acceptodds
Under review as a conference paper at ICLR 2027

Complementary Learnware Reuse via Distribution Filling

Abstract

The growing availability of AI models calls for systems that can organize and combine their capabilities for new tasks. The learnware paradigm supports this goal by pairing each model with a specification describing its capabilities, enabling identification and reuse without sharing the training data of developers. Since models are developed for different tasks and conditions, no single learnware may adequately serve a new user: each may cover only part of the required input distribution, together with regions irrelevant to the task. When different learnwares cover complementary parts, combining their useful portions can enable the system to solve tasks that no single learnware can adequately handle. The challenge is to determine how learnwares can jointly serve different parts of the user distribution while reducing the influence of irrelevant regions. By modeling learnware utility at individual user inputs, we characterize how well a learnware set serves the user distribution and where an additional learnware can fill gaps in its coverage. This perspective motivates distribution filling, which jointly reweights regions across learnware specifications, so that they collectively match the user distribution while irrelevant regions are down-weighted. For each learnware, the resulting weights quantify its overall relevance, which is combined with its compatibility with individual inputs to determine how much it contributes to the prediction on each input. Experiments show improved predictive performance over existing multi-learnware reuse methods and further demonstrate favorable accuracy–efficiency trade-offs for our method.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
learnware, model reuse, model combination, distribution matching, kernel mean embedding

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_vN88OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies how to reuse multiple learnwares when each one only partly meets the user's needs. It proposes Distribution Filling, which reweights points within the specifications to match the user distribution. The resulting weights are combined with local compatibility to select a model or combine predictions for each input. Experiments on tabular data, images, and sales prediction show improvements over existing learnware reuse methods.

Critical strengths
  1. The problem is clear and the method is easy to follow. The paper considers cases where each learnware only partly meets the user's needs. It reweights points within the specifications to combine useful parts from different learnwares and reduce the influence of irrelevant regions. It only needs to solve a convex optimization problem once for each user task, without training an extra selector.

  2. The experiments are extensive, which cover tabular data, images, and a sales prediction task. The paper also includes ablation studies, parameter analysis, and runtime comparisons. These experiments provide useful support for the method.

Critical weaknesses
  1. The contribution of pointwise filling to the overall improvement is unclear. The method shows large gains over JobSelector, but only gains 0.14–0.71 percentage points over the simple baseline using local compatibility alone. The paper compares whole-specification and pointwise weighting on some datasets, but this direct comparison is missing from the main benchmark.

  2. The link between distribution matching and prediction ability is not well established. A model may cover an input region but still make poor predictions there. The paper acknowledges this difference, but does not sufficiently explain when the distribution-based weights are reliable and when they may fail.

Questions for authors
  1. Could you compare whole-specification weighting plus local compatibility with pointwise weighting plus local compatibility on the main benchmark? This would help clarify the contribution of pointwise filling.

  2. Under what conditions does distribution coverage provide useful evidence of prediction ability? Could you discuss a case where this relationship may fail and how it would affect the method?

  3. Besides the accuracy–cost trade-off metioned in the paper, are there other practical considerations for choosing DF–Select over DF–MoE, for example when the candidate models’ probabilities are not well calibrated?

4 comments
AuthorsOpenReview ↗

Response to Reviewer vN88 (Part 1/2)

Thank you for recognizing the clarity and soundness of our work and the extensive evaluation, and for the constructive suggestions on clarifying DF's contribution and use. We address each concern below.

1. "The contribution of pointwise filling to the overall improvement is unclear."

As shown in the original Table 9, DF–MoE improves over Uniform-local in 49/60 dataset-scenario pairs. The original Fig. 4 and Appendix G.2 further compare pointwise and whole-specification weighting at fixed local compatibility, with gains of 18.40/7.36 points on CJS and 1.95/0.84 on EuroSAT (subsumption/instance recurrence).

Learnware pools accumulate models developed independently for different historical tasks, so the prevalence of a capability in the pool need not reflect its importance to a new user. This mismatch arises naturally across sites or operating regimes: in the original Fig. 1, the factory needs complementary low- and high-load capabilities despite overlapping source coverage of medium-load conditions.

To clarify DF's contribution to complementary reuse, we conducted an additional experiment on 13 tabular datasets with five candidates over six classes A–F. Each of D and F is covered by only one candidate; the other classes by several. We varied total D/F demand, split equally, while fixing predictions and local compatibility. Results are dataset-averaged accuracy (%) and paired gains (pp), reported as mean ± sample SD over three seeds.

Total D/F demandUniform-localWhole-localDF–MoEGain vs Uniform-local (pp)
33.3% (balanced)79.01 ± 0.2677.66 ± 0.4279.25 ± 0.31+0.24 ± 0.06
50%72.76 ± 0.3273.86 ± 0.4575.19 ± 0.49+2.43 ± 0.30
80%61.51 ± 0.4468.59 ± 0.7570.38 ± 0.45+8.87 ± 0.65
90%57.76 ± 0.4866.86 ± 0.7068.77 ± 0.89+11.02 ± 0.91

These results show that DF adds substantial value beyond local compatibility when demand concentrates on reliable capabilities supplied by few learnwares. Local compatibility indicates where a model is applicable; DF adapts its task-level contribution to user demand and coverage supplied by other candidates. Here, DF raises the D/F candidates' weights, directing local fusion toward their reliable complementary predictions. We added the results and mechanism analysis to Section 5.5 (Fig. 4) and Appendix G.3 (Fig. 9 and Table 13) of the revision.

2. "The link between distribution matching and prediction ability is not well established."

"Under what conditions does distribution coverage provide useful evidence of prediction ability? Could you discuss a case where this relationship may fail and how it would affect the method?"

Section 4.1 (Eqs. (2)–(3)) values complementary capabilities, motivating DF's pointwise allocation and local compatibility. Proposition 1 explains the allocation mechanism. Our added analysis bounds the prediction loss of the final DF–Select and DF–MoE rules using the fitted reuse weights in Eqs. (8)–(9). It separates regional model error, reuse-weight mismatch, and an additional term for selecting one model, making explicit when complementary coverage supports reliable prediction.

We assume (i) the user density is a mixture pu=∑iaipirelp_u=\sum_i a_i p_i^{\mathrm{rel}} of relevant source regions, which may overlap; and (ii) model ii has expected loss at most ϵi\epsilon_i on its region under the user's label relationship, for a loss in [0,1][0,1]. The first assumption captures tasks assembled from portions of source tasks. The second assumption is reasonable when models perform well on the relevant source regions and the input–label relationships remain stable from source to user, so that their regional competence can transfer to the user task.

The reference weight ri†(x)=aipirel(x)/pu(x)r_i^\dagger(x)=a_i p_i^{\mathrm{rel}}(x)/p_u(x) measures region ii's share at input xx. Let δ=EX[12∑i∣ri(X)−ri†(X)∣]\delta=\mathbb E_X[\frac12\sum_i|r_i(X)-r_i^\dagger(X)|] measure the discrepancy of the actual DF weights from these reference weights. Theorem 1 bounds DF–Select for any such loss, including zero–one loss:

RSelect≤∑iaiϵi+δ+EX[1−max⁡iri(X)].R_{\mathrm{Select}}\le\sum_i a_i\epsilon_i+\delta+\mathbb E_X[1-\max_i r_i(X)].

For aligned outputs in a common convex space and a convex loss, such as normalized Brier loss for probability fusion, the theorem also gives:

RMoE≤∑iaiϵi+δ.R_{\mathrm{MoE}}\le\sum_i a_i\epsilon_i+\delta.

Response 2 continues in Part 2.

AuthorsOpenReview ↗

Response to Reviewer vN88 (Part 2/2)

Continuation of response 2

2. "The link between distribution matching and prediction ability is not well established."

The first term bounds weighted regional prediction error; δ\delta measures reuse-weight mismatch. The extra DF–Select term accounts for committing to one model. DF–MoE has low convex-loss risk when regional errors and weight mismatch are small; DF–Select has low risk when its weights are also concentrated. With predictive utility defined as one minus prediction loss, the DF–Select bound also gives conditions for high achieved utility within the framework of Section 4.1.

As an illustrative failure case, suppose the same input region has label A in the source but label B for the user. Its input distribution can match perfectly while the source classifier still predicts A. DF may then select that model or assign substantial fusion weight to its incorrect output. The model's regional error is then large despite a good input match. We added the assumptions, risk bounds and their interpretation to Section 4.4 (Assumptions 1–2, Theorem 1, Eqs. (11)–(13)) of the revision, with the proof in Appendix C.2.

3. "Could you compare whole-specification weighting plus local compatibility with pointwise weighting plus local compatibility on the main benchmark?"

Thank you for this suggestion. We added a paired experiment on reconstructed main-benchmark tasks comparing whole-specification weighting (Whole-local) with pointwise weighting (DF–MoE), holding candidates, local compatibility and probability fusion fixed. Accuracy (%) is averaged over 17 tabular or 13 image datasets.

SettingWhole-localDF–MoEGain (pp)
Tabular / Subsumption84.5189.20+4.70
Tabular / Instance recurrence81.3482.44+1.09
Image / Subsumption89.2189.88+0.68
Image / Instance recurrence75.6475.75+0.10

DF–MoE improves over Whole-local in 45/60 dataset-scenario pairs, with the largest mean gain under tabular subsumption. This supports the benefit of adjusting useful and irrelevant source regions separately while keeping local compatibility fixed. We added this comparison to Appendix G.2 (Table 9) of the revision.

4. "...are there other practical considerations for choosing DF–Select over DF–MoE, for example when the candidate models’ probabilities are not well calibrated?"

As described in Section 4.3 (Eqs. (9)–(10)), DF–Select chooses a model using input-distribution information and returns its class prediction, whereas DF–MoE combines aligned probabilities. Reliability of the candidates' probabilities is therefore an additional practical consideration.

We tested a simple overconfidence setting across all 60 main-benchmark dataset-scenario pairs. We first temperature-scaled candidate probabilities using labeled samples separate from evaluation, then increased all candidates' confidence by dividing their temperatures by four, keeping their predicted classes and reuse weights fixed. Results are averaged over all 60 pairs.

Probability settingDF–MoE accuracy (%)DF–Select accuracy (%)
Before confidence increase85.1581.48
After confidence increase84.7581.48

Increasing confidence reduces DF–MoE accuracy by 0.40 percentage points while DF–Select remains unchanged. DF–Select is therefore a practical option when class predictions are trusted but confidence values are unreliable.

We hope these clarifications and revisions address your concerns.

Thank you for your detailed response. It has addressed my concerns. I will keep my rating at 3.

Thank you for reviewing our responses and confirming that they have addressed your concerns. We appreciate your time and constructive feedback, which helped us improve the paper.

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

The manuscript studies how to leverage a collection of specialised machine learning models for new user tasks. Within the learnware framework, each model is paired with a compact statistical description of its training inputs, which is matched to a user description built from unlabelled data. The particular issue the authors tackle is identifying and combining models when each covers only part of the user’s task. A closely related method is JobSelector, which assigns a weight to each model's statistical description. The authors extend the distribution-matching approach by assigning weights to representative points within each description and optimising these weights across the candidates. The point weights are summed to compute each model’s overall relevance, which is combined with local similarity from the original descriptions to select a model or combine several models’ predictions. The authors show that an ideal coverage measure based on fixed utility functions has diminishing returns. Separately, they show that their regularised distribution-fitting objective has a unique solution and characterise the resulting point weights, without proving that these weights optimise predictive utility. The method is evaluated using models trained on class subsets across tabular and image classification datasets, and in a sales forecasting scenario, with comparisons against existing learnware methods.

Critical strengths
  1. The paper provides practical insights into combining a collection of ML models in cases where each model covers only part of the user’s task. The proposed method, evaluated empirically across several settings, allows useful and irrelevant regions within a model’s statistical description to receive separate weights.

  2. The evaluation is broad and covers tabular and image classification, as well as a sales forecasting problem. The appendix also includes component ablations that are useful for understanding which parts of the proposed method bring improvements.

Critical weaknesses
  1. There is a disconnect between the ideal utility formulation and the proposed and evaluated method: the utility formulation reflects predictive usefulness and deployment requirements, while the algorithm assigns scores through input distribution matching. The proofs are helpful in understanding the ideal coverage measure and fitting objective, but they do not establish a connection between the resulting weights and the ideal utility objective or predictive performance. The authors acknowledge this limitation.

  2. The additional value of the algorithm appears insufficient, given that a simpler method using local similarity achieves comparable average performance. While the appendix provides good ablations over the components that bring improvement, the modest average gains and omission of this simple baseline from the sales forecasting experiment leave its additional practical utility unclear.

Questions for authors
  1. Under what assumptions can the fitted weights and resulting reuse scores be connected to predictive utility for DF–Select or DF–MoE?
  2. How reliable and practically meaningful are the gains over Uniform-local, and can you include this baseline in the sales forecasting experiment?
3 comments
AuthorsOpenReview ↗

Response to Reviewer UUEX (Part 1/2)

Thank you for recognizing the practical insights of our work, the broad evaluation, and the component ablations. We address each concern below.

1. "There is a disconnect between the ideal utility formulation and the proposed and evaluated method."

"Under what assumptions can the fitted weights and resulting reuse scores be connected to predictive utility for DF–Select or DF–MoE?"

Section 4.1 (Eqs. (2)–(3)) values complementary capabilities, motivating DF's pointwise allocation and local compatibility. Proposition 1 explains the allocation mechanism. Our added analysis bounds the prediction loss of the final DF–Select and DF–MoE rules using the fitted reuse weights in Eqs. (8)–(9). It separates regional model error, reuse-weight mismatch, and an additional term for selecting one model, making explicit when complementary coverage supports reliable prediction.

We assume (i) the user density is a mixture pu=∑iaipirelp_u=\sum_i a_i p_i^{\mathrm{rel}} of relevant source regions, which may overlap; and (ii) model ii has expected loss at most ϵi\epsilon_i on its region under the user's label relationship, for a loss in [0,1][0,1]. The first assumption captures tasks assembled from portions of source tasks. The second assumption is reasonable when models perform well on the relevant source regions and the input–label relationships remain stable from source to user, so that their regional competence can transfer to the user task.

The reference weight ri†(x)=aipirel(x)/pu(x)r_i^\dagger(x)=a_i p_i^{\mathrm{rel}}(x)/p_u(x) measures region ii's share at input xx. Let δ=EX[12∑i∣ri(X)−ri†(X)∣]\delta=\mathbb E_X[\frac12\sum_i|r_i(X)-r_i^\dagger(X)|] measure the discrepancy of the actual DF weights from these reference weights. Theorem 1 bounds DF–Select for any such loss, including zero–one loss:

RSelect≤∑iaiϵi+δ+EX[1−max⁡iri(X)].R_{\mathrm{Select}}\le\sum_i a_i\epsilon_i+\delta+\mathbb E_X[1-\max_i r_i(X)].

For aligned outputs in a common convex space and a convex loss, such as normalized Brier loss for probability fusion, the theorem also gives:

RMoE≤∑iaiϵi+δ.R_{\mathrm{MoE}}\le\sum_i a_i\epsilon_i+\delta.

The first term bounds weighted regional prediction error; δ\delta measures reuse-weight mismatch. The extra DF–Select term accounts for committing to one model. DF–MoE has low convex-loss risk when regional errors and weight mismatch are small; DF–Select has low risk when its weights are also concentrated. With predictive utility defined as one minus prediction loss, the DF–Select bound also gives conditions for high achieved utility within the framework of Section 4.1. We added this analysis to Section 4.4 (Assumptions 1–2, Theorem 1, Eqs. (11)–(13)) of the revision, with the proof in Appendix C.2.

AuthorsOpenReview ↗

Response to Reviewer UUEX (Part 2/2)

2. "The additional value of the algorithm appears insufficient..."

"How reliable and practically meaningful are the gains over Uniform-local?"

As shown in the original Table 9, DF–MoE improves over Uniform-local in 49/60 dataset-scenario pairs. The original Fig. 4 and Appendix G.2 further compare pointwise and whole-specification weighting at fixed local compatibility, with gains of 18.40/7.36 points on CJS and 1.95/0.84 on EuroSAT (subsumption/instance recurrence).

Learnware pools accumulate models developed independently for different historical tasks, so the prevalence of a capability in the pool need not reflect its importance to a new user. This mismatch arises naturally across sites or operating regimes: in the original Fig. 1, the factory needs complementary low- and high-load capabilities despite overlapping source coverage of medium-load conditions.

To clarify DF's contribution to complementary reuse, we conducted an additional experiment on 13 tabular datasets with five candidates over six classes A–F. Each of D and F is covered by only one candidate; the other classes by several. We varied total D/F demand, split equally, while fixing predictions and local compatibility. Results are dataset-averaged accuracy (%) and paired gains (pp), reported as mean ± sample SD over three seeds.

Total D/F demandUniform-localWhole-localDF–MoEGain vs Uniform-local (pp)
33.3% (balanced)79.01 ± 0.2677.66 ± 0.4279.25 ± 0.31+0.24 ± 0.06
50%72.76 ± 0.3273.86 ± 0.4575.19 ± 0.49+2.43 ± 0.30
80%61.51 ± 0.4468.59 ± 0.7570.38 ± 0.45+8.87 ± 0.65
90%57.76 ± 0.4866.86 ± 0.7068.77 ± 0.89+11.02 ± 0.91

These results show that DF adds substantial value beyond local compatibility when demand concentrates on reliable capabilities supplied by few learnwares. Local compatibility indicates where a model is applicable; DF adapts its task-level contribution to user demand and coverage supplied by other candidates. Here, DF raises the D/F candidates' weights, directing local fusion toward their reliable complementary predictions. We added the results and mechanism analysis to Section 5.5 (Fig. 4) and Appendix G.3 (Fig. 9 and Table 13) of the revision.

3. "Can you include this baseline in the sales forecasting experiment?"

Thank you for pointing this out. We added Uniform-local using the same candidates and predictions on all 240 sales tasks. DF–MoE reduces overall mean absolute error (MAE) by 1.05%, rising to 1.29% under skewed demand, relative to Uniform-local. This is consistent with global relevance adapting model weights to uneven user demand. The overall reduction occurs in all three seeds. We added the results and analysis to Appendix H.2 (Table 18 and Fig. 15) of the revision.

We hope these clarifications and revisions address your concerns.

AuthorsOpenReview ↗

Follow-up on Our Response

Dear Reviewer UUEX,

With the discussion period nearing its end, we would like to follow up on our rebuttal and see whether it has adequately addressed your concerns. We hope our responses and additional analyses have provided useful clarification. If you have any remaining questions or suggestions, we would be happy to address them before the discussion closes.

Thank you again for your valuable feedback, time, and consideration.

Best regards,

The Authors

Reviewer_3xfCOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper address resuing models from a learnware market when no single model covers the user's tasks and each model also covers irrelevant regions. This work use Reduced kernel Mean embedding to represent a data distribution and measure how similar two distributions are.

This paper proposes distribution filling (DF) to compute global relevance πi\pi_i and local compatibility qi(x)q_i(x). Then use the πiqi(x)\pi_i q_i(x) as the score to select model. It either use DF-select, using a single model with highest score, or DF-MoE, weighted averaging all models.

On 17 tabular and 13 image datasets. DF out perform baseline (Jobselector).

Critical strengths
  • The subsumption problem is well motivated.
  • The proposed method is simple and convex. And prop. 1 explains how mass is allocated.
  • This work also provides broad evaluations, including timing, ablation, sensitivity.
  • The proposed method is faster than baseline.
Critical weaknesses
  • The core contribution from the proposed component (DF) is marginal. On the main benchmark, Table 9 shows DF-MoE outperforms uniform global weights with the same local compatibility by only ~0.4 point. Most of the gain comes from local compatibility (Figure 4).

  • Limited novelty. Reweighting points to match a target distribution is kernel mean matching.

Questions for authors
  • Local compatibility is a kernel density estimate with one fixed bandwidth for all datasets. How sensitive are the results to the bandwidth selection?
3 comments
AuthorsOpenReview ↗

Response to Reviewer 3xfC (Part 1/2)

Thank you for recognizing our motivation, the simplicity and convexity of our method, and the breadth and efficiency of the evaluation. We address each concern below.

1. "The core contribution from the proposed component (DF) is marginal."

As shown in the original Table 9, DF–MoE improves over Uniform-local in 49/60 dataset-scenario pairs. The original Fig. 4 and Appendix G.2 further compare pointwise and whole-specification weighting at fixed local compatibility, with gains of 18.40/7.36 points on CJS and 1.95/0.84 on EuroSAT (subsumption/instance recurrence).

Learnware pools accumulate models developed independently for different historical tasks, so the prevalence of a capability in the pool need not reflect its importance to a new user. This mismatch arises naturally across sites or operating regimes: in the original Fig. 1, the factory needs complementary low- and high-load capabilities despite overlapping source coverage of medium-load conditions.

To clarify DF's contribution to complementary reuse, we conducted an additional experiment on 13 tabular datasets with five candidates over six classes A–F. Each of D and F is covered by only one candidate; the other classes by several. We varied total D/F demand, split equally, while fixing predictions and local compatibility. Results are dataset-averaged accuracy (%) and paired gains (pp), reported as mean ± sample SD over three seeds.

Total D/F demandUniform-localWhole-localDF–MoEGain vs Uniform-local (pp)
33.3% (balanced)79.01 ± 0.2677.66 ± 0.4279.25 ± 0.31+0.24 ± 0.06
50%72.76 ± 0.3273.86 ± 0.4575.19 ± 0.49+2.43 ± 0.30
80%61.51 ± 0.4468.59 ± 0.7570.38 ± 0.45+8.87 ± 0.65
90%57.76 ± 0.4866.86 ± 0.7068.77 ± 0.89+11.02 ± 0.91

These results show that DF adds substantial value beyond local compatibility when demand concentrates on reliable capabilities supplied by few learnwares. Local compatibility indicates where a model is applicable; DF adapts its task-level contribution to user demand and coverage supplied by other candidates. Here, DF raises the D/F candidates' weights, directing local fusion toward their reliable complementary predictions. We added the results and mechanism analysis to Section 5.5 (Fig. 4) and Appendix G.3 (Fig. 9 and Table 13) of the revision.

AuthorsOpenReview ↗

Response to Reviewer 3xfC (Part 2/2)

2. "Limited novelty. Reweighting points to match a target distribution is kernel mean matching."

DF shares KMM's RKHS matching principle (Appendix A). Classical KMM weights samples for predictor training. Our contribution is to infer fixed models' task-level relevance through joint regional allocation of their specifications, enabling complementary reuse from unlabeled user inputs. A model may cover useful parts of the user task together with irrelevant regions; one weight for its entire specification scales both together (Eq. (5)). DF jointly adjusts regional contributions across candidates (Eq. (6)), allowing the match to combine useful portions of different models. Each candidate's allocated mass determines its overall relevance. We retain the original specification for local compatibility because a region can receive little matching mass simply because other candidates already cover it. Combining these quantities gives input-wise reuse weights (Eqs. (8)–(9)) for selection or fusion, without retraining candidates or learning an additional selector.

Fig. 3 illustrates the resulting representational gain: whole-specification mixtures retain the irrelevant mass shared by all candidates, whereas DF reallocates it to useful regions. At 90% extra mass, independent-sample MMD² falls from 0.438 (Whole) to 0.0020 (DF). The fixed-local controls in the original Fig. 4 further show that this finer allocation improves prediction with the same models and local compatibility.

3. "How sensitive are the results to the bandwidth selection?"

Thank you for raising this point. Bandwidth controls how local the compatibility scores are, so we added a sensitivity sweep while fixing specifications, candidates, global weights and model predictions. The table reports DF–MoE accuracy (%), averaged over 17 tabular or 13 image datasets. The bandwidth ratio is relative to the default. Abbreviations follow the original paper.

Bandwidth ratioTabular Sub.Tabular IRImage Sub.Image IR
0.2590.1982.9086.2173.00
0.590.3482.9987.5573.94
0.7590.1282.8989.0475.07
1 (default)89.2082.4489.8875.75
1.586.8780.2489.2274.20
285.3577.2786.9171.01
482.9171.4882.5263.33

At the sampled ratios 0.75–1.5, all four averages stay within 2.33 points of the default. Tabular accuracy peaks at ratio 0.5, while both image settings peak at the default. Performance is relatively stable near the default, but overly broad bandwidths reduce accuracy, consistent with weaker local discrimination. We added this analysis to Section 5.5 and Appendix G.5 (Fig. 12) of the revision.

We hope these clarifications and revisions address your concerns.

AuthorsOpenReview ↗

Follow-up on Our Response

Dear Reviewer 3xfC,

With the discussion period nearing its end, we would like to follow up on our rebuttal and see whether it has adequately addressed your concerns. We hope our responses and additional analyses have provided useful clarification. If you have any remaining questions or suggestions, we would be happy to address them before the discussion closes.

Thank you again for your valuable feedback, time, and consideration.

Best regards,

The Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Rebuttal and Key Revisions

Summary of Rebuttal

We thank the reviewers for their thoughtful feedback. We appreciate their recognition of the practical importance of reusing models with partially overlapping capabilities (3xfC, UUEX, vN88), the simple and efficient convex formulation (3xfC, vN88), and the broad experimental evaluation (3xfC, UUEX, vN88). Below, we summarize how we addressed the main concerns.

  1. Clarification of the Contribution of DF (3xfC, UUEX, vN88). We added controlled experiments that vary user demand while keeping candidate predictions and local compatibility fixed. They show that DF provides substantial gains when demand concentrates on reliable capabilities supplied by few models. This captures a practical mismatch between models developed for historical tasks and the capabilities needed by a new user. Local compatibility identifies where a model applies, while DF adjusts its overall contribution according to user demand and the coverage supplied by other candidates.

  2. New Analysis Connecting Coverage to Predictive Performance (UUEX, vN88). We added risk bounds for the actual DF–MoE and DF–Select prediction rules, with convex losses for DF–MoE. These bounds show that low regional prediction errors and accurate reuse weights give low predictive-risk bounds. For DF–Select, weights concentrated on one model per input reduce the additional selection-risk term. The selection bound also connects to predictive utility through one minus prediction loss.

  3. Clarification of Novelty Beyond Kernel Mean Matching (3xfC). We explained that our contribution is to infer the task-level relevance of fixed models through joint regional allocation of their specifications, enabling complementary reuse from unlabeled user inputs. By weighting useful and irrelevant regions separately, DF combines useful capabilities from different models. The resulting model relevance is combined with local compatibility for input-wise selection or fusion, without retraining the candidates or learning an additional selector.

  4. Expanded Experimental Validation (3xfC, UUEX, vN88). We added the requested whole-specification comparison under fixed local compatibility, showing that DF improves predictive performance over whole-specification weighting and supporting the benefit of weighting useful and irrelevant regions separately (vN88). We also included Uniform-local in sales forecasting, where DF–MoE achieves lower overall MAE in all three seeds, demonstrating gains beyond local compatibility in this application (UUEX). Further experiments clarify practical choices. The bandwidth sweep shows relative stability near the default and reduced accuracy with overly broad kernels (3xfC). The confidence experiment in our response shows that increasing confidence while preserving class predictions affects DF–MoE but leaves DF–Select unchanged, supporting selection when confidence estimates are unreliable (vN88).

Summary of Changes in Revision

  • Section 4.4 adds the assumptions, prediction-risk bounds, and interpretation connecting regional competence and fitted reuse weights to predictive performance (Theorem 1). Appendix C.2 provides the proof.
  • Section 5.5 adds the controlled user-demand experiment and explains when DF contributes beyond local compatibility (Fig. 4). It also summarizes the direct comparison with whole-specification weighting and the bandwidth sensitivity analysis. Appendices G.2, G.3, and G.5 provide the detailed results and mechanism analysis.
  • Section 5.4 and Appendix H.2 incorporate the Uniform-local comparison for sales forecasting. Table 18 reports RMSE, MAE, and MAE ranks for all methods, with further analysis in Fig. 15.

We believe these clarifications, additional experiments, and revisions fully address the concerns raised by the reviewers. We thank the reviewers and the Area Chair for their time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.