformation assigns diminishing credit to additional patterns within each module, tempering count-driven score growth.

Revision. We will add these results to Section 4.2.2 (Figure 3a) and Appendix B.1/B.3 (Table 7), and clarify aggregation in Section 3.3.2.

W2 / Q2 — Contribution of variable cardinality

We thank the reviewer for this suggestion. We add a complete-search single-block baseline to assess the contribution of variable cardinality.

We fix block cardinality to one at every position while retaining block-identity search. The source pool, CLAS score, adapters, 30M-parameter/6G-FLOP limits, total search budget, and selection rule remain identical. Both searches use NSGA-III with a population of 96 for 160 generations. We select the highest-CLAS composition from each feasible Pareto front, then fully fine-tune both selected architectures on ImageNet-1K for 100 epochs using the same recipe and training seed 0.

Table C.2: Complete-search single-block comparison under the 30M-parameter/6G-FLOP limits.

CardinalityParams (M) ↓FLOPs (G) ↓Top-1 (%) ↑
Fixed at 117.985.6678.46
Variable >=118.263.9579.51

Variable cardinality improves the searched architecture. Under matched search and training settings, it achieves 1.05 percentage points higher Top-1, with 30.3% fewer FLOPs and 1.5% more parameters.

Revision. We will add this comparison and its protocol to Section 4.2.1 (Table 4).

Thank you again for your thoughtful suggestions. The additional comparisons have made our evaluation more complete.

Reviewer_5LorOpenReview ↗

Response to Rebuttal

Thanks for doing these tests. The results after 100 epochs answer my question about CLAS and the square root. The single-block test also shows that allowing more blocks helps. I have increased my score to 4.

Dear Reviewer 5Lor,

Thank you for considering our rebuttal and updating your score. We are glad that the additional experiments addressed your concerns. Your constructive feedback has helped us strengthen the evaluation.

Thank you again for your time.

Reviewer_5WChOpenReview ↗
Rating
4(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper considers the problem of variable-cardinality reassembly in pre-trained models. Specifically, the paper focuses on jointly determining block identity and participation cardinality at each block position. The authors introduce a new performance proxy which accounts for scale induced bias (arising uniquely in the variable cardinality scenario) and propose a Pareto-guided search over the proxy score, parameters and FLOPs. Their framework, HeteroWeave, is able to accommodate heterogeneous source architectures via a lightweight interface adaptation mechanism. Evaluations across four tasks demonstrate non-monotonic improvements with cardinality, emphasizing the importance of search rather than participation maximization. On classification, HeteroWeave yields reassembled architectures that simultaneously improve accuracy, while lowering FLOPs and parameter count compared to the single-block assembly baseline.

Critical strengths

I would like to commend the authors for the excellent exposition and overall presentation of the paper. It was a joy to read the paper, and I found each section, figure, table, to have been made with extra attention to the detail.

Formally, all the components including block library and composition encoding, interface adaptation and fusion, as well as multi-objective search are very clearly described. The formulation is sound, while the experiments reveal nuances of such a difficult combinatorial search.

The study reveals several valuable insights, including how higher cardinality does not always translate into performance gains, how composition effects transfer to different task types, and the effectiveness of CLAS as a reliable ranking signal.

Lastly, the study incorporates heterogeneous architectures, and joint-search over params and FLOPs, which widens the applicability of HeteroWeave. Given these overall strengths, I recommend acceptance.

Critical weaknesses

It is slightly unclear why higher score on binary response patterns should act as a good performance proxy. An intuition behind this score is missing.

Questions for authors
  1. How does HeteroWeave compare/relate to similar developments in adjacent domains such as FlexRank [1] and FlexMerge [2] which also consider model reassembly?

References

[1] Zaccone, R., Laskaridis, S., Ciccone, M. and Horváth, S. (2026) FlexRank: Nested Low-Rank Knowledge Decomposition for Adaptive Model Deployment, Proceedings of the 43rd International Conference on Machine Learning, Proceedings of Machine Learning Research, 306, pp. 152741–152765.

[2] Dhasade, A., Jhunjhunwala, D., Vujasinovic, M., Joshi, G. and Kermarrec, A.M., 2026, April. Navigating the accuracy-size trade-off with flexible model merging. In International Conference on Learning Representations (Vol. 2026, pp. 81010-81041).

3 comments
AuthorsOpenReview ↗

Response to W1 and Q1: CLAS Intuition and Comparison with FlexRank and FlexMerge

W1 — CLAS Intuition

Thank you for raising this point. We explain how binary activation patterns capture response diversity.

(a) How are binary activation patterns constructed? CLAS encodes positive activations as 1 and non-positive activations as 0, forming a response pattern across images (Table B.1).

Table B.1. Illustrative example of binary activation patterns.

FeatureImage AImage BImage CImage DBinary pattern
Feature 12.0 → 10.8 → 1−0.3 → 0−1.2 → 01100
Feature 20.4 → 13.1 → 1−0.7 → 0−0.2 → 01100
Feature 31.5 → 1−0.6 → 00.9 → 1−0.4 → 01010

(b) Why count distinct patterns? Features 1 and 2 share 1100, while Feature 3 adds 1010, yielding two distinct patterns. Further 1100 features add no new pattern.

(c) Why might a higher score predict better performance? More distinct responses may provide downstream learning with richer cues for distinguishing inputs. Repeated patterns add no new distinctions under the binary encoding.

Revision. We will incorporate this explanation into Section 3.3.2 (Low-Cost Score Estimation).

Q1 — Comparison with FlexRank and FlexMerge

Thank you for bringing up these relevant works. They share the goal of reusing pretrained computation to meet different performance and deployment-resource requirements. Table B.2 compares how they reuse pretrained computation to meet performance and resource requirements.

Table B.2. Comparison with FlexRank and FlexMerge.

MethodPool sizeHeterogeneous sourcesOperation unitVariable block cardinalitySearchOptimization goals
FlexRank [1]1NoLow-rank factorsNoDPPerformance ↑ under parameter budgets
FlexMerge [2]MNoBlock task vectorsNoGreedySimilarity ↑ with a target storage size
HeteroWeave (ours)MYesPretrained blocksYesNSGA-IIIPerformance ↑, Params. ↓, FLOPs ↓

[1] Zaccone, R., Laskaridis, S., Ciccone, M. and Horváth, S. (2026) FlexRank: Nested Low-Rank Knowledge Decomposition for Adaptive Model Deployment, Proceedings of the 43rd International Conference on Machine Learning, Proceedings of Machine Learning Research, 306, pp. 152741–152765.

[2] Dhasade, A., Jhunjhunwala, D., Vujasinovic, M., Joshi, G. and Kermarrec, A.M., 2026, April. Navigating the accuracy-size trade-off with flexible model merging. In International Conference on Learning Representations (Vol. 2026, pp. 81010-81041).

(a) Heterogeneous sources and variable cardinality. FlexRank extracts nested low-rank submodels from a single pretrained model, while FlexMerge merges block-level task vectors from same-architecture task models. HeteroWeave operates on a heterogeneous model pool, jointly searching block identity and participation cardinality at each position, allowing multiple heterogeneous pretrained blocks to contribute jointly.

(b) Explicit three-objective joint search. FlexRank searches low-rank configurations under budget constraints, while FlexMerge selects block groups by task-vector similarity to reach a target deployment size. Neither method explicitly treats performance, parameter count, and FLOPs simultaneously as three independent search objectives. HeteroWeave jointly optimizes these objectives through NSGA-III to obtain nondominated architectures.

(c) Multi-budget deployment. FlexRank selects submodels with different rank configurations from trained shared weights. FlexMerge obtains block-sharing configurations along its merging sequence to meet a target storage size. HeteroWeave selects architectures satisfying parameter and FLOP budgets from the Pareto candidate set produced by a single search.

Revision. We will add FlexRank and FlexMerge to Table 1 in Section 2, “Related Work,” and clarify these distinctions in the accompanying discussion.

We appreciate the opportunity to clarify these points and hope that our responses address the reviewer’s concerns.

I thank the authors for providing the intuition behind observing binary activation patterns as performance proxy. The discussion of FlexMerge and FlexRank also clearly delineates the differences and similarities.

After reading the concerns from other reviewers on CLAS score and the authors' responses to it, I am more convinced about its benefits. In particular, the authors' responses adequately show the improvement due to square root and that the ranking utility holds beyond short training.

Thus, I will raise my score. All the best!

Dear Reviewer 5WCh,

Thank you for your thoughtful feedback and for updating your score. We are glad that our clarifications and additional experiments addressed your concerns. If you have any further questions or suggestions, we would be happy to respond.

Thank you again for your time and constructive review.

Reviewer_AHNyOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

HeteroWeave searches for combinations of heterogeneous pretrained blocks, allowing multiple blocks at each position. It uses an activation-based proxy and multi-objective search to balance accuracy, parameters, and FLOPs.

Critical strengths

1, The ImageNet results show a promising accuracy–resource trade-off, with consistent gains across three seeds.

2, The capacity-matched serial versus parallel comparison is useful for examining the role of composition.

Critical weaknesses

1, The core experimental setup is hard to follow. The main text does not clearly list the source models, pretrained checkpoints, or selected block compositions. These details are central to understanding what is being reused and where the gains come from.

2, The justification for CLAS needs stronger evidence. Taking a square root within each module may reduce score growth, but summing over more modules could still favor larger compositions. Also, the ranking correlations are measured against short-training accuracy rather than final performance.

3, The practical benefit is not fully established. The selected model still requires substantial training. Comparisons with directly adapting suitable pretrained models, including total search and training cost and actual inference latency, would make the value clearer.

4,The broader task evidence is mixed. Detection gains are negligible relative to the reported variation. Retrieval uses frozen CLIP as the baseline, so it is difficult to separate the benefit of composition from additional adaptation.

Questions for authors

See above

3 comments
AuthorsOpenReview ↗

Response to W1–W2: Experimental Setup and CLAS Ranking

W1 — Experimental setup

Thank you for this helpful suggestion. The main text points to Appendix B. We agree that a clearer overview would help readers follow the setup. Table A.1 summarizes the main model in Table 2 under the 30M-parameter/6G-FLOP limits.

Table A.1. Experimental setup.

ItemDetailsConfiguration
Source modelsSWSL ResNeXt-50, ResNet-50, ViT-S/16.Architecture
Pretrained checkpointsSWSL ResNeXt-50 release, torchvision ResNet-50 checkpoint, and AugReg ViT-S/16 pretrained on ImageNet-21K and fine-tuned on ImageNet-1K.Checkpoints
Block compositionSWSL ResNeXt-50 → ResNet-50 → (ResNet-50 + ViT-S/16) → ViT-S/16. Cardinalities [1,1,2,1].Architecture
Training100-epoch full fine-tuning, AdamW, learning rate 0.001, weight decay 0.05, effective batch 1,024, seeds 0/11/23.Training

“+” denotes joint participation at the same position. Appendix B.1–B.2, Tables 6–7, and the repository's Paper model inventory and reproducibility guides cover the other compositions and reproduction procedures. Repository

Revision. We will summarize the setup in Section 4.1.1, with details in Appendix B.1–B.2 (Tables 5–6).

W2 — CLAS ranking

Thank you for raising these important concerns. We provide evidence on size sensitivity and final-performance ranking.

(a) Ranking beyond model size

We examine reassembled architectures from multi-objective and random search, each trained for 100 epochs on ImageNet-1K. To assess whether CLAS distinguishes architectures with similar resource costs, we exhaustively match pairs whose parameter counts and FLOPs each differ by at most 5%, using only resource measurements for matching. Two pairs qualify.

Table A.2. Resources, CLAS, and final accuracy. Top-1 denotes validation accuracy after 100 epochs on ImageNet-1K.

PairArchitectureParams (M) ↓FLOPs (G) ↓CLAS (×10³) ↑Top-1 (%) ↑
1A118.513.839.22 ± 0.1778.41
1A218.263.9513.26 ± 0.3879.46
2A327.943.908.79 ± 0.0777.54
2A427.333.9710.19 ± 0.2178.69
  • CLAS distinguishes architectures with similar resource costs. Higher CLAS corresponds to 1.05 and 1.15 percentage points higher final Top-1 within the two pairs.
  • Higher CLAS and accuracy at lower resource costs. A1 and A2 achieve higher CLAS and final Top-1 than A3 and A4, respectively, with fewer parameters and FLOPs.
  • The square root improves ranking. The square root assigns diminishing credit to additional patterns within each module. Retaining it improves final-accuracy correlation from 0.71 ± 0.01 to 0.81 ± 0.02 across the 20 architectures (Table A.3).

(b) Final-performance ranking

We evaluate all proxies against the completed 100-epoch ImageNet-1K results of 20 reassembled architectures, including a CLAS variant without the square root.

Table A.3. Spearman correlation with final Top-1.

ProxySpearman ρ ↑
SynFlow−0.11 ± 0.00
Fisher0.33 ± 0.08
SNIP0.40 ± 0.03
GradNorm0.43 ± 0.03
ZiCo0.59 ± 0.01
NASWOT0.68 ± 0.01
Jacov0.75 ± 0.01
CLAS (Ours)0.81 ± 0.02
CLAS (no square root)0.71 ± 0.01

Final ranking. CLAS achieves the highest mean Spearman correlation among the evaluated proxies, supporting its ranking utility beyond short training.

Revision. We will add these comparisons to Section 4.2.2 (Figure 3a–b) and Appendix B.1/B.3 (Table 7), and clarify aggregation in Section 3.3.2.

Thank you again for your thoughtful comments, which helped us clarify the experimental setup and strengthen the evidence for CLAS.

AuthorsOpenReview ↗

Response to W3–W4: Practical Value and Cross-Task Evidence

W3 — Practical value

Thank you for raising this important question. We agree that the practical value of reassembly should be assessed alongside its search, fine-tuning, and deployment costs.

Table A.4, provided in a separate comment titled “Table A.4 — Practical Value Comparison,” compares HeteroWeave with direct adaptation of ResNet-50 and RegNetY-3.2GF, and includes DeRy [1], a key reassembly baseline that also applies end-to-end full fine-tuning to reassembled models.

On upstream ImageNet-1K, HeteroWeave fully fine-tunes its pretrained-block composition for 100 epochs. ResNet-50 and RegNet use existing ImageNet-trained weights with zero incremental upstream training cost. On the six downstream tasks, we reuse the same ImageNet-trained composition and follow the backbone-frozen protocol in Appendix B.1 for all three models.

(a) Better upstream trade-offs. HeteroWeave achieves 79.67% Top-1, exceeding DeRy, ResNet-50, and RegNet by 1.30, 2.98, and 0.15 percentage points, respectively, with the fewest parameters.

(b) Higher mean transfer accuracy. Across six downstream tasks, HeteroWeave achieves 86.80% mean Top-1, outperforming direct adaptation of RegNet and ResNet-50 by 2.86 and 4.45 percentage points, respectively.

(c) Competitive deployment costs. Compared with RegNet, HeteroWeave uses fewer parameters and reduces mean downstream latency by approximately 24%. Compared with ResNet-50, it achieves higher accuracy and fewer parameters at slightly higher latency.

(d) Build once, reuse across tasks. The composition serves all six downstream tasks, spreading its one-time search and upstream fine-tuning costs across applications.

(e) Greater architectural flexibility. Direct fine-tuning typically preserves backbone parameter count and FLOPs. HeteroWeave additionally searches block identities and cardinalities to meet accuracy and resource requirements.

Revision. We will add these comparisons to Section 4.3 (Figure 4) and Appendix B.3 (Table 10), with protocols in Appendix B.1/B.3.

[1] Yang et al. Deep Model Reassembly. NeurIPS, 2022.

W4 — Cross-task evidence

Thank you for highlighting the need for stronger cross-task evidence.

Beyond the original frozen-block setting, we fully fine-tune both baselines and HeteroWeave under the same protocol within each task for a fair comparison.

(a) Detection comparison

Table A.5: Detection performance under full fine-tuning on PASCAL VOC 2007.

ModelSeed 0Seed 1Seed 2Mean ± SD
Faster R-CNN (Baseline)73.3275.2574.2674.28 ± 0.96
HeteroWeave-D174.9275.3374.5674.93 ± 0.39
HeteroWeave-D275.6975.9675.3875.68 ± 0.29
HeteroWeave-D374.8775.8875.2175.32 ± 0.51

D1–D3 outperform the baseline in every paired seed, with an average gain of 1.03 points, indicating larger and directionally consistent detection gains from composition.

(b) Retrieval comparison

Table A.6: Retrieval performance under full fine-tuning on Flickr8K.

ModelSeed 0Seed 1Seed 2Mean ± SD
CLIP (Baseline)89.3889.7089.6489.57 ± 0.17
HeteroWeave-R189.6089.8689.9289.79 ± 0.17
HeteroWeave-R289.7889.8089.9189.83 ± 0.07
HeteroWeave-R389.7889.7689.8289.79 ± 0.03
  • Across seeds. All three HeteroWeave compositions outperform CLIP in every paired seed.
  • Mean gains. HeteroWeave improves performance by 0.21–0.25 percentage points over fully fine-tuned CLIP.

Revision. We will update Section 4.1.2 (Table 3) with these results and Appendix B.1 with the protocols.

We sincerely thank the reviewer again for the constructive feedback.

AuthorsOpenReview ↗

Table A.4 — Practical Value Comparison

Table A.4: Accuracy, resources, inference latency and computational costs across seven tasks.

TaskModelTop-1 (%) ↑Params (M) ↓FLOPs (G) ↓Latency (ms/batch) ↓Search (GPU-h)Train (GPU-h)
Upstream
ImageNet-1KResNet-5076.6925.564.1210.70 ± 0.0300
RegNetY-3.2GF79.5219.443.2121.74 ± 2.2200
DeRy [1]78.3724.894.4823.68 ± 2.075.5285.05
HeteroWeave (Ours)79.6718.263.9515.26 ± 0.912.28107.00
Downstream
Caltech-101ResNet-5092.6223.724.1210.62 ± 0.03011.03
RegNetY-3.2GF93.3718.083.2118.40 ± 0.11011.04
HeteroWeave (Ours)92.9817.913.9513.98 ± 0.0509.52
CIFAR-10ResNet-5089.3623.534.1210.68 ± 0.0109.87
RegNetY-3.2GF90.5517.943.2118.52 ± 0.0409.91
HeteroWeave (Ours)97.3517.883.9414.05 ± 0.0108.24
CIFAR-100ResNet-5070.3023.714.1210.69 ± 0.0109.97
RegNetY-3.2GF74.1018.073.2118.64 ± 0.30010.02
HeteroWeave (Ours)85.2017.913.9514.05 ± 0.0108.18
CUB-200-2011ResNet-5075.3723.924.1210.68 ± 0.01015.22
RegNetY-3.2GF79.1718.233.2118.52 ± 0.05015.24
HeteroWeave (Ours)78.9617.953.9514.04 ± 0.01013.04
DTDResNet-5073.6223.604.1210.68 ± 0.00015.88
RegNetY-3.2GF73.4017.993.2118.49 ± 0.02015.89
HeteroWeave (Ours)75.9017.893.9414.04 ± 0.00013.89
Oxford-IIIT PetsResNet-5092.8023.584.1210.68 ± 0.01013.81
RegNetY-3.2GF93.0217.983.2118.50 ± 0.05013.79
HeteroWeave (Ours)90.4117.893.9414.04 ± 0.01015.21
Mean (7 tasks)
ResNet-5081.5423.954.1210.68010.83
RegNetY-3.2GF83.3118.253.2118.97010.84
HeteroWeave (Ours)85.7817.963.9514.210.3325.01

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Rebuttal and Manuscript Revisions

Dear Area Chair and Reviewers,

Thank you for your time and helpful suggestions. Below, we summarize the key improvements from the review period, with the corresponding locations in the revised manuscript.

1. CLAS validation against fully trained performance (Reviewers AHNy and 5Lor). Across 20 architectures trained for 100 epochs on ImageNet-1K, CLAS achieves the highest mean Spearman correlation among the evaluated proxies (0.81 ± 0.02). Removing square-root aggregation reduces this to 0.71 ± 0.01. We added these results and resource-matched comparisons to Section 4.2.2, with protocols and scoring-seed results in Appendices B.1/B.3, Table 7.

2. Controlled evaluation of variable cardinality (Reviewer 5Lor). We added a complete-search single-block baseline, retaining the same source pool, proxy, adapters, resource limits, search budget, and selection rule. Under matched 100-epoch training with seed 0, variable cardinality improves Top-1 from 78.46% to 79.51%, with 30.3% fewer FLOPs and 1.5% more parameters (Section 4.2.1, Table 4).

3. Practical value and computational costs (Reviewer AHNy). We expanded comparisons with direct adaptation, reporting accuracy, parameters, FLOPs, measured latency, and construction costs. Reusing one composition across six downstream datasets yields 86.80% mean accuracy, exceeding RegNet by 2.86 points with approximately 24% lower mean latency (Section 4.3, Appendix B.3, Table 10).

4. Fairer cross-task comparisons and clearer exposition (Reviewers AHNy and 5WCh). Following AHNy’s suggestions, we updated retrieval and detection with matched full fine-tuning across three seeds (Section 4.1.2, Table 3; Appendix B.1) and clarified source checkpoints and compositions (Section 4.1.1; Appendices B.1–B.2). Following 5WCh’s suggestions, we clarified CLAS intuition (Section 3.3.2) and distinctions from FlexRank and FlexMerge (Section 2, Table 1).

Reviewers 5WCh and 5Lor confirmed that the additional evidence addressed their concerns and raised their ratings to clear acceptance. We appreciate the discussion and welcome any remaining questions.

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.