acceptodds
Under review as a conference paper at ICLR 2027

ElectrolyteFM: Unifying Electrolyte Property Prediction through Cross-Property Knowledge Learning

Abstract

Electrolyte formulation design requires balancing multiple physicochemical properties, yet existing models often focus on a limited subset. Learning each property in isolation can overlook transferable chemical information, whereas indiscriminate sharing can introduce cross-property interference. Our directed transfer analysis shows that jointly learning two property prediction tasks can improve or degrade prediction relative to separate training, with asymmetric transfer effects between the tasks. We propose ElectrolyteFM, a unified multi-property prediction model which can more accurately predict multiple properties of each electrolyte by effectively identifying and utilizing property-specific features and knowledge shared across properties. More specifically, ElectrolyteFM learns property-specific representations independently and captures cross-property knowledge through a separately trained expert pool. A router selects relevant shared information for each formulation and target property, and property-specific residual adapters convert this information into corrections to the corresponding representation for prediction. Experiments on Electrolyte12 show that ElectrolyteFM reduces normalized mean absolute error averaged across 12 electrolyte properties by 14.8% relative to the strongest electrolyte-specific baseline. On an independent sodium-electrolyte dataset unseen during training, it reduces conductivity mean absolute error by 6.7% relative to the best-performing baseline.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Electrolyte property prediction, multi-task learning, cross-property transfer, mixture-of-experts, adaptive knowledge fusion

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_a89NOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes ElectrolyteFM for predicting multiple electrolyte properties. It combines property-specific experts, a shared expert pool, formulation–property-conditioned routing, and residual adapters to exploit cross-property information while mitigating negative transfer. Experiments on Electrolyte12 and an independent sodium-electrolyte dataset demonstrate competitive prediction performance and promising generalization.

Critical strengths
  1. The paper tackles an important problem in electrolyte formulation design by developing a unified framework for predicting multiple physicochemical properties.
  2. The directed cross-property transfer analysis reveals beneficial, harmful, and asymmetric transfer effects among the 12 electrolyte properties, which provides empirical motivation for selective knowledge sharing.
  3. ElectrolyteFM integrates property-specific experts, a shared expert pool, formulation–property-conditioned routing, and residual adapters to combine shared and property-specific knowledge.
  4. Experiments on Electrolyte12 and an independent sodium-electrolyte dataset show competitive multi-property prediction performance and promising zero-shot generalization, with few-shot adaptation further improving predictive performance.
Critical weaknesses
  1. ElectrolyteFM, although designing exclusive branches to store common knowledge and property-specific knowledge respectively, still enables these branches to learn redundant representation automatically. One cannot tell how much these branches are learning complementary representation instead of redundant one.
  2. SpKL and ShKL pursue individual objectives and contain individual differentiable parts, which makes disentangling individual property learning and cross-property knowledge sharing challenging.
  3. Such directed transfer scores on the test set encourage designing the following framework. However, such results may not indicate how to design our model and select hyperparameters, which could lead to test-set leakage.
  4. Such ablation studies are unable to well illustrate how our formulation–property-conditioned router can minimize negative transfer.
  5. The multi-task learning bases MMoE (2018) and PLE (2020) may not represent the state-of-the-art methods. Adding state-of-the-art techniques as baselines would make our comparison more complete.
Questions for authors
  1. Further analysis is needed to determine whether the shared and property-specific branches learn complementary representations or merely redundant ones.
  2. Controlled experiments under identical training conditions are needed to isolate the respective contributions of SpKL and ShKL.
  3. The authors should clarify whether the directed transfer scores computed on the test set influenced model design or hyperparameter selection, and explain how potential test-set leakage was avoided.
  4. Comparisons with fixed routing and property-only routing are needed to better establish that the formulation–property-conditioned router effectively mitigates negative transfer.
  5. Comparisons with more recent multi-task learning methods are needed to better establish the advantages of ElectrolyteFM over existing approaches.
6 comments

Thank you for the review. We address your concerns through additional representation analyses, controlled ablations, and comparisons with recent multi-task learning methods. Unless otherwise specified, results are reported as mean ± SD over five seeds.

Q1: Representation Complementarity (W1/Q1)

A1: Thank you for the question. We conducted a linear CKA analysis on the same validation samples:

Representation ComparisonLinear CKA
Private vs. Shared0.508–0.818
Private (Cross-Seed)0.852–0.985
Shared (Cross-Seed)0.799–0.967

These results indicate representational differences while retaining some overlap. Importantly, our objective is predictive complementarity rather than strictly disentangled representations. The full model outperforms either branch alone (Table 4), suggesting that combining the two representations provides additional predictive value beyond either individual branch.

Together, the CKA analysis and predictive ablations support the usefulness of learning and integrating property-specific and shared knowledge.

Q2: Contributions of SpKL and ShKL (W2/Q2)

A2: Thank you for the question. As detailed in Sections 3.3–3.5, SpKL independently learns property-specific representations from single-property supervision, while ShKL learns shared experts using multi-property supervision. During S2SKF, both learned branches are frozen, and only the residual adapters and prediction heads are optimized. This design separates the optimization of property-specific and shared knowledge without requiring strictly disentangled representations.

We evaluate their individual and combined predictive contributions through branch ablations. Table 4 shows that the full model outperforms either branch alone. We further repeated this comparison over five seeds (Appendix G.2, Table 15):

ConfigurationMacro NMAE ↓
SpKL Only0.1857 ± 0.0022
ShKL Only0.1654 ± 0.0018
Full ElectrolyteFM0.1569 ± 0.0002

Relative to STL, their Macro NMAE changes are +11.25%, −0.90%, and −5.99%, respectively. Although ShKL alone improves aggregate performance, the full model achieves further gains and outperforms STL on 9/12 properties, compared with 4/12 for ShKL. This indicates that shared-to-specific fusion benefits a broader range of properties rather than merely improving the aggregate metric.

To further examine the staged learning strategy, we compare it with direct joint optimization under the same final architecture. Even after 60 epochs, joint optimization underperforms staged training on both ID (0.1615 vs. 0.1569) and OOD (3.397 vs. 2.591 mS/cm).

Unlike joint optimization, our approach preserves independently learned property-specific representations while integrating shared knowledge through frozen residual fusion. These results support the contributions of both branches and the effectiveness of staged learning.

Please refer to Section 4.4.2 (Table 4) and Appendix F.2 (Table 12) for more details. We have added the above new five-seed results in Appendices G.2–G.3 (Tables 15–16).

Q3: Directed Transfer Analysis and Potential Test-Set Leakage (W3/Q3)

A3: Thank you for raising this concern. We clarify the role of the directed transfer analysis and its relationship to ElectrolyteFM's development.

  • All Electrolyte12 experiments follow the same fixed Train/Val/Test splits and model-selection protocol. Each model is trained on Train, tuned and early-stopped on Val, and evaluated on held-out Test. This ensures consistent data usage and comparable evaluation across methods.
  • The directed transfer analysis in Figure 1 and Section 3.1 follows the same protocol. All 78 diagnostic models (12 single-property and 66 pairwise models) are trained on Train and selected using Val (Appendix C.1). Reporting transfer scores on held-out Test therefore avoids validation-set model-selection bias in evaluating these diagnostic models and provides an independent characterization of cross-property transfer.
  • The transfer analysis motivates ElectrolyteFM at the high-level architectural level. The coexistence of beneficial, harmful, and asymmetric transfer suggests that indiscriminate sharing may introduce interference, motivating separate property-specific and shared knowledge learning followed by selective residual fusion.
  • Importantly, the numerical transfer scores are not used to configure the model, determine expert allocation, or select hyperparameters. ElectrolyteFM is trained and selected independently, and its routing weights are learned from training data rather than derived from the transfer matrix.
  • Moreover, ElectrolyteFM achieves the lowest mean prediction error on an independent sodium-electrolyte OOD dataset entirely excluded from the transfer analysis. This provides additional evidence that the framework generalizes beyond the Electrolyte12 test set.

Q4: Routing and Negative Transfer (W4/Q4)

A4: Thank you for the question. We clarify that negative-transfer mitigation arises from the combined SpKL–ShKL–S2SKF framework, rather than routing alone. SpKL preserves independently learned property-specific knowledge, ShKL captures reusable cross-property information, and S2SKF integrates them through frozen residual fusion. The grouped results in Table 11 provide supporting evidence.

To isolate the router's contribution, we additionally compare four routing strategies:

Routing StrategyMacro NMAE ↓
Formulation + Property (Ours)0.1569 ± 0.0002
Formulation-Only0.1579 ± 0.0014
Property-Only0.1592 ± 0.0014
Uniform Weights0.1586 ± 0.0017

Property-only uses softmax⁡(bt)\operatorname{softmax}(b_t), while uniform routing uses fixed weights 1/E1/E. All strategies follow the same data splits and validation-based model selection protocol.

The full formulation–property-conditioned router achieves the lowest mean error. Notably, property-only routing does not outperform uniform routing, whereas incorporating formulation information yields better average performance. This supports the value of input-dependent expert allocation beyond fixed property preferences.

Thus, we attribute negative-transfer mitigation to the overall knowledge learning and fusion framework, while the routing ablation demonstrates the additional predictive benefit of formulation–property conditioning.

We have revised our paper and added the above results to Appendix G.5 (Table 18), including routing protocols; grouped results are included in Appendix F.1 (Table 11).

Q5: Recent Multi-Task Baselines (W5/Q5)

A5: Thank you for the suggestion. MMoE and PLE are representative expert-sharing architectures, and our original paper already includes recent electrolyte-specific prediction methods. To broaden the multi-task comparison, we additionally evaluate FairGrad (ICML 2024) and PIVRG (CVPR 2025), which address multi-task optimization from a different perspective.

MethodMacro NMAE ↓OOD MAE (mS/cm) ↓
STL0.1669 ± 0.00143.233 ± 0.357
MMoE (2018)0.1627 ± 0.00133.407 ± 0.652
FairGrad (2024)0.1610 ± 0.00263.428 ± 0.701
PIVRG (2025)0.1618 ± 0.00233.491 ± 0.789
PLE (2020)0.1611 ± 0.00012.731 ± 0.080
ElectrolyteFM (Ours)0.1569 ± 0.00022.591 ± 0.227

ElectrolyteFM achieves the lowest mean error on both metrics, outperforming the additional recent multi-task baselines on Electrolyte12 and maintaining competitive generalization on the independent sodium-electrolyte dataset.

Across five paired seeds, its Macro NMAE improvement over FairGrad is 0.00409 ± 0.00277, providing further evidence of its average predictive advantage across different random seeds.

These comparisons extend our evaluation beyond classical expert-sharing architectures and support the effectiveness of staged knowledge learning and fusion relative to recent multi-task optimization approaches.

We have revised our paper and added the above results to Appendix G.1 (Table 14).

References

  • Ma et al. Modeling Task Relationships in Multi-task Learning with Multi-gate Mixture-of-Experts. KDD, 2018.
  • Tang et al. Progressive Layered Extraction (PLE). RecSys, 2020.
  • Ban & Ji. Fair Resource Allocation in Multi-Task Learning. ICML, 2024.
  • Qin et al. Revisiting Fairness in Multitask Learning: A Performance-Driven Approach for Variance Reduction. CVPR, 2025.

Dear Reviewer,

As the rebuttal and discussion process draws to a close, we would like to emphasize the importance of addressing any outstanding points before the deadline. Please ensure that the authors’ responses have been considered and share any final observations while the discussion remains open.

Thank you for your valuable time and engagement.

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

This paper proposes ElectrolyteFM, a unified multi-property prediction model that predicts multiple properties of each electrolyte by learning both shared and property-specific knowledge. More specifically, ElectrolyteFM learns property-specific representations independently and captures cross-property knowledge through a separately trained expert pool. Empirical experiments show that ElectrolyteFM exhibits very competitive performance across 12 electrolyte properties prediction.

Critical strengths
  1. This method achieves very strong predictive performance compared to many of the electrolyte-specific baseline models. As shown in Table 1, the proposed model achieves the best MAE on 10/12 properties. The improvements on conductivity prediction and excess molar enthalpy are particularly large. The OOD evaluation results from the independent sodium-electrolyte dataset are also very impressive.

  2. The authors claim that the model can "effectively identify and utilize property-specific features and knowledge shared across properties", which is a compelling idea, as in nature different properties are different aspects of the same entity. So they naturally have what are unique to each of them, and what are common and foundational. The problem is that combining these can bring both bonus and noises, depending on how you model them. I really appreciate that the authors have done very thorough analysis to support this to support this idea, including the directed transfer analysis, which shows both positive and negative transfer between properties, the ablation studies, and how different properties each relies more heavily on different shared experts (as shown in Figure 4).

Critical weaknesses

As I mentioned in the strengths section, the proposed model achieves strong performance compared to electrolyte-specific models. However, the main idea of ElectrolyteFM is closely related to the existing multi-task learning methods (such as MMoE and PLE), which also use shared and task-specific representations. In Table 3, the authors compare ElectrolyteFM with STL and existing multi-task learning strategies, including MMoE and PLE, and the improvement ElectrolyteFM brings in this aspect is relatively small.

I do not think the evaluation, in this aspect, is thorough enough to give a complete picture of how good the proposed learning strategy works. The results in Table 3 are reported as average across all 12 properties. I saw that Appendix Table 11 provides the performance of experimental and computed properties, but the individual performance results for each property is not present in the paper. Also, most of the main experimental results do not provide standard deviations, confidence intervals, or any other measures of uncertainty. As the improvements over multi-task methods are relatively small, I think it is important to evaluate how robust the results are across different random seeds.

Although Table 4 compares the private branch, shared branch, and complete model, the experimental results from directly evaluating the benefit of freezing the two branches during fusion are not presented in the paper.

Questions for authors
  1. How can the robustness of the reported performance improvements be assessed given the relatively small improvements over existing multi-task learning methods?

  2. What are the individual performance results like for each property under STL and different multi-task learning strategies?

  3. What is the benefit of separately training and freezing the property-specific and shared branches? Can the authors compare ElectrolyteFM with a jointly trained version without separately training and freezing the branches?

4 comments

Thank you for your positive evaluation of ElectrolyteFM's predictive performance, OOD generalization, and directed transfer and expert-routing analyses. We address your concerns through additional five-seed evaluations, per-property analyses, and controlled training-strategy ablations. Unless otherwise specified, results are reported as mean ± SD.

Q1: Performance Improvements Compared to Other Multi-task Learning Methods and Statistical Robustness (W1, W2 / Q1)

A1: Thank you for the question. Unlike MMoE and PLE, ElectrolyteFM uses separate property-specific experts and a shared expert pool with formulation–property-conditioned routing. During training, the private and shared branches are learned independently and then frozen for residual fusion, rather than jointly optimized.

We evaluated STL, MMoE, PLE, and ElectrolyteFM across five random seeds, including OOD testing on the same 106 independent sodium-electrolyte formulations.

MethodID Macro NMAE ↓OOD MAE (mS/cm) ↓
STL0.1669 ± 0.00143.233 ± 0.357
MMoE0.1627 ± 0.00133.407 ± 0.652
PLE0.1611±0.0001‾\underline{0.1611 \pm 0.0001}2.731±0.080‾\underline{2.731 \pm 0.080}
ElectrolyteFM (Ours)0.1569 ± 0.00022.591 ± 0.227
  • Across five matched seeds, ElectrolyteFM reduces ID macro NMAE relative to PLE by 0.00424 ± 0.00014 (mean ± SD of paired differences; 95% CI: [0.00407, 0.00442]). The confidence interval excludes zero, indicating that the improvement is consistent across training runs despite its modest magnitude.
  • On the independent OOD dataset, ElectrolyteFM also achieves the lowest mean MAE compared with other multi-task methods, improving over PLE by 5.1%. Compared with the baseline SCAN (2.964±0.1612.964 \pm 0.161 mS/cm), ElectrolyteFM (2.591±0.2272.591 \pm 0.227 mS/cm) reduces mean MAE by 0.373 mS/cm (12.6%), with a paired bootstrap 95% CI of [0.048, 0.792] mS/cm.

Together, these results provide additional evidence for the reproducibility of ElectrolyteFM's ID improvements and its generalization to unseen electrolyte formulations.

We have revised our paper and added the new experimental results to Appendix G.1 (Table 14) and Appendix H.1–H.2 (Tables 19–20).

Q2: Per-Property Performance and Negative Transfer (W1, W2 / Q2)

A2: Thank you for the suggestion. We added five-seed MAE comparisons for all 12 properties across STL, MMoE, PLE, and ElectrolyteFM, using identical property-wise test samples.

PropertySTLMMoEPLEElectrolyteFM
Experimental labels
Cond. (mS/cm)1.213 ± 0.0100.993 ± 0.0520.922±0.039‾\underline{0.922 \pm 0.039}0.895 ± 0.008
Visc. (mPa·s)121.239 ± 12.195109.193 ± 20.93189.095 ± 1.18792.443±6.780‾\underline{92.443 \pm 6.780}
HEH^E (J/mol)91.401 ± 10.65564.087 ± 6.24157.061 ± 1.26960.897±8.146‾\underline{60.897 \pm 8.146}
VEV^E (cm³/mol)0.065 ± 0.0020.052 ± 0.0020.048±0.001‾\underline{0.048 \pm 0.001}0.047 ± 0.002
Computed labels
Dens. (g/cm³)0.0061 ± 0.00030.0066 ± 0.00020.0059±0.0001‾\underline{0.0059 \pm 0.0001}0.0052 ± 0.00004
D+D_{+} (10⁻¹¹ m²/s)4.160 ± 0.0754.391 ± 0.0674.427 ± 0.0354.246±0.102‾\underline{4.246 \pm 0.102}
D−D_{-} (10⁻¹¹ m²/s)3.897 ± 0.0583.797±0.092‾\underline{3.797 \pm 0.092}3.896 ± 0.1073.680 ± 0.043
DsD_s (10⁻¹¹ m²/s)9.020 ± 0.0868.860±0.146‾\underline{8.860 \pm 0.146}9.051 ± 0.1088.713 ± 0.171
CN+−\mathrm{CN}_{+-}0.942±0.003‾\underline{0.942 \pm 0.003}0.953 ± 0.0100.958 ± 0.0030.936 ± 0.004
CN+s\mathrm{CN}_{+s}1.802 ± 0.0101.823 ± 0.0401.886 ± 0.0191.809±0.018‾\underline{1.809 \pm 0.018}
Oxid. stab. (eV)0.215±0.003‾\underline{0.215 \pm 0.003}0.220 ± 0.0030.207 ± 0.0030.217 ± 0.006
HvapH_{\mathrm{vap}} (J/mol)483.3±15.3‾\underline{483.3 \pm 15.3}548.7 ± 20.3495.7 ± 5.8471.1 ± 7.3
Macro NMAE ↓0.1669 ± 0.00140.1627 ± 0.00130.1611±0.0001‾\underline{0.1611 \pm 0.0001}0.1569 ± 0.0002
Best count ↑2/120/123/127/12
  • ElectrolyteFM achieves the best mean MAE on 7/12 properties and outperforms PLE on 9/12, demonstrating improvements across multiple properties rather than a single dominant target. Notably, it improves predictions for density, anion and solvent diffusivity, and ion coordination, although PLE remains stronger on viscosity, excess molar enthalpy, and oxidation stability.
  • We further examine performance degradation relative to STL. MMoE and PLE degrade on 6 and 5 properties, respectively, whereas ElectrolyteFM degrades on only 3 properties. This is consistent with the heterogeneous positive and negative transfer effects observed in Figure 1, and suggests that ElectrolyteFM better balances shared knowledge utilization and property-specific prediction, helping alleviate cross-property interference.

We have revised our paper and added the above results to Appendix H.1 (Table 19).

Q3: Benefits of Staged Pretraining and Frozen Fusion (W1, W3 / Q3)

A3: Thank you for the suggestion. ElectrolyteFM separately learns private and shared representations, then integrates routed shared knowledge through property-specific residual adapters while keeping both branches frozen. This design aims to incorporate cross-property information without directly modifying the independently learned private representations during fusion.

To isolate the benefits of staged learning and freezing, we compare three strategies with identical final architectures and the same pretrained encoder initialization:

  • A. Pretrained + Frozen (Ours): Independently pretrain private and shared branches, then freeze both during residual fusion.
  • B. Pretrained + Joint: Retain independent branch pretraining, but jointly fine-tune both branches and fusion modules.
  • C. Direct Joint Training: Skip separate branch pretraining and directly optimize the complete architecture.

A and B use 40 branch-pretraining epochs (20 SpKL + 20 ShKL) followed by 20 fusion epochs. C is evaluated with 20 and 60 epochs to examine whether longer direct joint optimization closes the performance gap. All strategies use the same data splits and validation-based model selection. Results are averaged over five seeds.

Training StrategyID Macro NMAE ↓OOD MAE (mS/cm) ↓Trainable Params (Final Stage)
A: Pretrained + Frozen (Ours)0.1569 ± 0.00022.591 ± 0.227595,212
B: Pretrained + Joint0.1574±0.0019‾\underline{0.1574 \pm 0.0019}3.359±0.458‾\underline{3.359 \pm 0.458}3,384,104
C: Joint (20 epochs)0.1630 ± 0.00123.503 ± 0.1933,384,104
C: Joint (60 epochs)0.1615 ± 0.00103.397 ± 0.1383,384,104
  • Benefit of independent pretraining (B vs. C): Strategy B achieves lower ID macro NMAE than direct joint training for 60 epochs (0.1574 vs. 0.1615). Even with additional training epochs, C does not achieve B's performance. This suggests that independently pretrained representations provide a more effective starting point for subsequent fusion than direct joint optimization.
  • Benefit of freezing (A vs. B): Freezing the pretrained branches further improves mean ID performance (0.1569 vs. 0.1574) and substantially improves OOD generalization, reducing MAE from 3.359 to 2.591 mS/cm. Moreover, B exhibits approximately 9.5 times the ID cross-seed standard deviation of A, while A requires 82.4% fewer trainable parameters during the final stage.

These comparisons support the effectiveness of ElectrolyteFM's staged learning strategy. Its advantages are not solely attributable to combining private and shared experts, but also to learning their representations separately and integrating them through frozen residual fusion. This strategy preserves the learned branches during fusion and achieves stronger predictive performance with fewer final-stage trainable parameters.

We have revised our paper and added the above results to Appendix G.3 (Table 16).

References

  1. MMoE (KDD 2018). Ma et al. Modeling Task Relationships in Multi-task Learning with Multi-gate Mixture-of-Experts. Proceedings of KDD, pp. 1930–1939, 2018.
  2. PLE (RecSys 2020). Tang et al. Progressive Layered Extraction (PLE): A Novel Multi-Task Learning (MTL) Model for Personalized Recommendations. Proceedings of RecSys, pp. 269–278, 2020.

Dear Reviewer,

As the rebuttal and discussion process draws to a close, we would like to emphasize the importance of addressing any outstanding points before the deadline. Please ensure that the authors’ responses have been considered and share any final observations while the discussion remains open.

Thank you for your valuable time and engagement.

Dear Reviewers,

Thank you for your valuable feedback. We have carefully addressed all your concerns and provided additional experiments and analyses in our rebuttal.

As the discussion period approaches its end, we would greatly appreciate it if you could review our responses and let us know whether any concerns remain. We would be happy to provide further clarification if needed.

Thank you again for your time and consideration.

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

ElectrolyteFM addresses prediction of 12 electrolyte properties under heterogeneous, partially observed supervision. Its contributions are a directed cross-property transfer analysis, the curated Electrolyte12 benchmark, and a model combining independently trained property-specific experts with a separately trained shared expert pool. Property- and formulation-conditioned routing selects shared features, which residual adapters incorporate into frozen private representations. Validation includes six predictive baselines, STL/MMoE/PLE comparisons, branch and expert ablations, and external sodium-electrolyte conductivity prediction with few-shot adaptation. The model reports the best MAE on 10/12 properties, a 14.8% macro NMAE reduction over Bamboo-Mixer, and a 6.7% external conductivity MAE reduction over SCAN.

Critical strengths

+Broad evaluation with encouraging practical results. The model improves all four experimentally evaluated properties and performs well across eight computed targets (Table 1). The independent 106-formulation sodium-electrolyte evaluation provides evidence beyond the main benchmark, while the few-shot protocol selects regularization using ID validation data rather than external query labels. These results support the approach’s practical promise.

  • A coherent approach to selective sharing, evaluated against relevant alternatives. Separately learning private and shared representations, then fitting zero-initialized residual adapters, provides a sensible way to retain task-specific representations while incorporating shared information. Comparisons with STL, MMoE, and PLE strengthen the evaluation, and the complete model outperforms either branch alone (Tables 3–4).
Critical weaknesses
  • The ablations do not isolate the distinctive methodological contribution. Tables 4 and 12 compare private-only, shared-only, and full models, but do not compare residual fusion with simple concatenation or prediction ensembling, or staged freezing with otherwise comparable joint training. Expert removal establishes that the fitted model uses each expert; it does not establish that the proposed routing and fusion are necessary to reduce negative transfer. Since the contribution centers on this training and fusion arrangement, these controls materially affect its significance.

  • The robustness of the most relevant improvements is unclear. The advantage over PLE is 2.5% in macro NMAE, while the external advantage over SCAN is 0.186 mS/cm on 106 formulations (Tables 2–3). Neither comparison includes reported training-seed variability or confidence intervals. Figure 3 reports variability across support/query episodes, which does not establish robustness to model training randomness. The experimental enthalpy and excess-volume test sets also contain only 62 and 102 observations, respectively. These limitations make it difficult to assess how reliably the reported improvements would reproduce.

Questions for authors
  • Requested validation: Can you report repeated-seed results for ElectrolyteFM and PLE, together with paired uncertainty estimates for their performance difference and for the external conductivity comparison against SCAN?

  • Requested mechanism evidence: Does the advantage persist against simple private/shared concatenation or prediction ensembling, and against a comparable jointly trained fusion model? Please specify parameter counts, training budgets, and model-selection procedures for these comparisons.

4 comments

Thank you for the review. We appreciate the recognition of our broad evaluation and selective knowledge-sharing design. To address the concerns about statistical robustness (W2/Q1) and methodological contributions (W1/Q2), we conduct additional five-seed evaluations and controlled ablations. Results are reported as mean ± SD unless otherwise specified.

Q1: Statistical Robustness (W2/Q1)

A1: Thank you for your question. We extend our evaluation with five random seeds and paired confidence intervals to assess whether the reported improvements are consistent across training runs and test formulations.

(a) ID performance (five seeds).

ModelMacro NMAE ↓
STL0.1669 ± 0.0014
PLE0.1611 ± 0.0001
ElectrolyteFM (Ours)0.1569 ± 0.0002

ElectrolyteFM achieves the lowest macro NMAE with small cross-seed variability. The paired seed-wise improvement over PLE is 0.00424 ± 0.00014 (95% CI: [0.00407, 0.00442]). The confidence interval excludes zero, providing further evidence that the improvement is robust to training randomness. At the property level, ElectrolyteFM outperforms PLE on 9/12 properties, including 7/8 computed properties, indicating that the overall improvement is not confined to a single prediction target.

(b) External OOD conductivity (106 formulations, five seeds).

MethodMAE ↓ (mS/cm)MAE Reduction95% CI
SCAN2.964 ± 0.161––
ElectrolyteFM2.591 ± 0.2270.373[0.048, 0.792]

On the independent sodium-electrolyte dataset, ElectrolyteFM maintains a 12.6% lower MAE than SCAN on average across five seeds. The paired bootstrap confidence interval remains positive, providing additional evidence that the external generalization improvement is not solely attributable to a particular training seed or test formulation.

(c) Small experimental test sets (vs. Bamboo-Mixer).

To address the limited test-set sizes for excess enthalpy and excess volume, we further quantify uncertainty in their MAE improvements relative to Bamboo-Mixer.

PropertyNMAE ReductionRelative Reduction95% CI
HEH^E (J/mol)6280.757.00%[54.3, 109.8]
VEV^E (cm³/mol)820.016526.10%[0.0098, 0.0239]

Both confidence intervals exclude zero, supporting the reliability of these improvements despite the limited number of independent formulations.

Uncertainty estimation. For the bootstrap analyses in (b) and (c), we resample both training seeds and formulation groups while preserving paired model comparisons. Repeated measurements from the same formulation are grouped during resampling to avoid treating correlated observations as independent. Here, N denotes independent formulation groups; the VEV^E test set contains 102 records from 82 formulations. This procedure accounts for uncertainty arising from both training randomness and test-set composition.

We have revised our paper and added the above results to Appendix G.1 (Table 14) and Appendix H.1–H.2 (Tables 19–21).

Q2: Methodological Contributions (W1/Q2)

A2: Thank you for the valuable suggestions. We conduct three additional ablations examining fusion alternatives, staged freezing, and routing strategies.

(a) Fusion alternatives.

We compare S2SKF against simple feature concatenation and prediction ensembling using identical frozen SpKL/ShKL representations. This controls for the upstream representations and allows us to examine whether the proposed residual fusion provides benefits beyond simply combining private and shared predictions or features.

Fusion MethodMacro NMAE ↓Additional Params
Uniform Ensembling0.1660 ± 0.00160
Validation-Weighted Ensembling0.1615 ± 0.000612
Concatenation0.1610 ± 0.0010592,140
S2SKF (Ours)0.1569 ± 0.0002595,212

S2SKF achieves the lowest macro NMAE, outperforming both prediction-level ensembling and feature concatenation. Importantly, concatenation and S2SKF have comparable additional parameter counts, suggesting that the improvement cannot be explained merely by increased fusion capacity.

For a fair comparison, the trainable fusion methods use up to 20 epochs with validation-based early stopping. Prediction ensembling requires no additional model training: uniform ensembling averages branch predictions, while validation-weighted ensembling selects weights using validation data.

These results support the effectiveness of the proposed representation-level residual fusion over simpler feature- and prediction-level alternatives.

We have revised our paper and added the above results to Appendix G.4 (Table 17).

(b) Staged freezing vs. joint optimization.

To investigate whether the proposed staged training and freezing are necessary, we compare three strategies with identical final architectures and pretrained encoder initialization:

  • A (Ours): Pretrain SpKL and ShKL separately, then freeze their representations and train the fusion modules.
  • B (Pretrained + Joint): Use the same pretrained SpKL/ShKL initialization as A, but jointly fine-tune the upstream branches and fusion modules.
  • C (Joint from Scratch): Skip separate branch pretraining and jointly optimize the architecture from the start.

A and B use 40 pretraining + 20 fusion epochs. C is evaluated with both 20 and 60 epochs to examine whether additional joint-training epochs can close the performance gap. All strategies use the same data splits and validation-based model selection.

StrategyMacro NMAE ↓OOD MAE (mS/cm) ↓Trainable Params (Final Stage)
A: Pretrained + Frozen (Ours)0.1569 ± 0.00022.591 ± 0.227595,212
B: Pretrained + Joint0.1574 ± 0.00193.359 ± 0.4583,384,104
C: Joint (20 epochs)0.1630 ± 0.00123.503 ± 0.1933,384,104
C: Joint (60 epochs)0.1615 ± 0.00103.397 ± 0.1383,384,104

The staged approach achieves the best ID and OOD performance. Comparing A and B, freezing the pretrained branches substantially improves external OOD conductivity prediction (2.591 vs. 3.359 mS/cm) while requiring 82.4% fewer final-stage trainable parameters. It also exhibits lower cross-seed variability.

Furthermore, direct joint optimization performs worse even when extended to 60 epochs, indicating that the advantage of staged training is not simply explained by a longer nominal training schedule.

These findings suggest that separately learning and subsequently freezing private/shared representations helps preserve useful pretrained knowledge during fusion, particularly for generalization to unseen electrolyte distributions.

We have revised our paper and added the above results to Appendix G.3 (Table 16).

(c) Routing ablation.

To isolate the contribution of the routing design, we compare four strategies while retraining ShKL and the subsequent fusion modules under each condition.

Routing StrategyMacro NMAE ↓
Formulation + Property (Ours)0.1569 ± 0.0002
Formulation-Only0.1579 ± 0.0014
Property-Only0.1592 ± 0.0014
Uniform Weights0.1586 ± 0.0017

The proposed formulation-property conditioned routing achieves the lowest macro NMAE and smallest cross-seed variability. Its advantage over formulation-only, property-only, and uniform routing supports the benefit of jointly considering formulation characteristics and target-property information when selecting shared experts.

These findings provide evidence for the routing design within the broader framework, where private/shared knowledge learning, conditional routing, and residual fusion work together to enable selective knowledge sharing.

We have revised our paper and added the above results to Appendix G.5 (Table 18).

Dear Reviewer,

As the rebuttal and discussion process draws to a close, we would like to emphasize the importance of addressing any outstanding points before the deadline. Please ensure that the authors’ responses have been considered and share any final observations while the discussion remains open.

Thank you for your valuable time and engagement.

Dear Reviewers,

Thank you for your valuable feedback. We have carefully addressed all your concerns and provided additional experiments and analyses in our rebuttal.

As the discussion period approaches its end, we would greatly appreciate it if you could review our responses and let us know whether any concerns remain. We would be happy to provide further clarification if needed.

Thank you again for your 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.