acceptodds
Under review as a conference paper at ICLR 2027

COIN: Coordinating Multimodal Updates through Functional Invariance

Abstract

Multimodal learning must not only train each modality well, but also coordinate how their updates affect the shared prediction. Joint training can underuse informative modalities, while improving each modality separately does not ensure that their updates work well together after fusion. We show that module-wise signals can be misleading: coordinate rescaling can create arbitrarily different learning speeds, and large module responses can cancel while producing little or no fused-output change. To study this interaction directly, we characterize each module by the first-order change it can induce in the fused output and identify a *functional invariant relation*: different module responses can produce the same combined output change. Among such equivalent responses, the minimum-energy allocation removes unnecessary module movement without changing the prediction to first order. Building on this principle, we derive **COIN**, which jointly chooses a task-improving output change and how to distribute it across modules. COIN accounts for both each module's relation to the task and its interaction with other modules through a small linear system. This view also explains why a module that cannot improve the task on its own may still help by correcting another module's response. We further bound total module response and derive the exact cost of ignoring cross-module interactions. Across evaluated benchmarks, COIN achieves the highest mean fusion accuracy among the compared methods.

Primary area:
optimization
Keywords:
multimodal learning, multimodal optimization, modality coordination, functional invariance

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

This paper proposes COIN, a multimodal optimization framework that coordinates parameter updates according to their first-order effects on the fused prediction. Rather than relying on modality-specific gradient magnitudes or learning speeds, the authors introduce module response spaces to characterize attainable changes in the shared output. They identify an output-preserving equivalence relation between module responses and derive a minimum-energy allocation principle to eliminate redundant functional movement.

Building on this formulation, COIN jointly determines task-aligned responses and their allocation across modules through a regularized least-squares objective. The resulting solution incorporates both module-task alignment and cross-module interactions through a small Gram-matrix linear system. The paper further provides theoretical results concerning uniqueness, basis invariance, bounded functional motion, and the cost of ignoring cross-module interactions.

Experiments on AVE, AVSBench, VGGSound50, and FOOD101 demonstrate improvements over several multimodal learning baselines. Additional analyses investigate allocation efficiency, module cooperation, response-space geometry, and finite-step behavior.

Critical strengths
  1. The paper studies multimodal optimization through the output changes induced by individual modules, rather than comparing their parameter-space gradients or learning speeds. This formulation highlights interactions that cannot be captured by module-wise statistics alone. This offers a useful perspective on cooperation between multimodal components.

  2. The proposed objective admits a well-defined solution, and the paper establishes several relevant properties, including minimum-energy decomposition, basis invariance for fixed response spaces, bounded functional motion, and an exact characterization of the cost of independent allocation. The mathematical arguments appear carefully developed, with important assumptions and limitations explicitly discussed.

  3. Experimentally, COIN achieves the highest reported mean fusion accuracy among the compared methods on all four benchmarks. The additional matched-progress comparisons, module removal experiments, and finite-step diagnostics provide further support for the proposed functional coordination mechanism.

Critical weaknesses
  1. The theoretical components presented are all existing. Although their application to multimodal update coordination is meaningful, the paper could more clearly articulate which theoretical insights are fundamentally new beyond these established constructions.

  2. The practical scalability of COIN remains insufficiently validated.

  3. The paper does not report measured training-time or memory overhead relative to competing methods. The experiments also primarily involve conventional two-modality classification architectures, leaving the applicability to larger contemporary multimodal models uncertain.

  4. There is no dedicated conclusion section, which is inconsistent with the standard structure of a full paper.

  5. Some reported accuracy statistics appear inconsistent with the stated test-set sizes and three-seed evaluation protocol. For example, Appendix J.1 (Table 4) reports 402 test samples for AVE, while Table 1 reports an average accuracy of 69.87% for ARL over three seeds. Assuming standard sample-level Top-1 accuracy, the total number of correct predictions should be an integer. However, 402 × 3 × 0.6987 = 842.6322, which is not an integer. This discrepancy cannot be explained by rounding to two decimal places, as the nearest valid averages are 69.82% and 69.90%. The authors should clarify the evaluation protocol and provide per-seed results to verify the numerical consistency and reliability of the reported findings.

Questions for authors

Please see the critical weaknesses above. If the authors can convincingly address these concerns, I would be willing to reconsider and potentially raise my score. If the authors fail to adequately address Weakness 5, I will lower my rating.

5 comments

W5: The AVE accuracy appears inconsistent with 402 test samples and three seeds. Please clarify the evaluation protocol and numerical consistency.

The AVE test loader previously loaded only 395 examples. We have fixed this issue, and the current evaluation uses all 402 test examples. For ARL, the three seeds correctly classified 275, 275, and 278 out of 395 examples, yielding 69.87 ± 0.44%.

W1: The theoretical tools are established. Which insights are specific to this work?

Our contribution is the task-dependent coordination of module responses in the fused output space. We build on generalized inverses, orthogonal projection, and regularized least squares, as discussed in Appendix I.

Table 2 distinguishes modality-strength modulation, objective-gradient reconciliation, and separated learning. Section 2 identifies the remaining coordination issue: module-local magnitudes or learning speeds do not determine how updates to different blocks combine in the fused prediction. COIN instead allocates their attainable output responses jointly.

Three results distinguish this setting:

  1. Task-dependent allocation: Proposition 4.3 gives identical module-local norms and task alignments but different optimal joint responses.
  2. Exact interaction cost: Theorem 4.4 quantifies the cost of independent allocation; its gap is zero precisely when Eb=0Eb=0.
  3. Conditional compensation: Appendix E.2 characterizes how a module can improve the joint response by compensating for other modules, even with zero standalone task alignment.

The new full-training controls connect this characterization to final performance: Full COIN achieves the highest mean accuracy among all six variants on all four datasets, including both single-direction controls. The margin over Independent is very small on VGGSound50 (61.52% versus 61.50%), while the gains are larger on the other three datasets.

The submission also evaluates recent external baselines, including MoNet (ICML 2026) and IBML (TPAMI 2026), alongside DGL, InfoReg, and MMPareto (Table 1). The submitted benchmark comparison reports higher fusion means for COIN than these external baselines. The new controlled comparisons address which part of the method contributes to that performance.

W2 & W3: Practical scalability, measured training cost, and applicability to larger multimodal models remain unclear.

We measured optimizer time and peak allocated memory, including derivatives for correction. The isolated profile uses the same H100 and averages three profiling windows per method.

DatasetBase (ms/step)Full (ms/step)Full / BaseBase memory (GiB)Full memory (GiB)
AVE156.919227.9861.453×6.9307.056
AVSBench85.921127.5351.484×3.8753.962
VGGSound5086.09192.1501.070×3.8753.962
FOOD10138.36248.7001.269×1.5252.539

The overhead varies across tasks: FOOD101 peak allocated memory increases by 1.014 GiB. Independent and Scalar have similar profiled time to Full, as fresh gradients and Jacobian–vector products dominate the additional computation (Appendix H.4).

With two candidates per module, k≤2Mk\leq2M; the five-module implementation has k≤10k\leq10. Gram construction costs O(NCk2)O(NCk^2) and the solve costs O(k3)O(k^3), alongside response estimation and parameter-direction storage. The reported measurements cover two-modality classification architectures; larger-scale and three-modality evaluations remain for future work.

W4: The paper has no dedicated conclusion section.

Useful module updates also need to cooperate through their effects on the shared prediction. COIN formalizes output-preserving reallocations, selects a minimum-energy joint response, and implements it through a small candidate-space solve (Sections 3–5). Fig. 1/Table 7 establish the local advantage over common rescaling, while the full-training ablations show the resulting accuracy gains across the evaluated datasets. The analysis characterizes local response spaces and the experiments test two-modality classification networks.

Reviewer_JqBeOpenReview ↗

Follow up

Thank you for the response. However, I still have several significant concerns that have not been adequately addressed.

  1. The incorrect loading of the AVE dataset affects not only ARL but also all other methods, including the results reported in this paper. Therefore, the corresponding experimental results need to be corrected. However, the authors have neither provided an explanation for this issue nor made the necessary corrections. As a result, I still have serious concerns about the validity and reliability of the reported experimental results.

  2. Regarding the authors' response to W4: In my previous review, I explicitly pointed out that:

    "There is no dedicated conclusion section, which is inconsistent with the standard structure of a full paper."

    The authors also correctly summarized this concern in their current response as:

    "W4: The paper has no dedicated conclusion section."

However, their actual response is difficult to understand. My interpretation is that the authors have provided the content they intend to include in a conclusion section. Specifically, there is no explicit indication, such as "This is our conclusion section," to clarify how this issue has been resolved. Instead, the response consists of a bolded summary statement followed by some additional content. I am therefore unsure whether the authors may have mistakenly provided an incorrect or misplaced response to this comment. Could the authors please clarify?

  1. Regarding the manuscript revisions: As of the time of writing this review, the changes described in the authors' response have not been reflected in the revised manuscript.

If I have misunderstood or overlooked anything in the authors' response or revisions, please feel free to point it out and provide further clarification.

Thank you for the follow-up. We have uploaded the revised manuscript, including the corrected results, additional experiments, and a dedicated conclusion.

AVE Evaluation and Baseline Reporting. We reproduced all baseline methods. Since some reproduced results were close to those reported in the original papers, we initially retained the published values for those methods. We have now updated the manuscript using our verified three-seed results, with all AVE evaluations standardized to the complete 402-example test set. For example, ARL has been corrected from 69.87 ± 0.44% (395 examples) to 69.90 ± 0.43% (402 examples). The updated results are reported in Tables 1 and 7. Our main findings remain unchanged.

Other revisions. Table 2 (p. 9) presents six complete-training controls, and Appendix J.8, Tables 11–13 (pp. 37–38) reports measured time and memory, sensitivity, and same-base masked-modality results. We have also added the requested Section 7, “Conclusion” (p. 10).

Theoretical contribution. COIN is not presented as a new projection or least-squares algorithm. Its contribution is the task-dependent joint allocation of module responses in fused-output space: Proposition 4.3 establishes compensating updates that module-local statistics cannot determine, and Theorem 4.4 quantifies the exact local cost of ignoring cross-module interactions. The complete-training results in Table 2 provide corresponding empirical evidence. We hope these concrete revisions address your concerns.

Reviewer_J6sJOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes COIN to coordinate parameter updates across modules in multimodal learning. COIN measures the first-order effect of each candidate update on the fused output. It jointly selects update coefficients to match a target direction and penalizes the total response energy. This optimization reduces to a small linear system in low-dimensional candidate spaces. The paper evaluates COIN on four multimodal benchmarks and studies update allocation, module removal, and the accuracy of the first-order approximation.

Critical strengths
  1. The paper shifts the coordination target from gradient magnitudes to each module’s effect on the fused output. This perspective explains how separate parameter blocks can still produce cancelling or compensating effects through a shared output.
  2. The method connects an explicit objective to a practical low-dimensional correction. Appendices J.3, J.4, J.6, and J.7 examine update allocation, redistribution after module removal, and agreement between first-order predictions and actual loss changes.
Critical weaknesses
  1. The paper compares joint and independent updates at saved checkpoints in Appendices J.3–J.4, but does not test their effect on final accuracy through full training. Table 1 lacks comparisons with independent allocation and uniform scaling under the same base training setup. The paper also does not test a variant with the cross-module interaction terms removed.
  2. The experiments fix the two candidate directions per module, calibration batch size, and correction frequency. The experiments vary the local step size, but it does not evaluate how these implementation choices affect final test accuracy.
Questions for authors
  1. Please report final accuracy for independent allocation, uniform scaling, and full COIN under matched compute budgets. What is the result of removing the cross-module interaction terms while keeping the candidate directions and correction schedule unchanged?
  2. How much training time and peak memory do COIN corrections add to the same base training?
  3. How sensitive is final accuracy to the calibration batch size and correction frequency? Are the same settings suitable across the four datasets?
4 comments

W1 & Q1: Please compare independent allocation, uniform scaling, and full COIN through complete training. What happens when cross-module interaction terms are removed?

We evaluate the requested controls through complete training under the same base setup, using three training seeds per dataset.

MethodAVEAVSBenchVGGSound50FOOD101
Base73.3893.3860.7891.30
Independent / no cross-module terms73.1392.7061.5091.29
Scalar71.8993.6560.7891.42
Base direction only72.8993.5160.3491.26
Fused gradient only73.1393.2461.2291.31
Full COIN74.3894.1461.5291.96

Independent allocation corresponds to removing the cross-module terms. Each within-module basis is orthonormal, so Gmm=IG_{mm}=I; removing the off-diagonal blocks in Eq. (12) gives aind=b/2a_{\mathrm{ind}}=b/2.

Scalar instead optimizes one common coefficient along bb:

\qquad a_{\mathrm{COIN}}=(G+I)^{-1}b.$$ When $b=0$, both are zero. The scalar control optimizes the common scale for the same local objective (Appendix J.3, Eq. (86)), while Full also adjusts the allocation direction. Independent, Scalar, and Full share candidate construction, calibration, correction schedule, and parameter lift. **Full has higher mean accuracy than Base, Independent, Scalar, and both single-direction controls on all four datasets.** The advantage over Independent is clearest on AVE, AVSBench, and FOOD101, while VGGSound50 is nearly tied (61.52% versus 61.50%). Theorem 4.4 quantifies the local objective benefit of retaining cross-module terms. Fig. 1/Table 7 also show **2.47–7.75% lower functional cost** at matched predicted progress, alongside different allocation directions. The new experiments extend this comparison to complete training. > Q2: How much training time and peak memory do COIN corrections add? **Independent, Scalar, and Full COIN have similar measured costs under the same correction schedule.** Profiling uses the same H100, three windows per method, and rotated method order. Time is normalized to Base; memory is peak allocated GPU memory. | Dataset | Independent time | Scalar time | Full time | Base memory (GiB) | Full memory (GiB) | |---|---:|---:|---:|---:|---:| | AVE | 1.455× | 1.450× | 1.453× | 6.930 | 7.056 | | AVSBench | 1.485× | 1.487× | 1.484× | 3.875 | 3.962 | | VGGSound50 | 1.068× | 1.071× | 1.070× | 3.875 | 3.962 | | FOOD101 | 1.260× | 1.284× | 1.269× | 1.525 | 2.539 | Independent, Scalar, and Full are within approximately **2%** of one another in profiled time on each dataset. The overhead relative to Base is larger: Full adds approximately **45.3%, 48.4%, 7.0%, and 26.9%**. Its additional peak memory is about **0.09–0.13 GiB** on the first three datasets and **1.014 GiB** on FOOD101. These measurements cover the optimizer computation associated with correction (Appendix H.4). Training length and correction schedule are held fixed across the compared variants. > W2 & Q3: How sensitive is final accuracy to candidate directions, calibration batch size, and correction frequency? Are the same settings used across datasets? **Complete-training sensitivity experiments evaluate candidate directions, correction frequency, and calibration batch size.** The single-candidate controls in the first table show higher mean accuracy when both the base delta and fresh fused gradient are retained. For frequency, we fix the fit size at 16 and vary the target number of corrections per epoch. All entries are three-seed mean test accuracy (%). | Target corrections/epoch | AVE | AVSBench | VGGSound50 | FOOD101 | |---|---:|---:|---:|---:| | 4 | 74.13 | 93.11 | **62.18** | 91.83 | | 8 (default) | **74.38** | **94.14** | 61.52 | **91.96** | | 16 | 74.13 | 94.05 | 61.38 | 91.91 | For calibration size, we retain the default frequency and vary the number of fit examples. | Fit examples | AVE | AVSBench | VGGSound50 | FOOD101 | |---|---:|---:|---:|---:| | 16 (default) | **74.38** | **94.14** | 61.52 | **91.96** | | 32 | 73.63 | 93.78 | **62.22** | 91.88 | | 64 | 71.39 | 93.24 | 61.90 | 91.92 | Across datasets, frequency sweeps have accuracy ranges of approximately **0.1–1.1 percentage points**. AVE is more sensitive to calibration size: its mean falls by about **3.0 points** from 16 to 64 fit examples. Neither a larger calibration batch nor more frequent correction consistently improves accuracy. **The shared default uses 16 fit examples and approximately eight corrections per epoch** (Appendix H.2), giving periods of **13, 14, 92, and 328 base updates** (Table 5). In the separate VGGSound50 sweeps, four corrections per epoch or 32 fit examples give higher mean accuracy. These experiments characterize final-accuracy sensitivity, complementing the local step-size study in Appendix J.7.

The authors addressed some of my concerns, so I raised my score from 2 to 3. However, I am still concerned about the instability of calibration batch size and correction frequency.

Thank you for reconsidering the paper and for raising the remaining sensitivity question. We have added the full-training study to Appendix J.8, Table 12 (p. 38). Importantly, one common default is used across all four datasets: 16 calibration examples and approximately eight corrections per epoch, without dataset-specific tuning.

Performance is relatively stable around this common configuration. Changing calibration size from 16 to 32 changes three-seed mean accuracy by no more than 0.75 percentage points across the four datasets. With calibration size fixed, varying correction frequency among 4, 8, and 16 per epoch yields ranges of just 0.13–1.03 points. The same default achieves the highest tested mean on three of four datasets in both sweeps. Table 12 also reports the larger calibration setting transparently; the results support stability near the shared default, not a claim that every setting is interchangeable.

The new Table 2 (p. 9) further shows that Full COIN has the highest mean among six complete-training variants on all four datasets, while Table 11 (p. 38) shows very similar measured optimizer time for Independent, Scalar, and Full. Together, these results distinguish the gain from joint allocation from changes in practical settings or correction cost. We hope this addresses the remaining concern.

Thanks, I get it.

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

This paper proposes COIN, a multimodal optimization framework that coordinates module updates in the fused-output space. By introducing functional invariance and minimum-energy allocation, COIN explicitly models cross-module interactions to reduce redundant updates. The authors provide theoretical guarantees and demonstrate improved fusion accuracy across four multimodal benchmarks.

Critical strengths
  1. The paper introduces a principled perspective on multimodal optimization by explicitly modeling cross-module interactions in the output space, supported by theoretical analyses of functional invariance and minimum-energy allocation.

  2. COIN achieves consistent improvements across four multimodal benchmarks.

Critical weaknesses
  1. Although the paper provides local analyses of joint allocation, it lacks end-to-end ablations comparing COIN with independent allocation under the same base training strategy. Consequently, the contribution of cross-module coordination to the final accuracy improvements remains unclear.

  2. The practical efficiency and scalability of COIN remain insufficiently evaluated. The method introduces additional gradient and Jacobian computations without clear empirical training-cost comparisons, and all experiments are limited to two modalities.

Questions for authors
  1. COIN models cross-module interactions through the alignment of their output responses, but does not explicitly capture how updating one modality changes the optimization dynamics of others. Consequently, it remains unclear whether the proposed coordination mechanism can directly mitigate modality competition, rather than merely improving the allocation of available updates.

  2. In Eq. (14), each module's candidate update space is restricted to the base optimizer update and the negative fused-task gradient. What justifies this two-direction restriction, and how sensitive is COIN's performance to the choice of candidate directions?

2 comments

W1: The paper lacks end-to-end comparisons with independent allocation under the same base training strategy.

We trained six variants end-to-end on all four datasets, reporting mean test accuracy (%) over three seeds. Full COIN performs best among these variants on every dataset.

MethodAVEAVSBenchVGGSound50FOOD101
Base73.3893.3860.7891.30
Independent73.1392.7061.5091.29
Scalar71.8993.6560.7891.42
Base direction only72.8993.5160.3491.26
Fused gradient only73.1393.2461.2291.31
Full COIN74.3894.1461.5291.96

All variants use the same architecture, supervision, optimizer, and training length. Base omits correction. Independent removes cross-module terms, giving a=b/2a=b/2; Scalar optimizes a common factor a=tba=tb (Eq. (86)); Full uses a=(G+I)−1ba=(G+I)^{-1}b. The single-direction variants retain the joint solve but use only one candidate. Full outperforms every control, although the margin over Independent is small on VGGSound50 (61.52 vs. 61.50). These complete-training comparisons complement the local interaction analysis in Theorem 4.4.

Table 1 also compares recent baselines—MoNet (ICML 2026), IBML (TPAMI 2026), DGL (ICCV 2025), and InfoReg (CVPR 2025). COIN has the highest reported fusion mean on all four benchmarks; its margins over the strongest baselines on AVSBench, VGGSound50, and FOOD101 are 1.08, 0.62, and 1.29 pp, respectively.

W2: Practical efficiency and scalability are insufficiently evaluated, and all experiments involve two modalities.

We measured optimizer time and peak allocated GPU memory on one H100, with three profiling windows per method and rotated method order. Time is normalized to Base.

DatasetFull / Base timeBase GiBFull GiB
AVE1.453×6.9307.056
AVSBench1.484×3.8753.962
VGGSound501.070×3.8753.962
FOOD1011.269×1.5252.539

Independent's time ratios are 1.455/1.485/1.068/1.260×, close to Full. The joint solve adds little cost beyond constructing candidate responses; fresh fused gradients and Jacobian–vector products dominate the overhead (Appendix H.4).

Q1: Does COIN directly mitigate modality competition, or only improve the allocation of available updates?

COIN coordinates available updates in fused-output space; unimodal supervision provides branch-specific learning signals. Table 2 distinguishes three coordination views:

Coordination viewWhat it addressesWhat remains to be determined
Modality-strength modulationRelative gradient magnitudes, learning strength, or information acquisitionMagnitude and speed do not determine whether output responses reinforce or counteract one another.
Objective-gradient reconciliationAgreement between unimodal and multimodal objectivesReconciling objectives does not by itself specify the joint output effects of updates to different parameter blocks.
Separated learning paths or timescalesPreserving module-specific supervision and learningAvailable updates still need to be combined effectively after fusion.

COIN uses both module–task and cross-module alignment (Eqs. (7)–(12)). Proposition 4.3 shows that a module with no standalone task alignment can still compensate for another module's response. In Fig. 2/Table 8, re-solving after removing a correction recovers about 70–84% of lost audit progress across fusion blocks.

Same-base masked-modality tests show higher visual/image branch accuracy:

DatasetBaseFull
AVE, video41.7943.03
AVSBench, video67.56869.19
VGGSound50, video42.3442.82
FOOD101, image58.0858.59

Here the other modality's encoded features are zeroed and predictions use the shared classifier. These results support improved branch predictions in the tested settings. Appendix G further clarifies that zero fused-task direction gives zero COIN correction.

Q2: Why use only the base update and the negative fused-task gradient?

They supply complementary information. The base delta carries supervision and optimizer history; the fresh negative fused gradient reflects the current fused objective after the base step (Eq. (14), Appendix H.2). The W1 ablation shows that using both candidates improves mean accuracy over either alone on all four datasets. Our derivation and lift (Eqs. (15)–(16)) apply to general finite candidate spaces; two directions offer a practical cost–capacity tradeoff.

Thank you for your feedback. The revised manuscript now includes the complete-training ablations and measured costs requested in your review.

Cross-module coordination (W1). Table 2 (p. 9) compares six variants with the same architecture, supervision, optimizer, and training length. Full COIN achieves the highest mean accuracy on all four datasets, improving over Independent by 1.25 points on AVE and 1.44 points on AVSBench. It also outperforms Scalar and both single-direction controls on every dataset. At matched predicted task progress, joint allocation reduces functional cost by 2.47–7.75% (Fig. 1/Table 8), showing a benefit beyond common step-size rescaling. Against external baselines in Table 1, COIN leads by 3.15, 1.08, 0.62, and 1.29 points across the four benchmarks.

Mechanism and candidates (Q1–Q2). COIN coordinates the fused-output effects of available module updates; it complements, rather than replaces, unimodal supervision. Proposition 4.3 identifies compensating responses that individual module statistics cannot capture. In Fig. 2/Table 9, re-solving recovers approximately 70–84% of the lost audit progress across fusion blocks. The two candidates—the actual base optimizer step and a fresh fused-task gradient—carry complementary information: Table 2 shows that using both outperforms either candidate alone on all four datasets. Table 13 additionally shows improvements in same-base masked visual/image predictions on all four tasks.

Cost (W2). Table 11 (p. 38) reports measured H100 optimizer time and GPU memory. Full takes 1.070–1.484× Base optimizer time, while its measured optimizer time differs from Independent by less than 1%. The joint solve uses at most 10 response coordinates; the additional work is dominated by gradient and Jacobian-response estimation shared by these corrected variants. Scalability beyond the evaluated architectures is identified as future work in Section 7.

We hope the revised experiments clarify the contribution and support a reassessment.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.