acceptodds
Under review as a conference paper at ICLR 2027

Energy Reinforced Diffusion for Structured Representation from Heterogeneous Spectral Observations

Abstract

Heterogeneous spectral observations provide high dimensional molecular signals for knowledge discovery and structured evidence learning. However, global representations often obscure informative relationships among local peak groups, while insufficient structural guidance makes these dependencies difficult to preserve during evidence exploration and policy optimization. Consequently, learned representations may rely on isolated peaks or background patterns, rather than informative group dependencies with consistent structural semantics. To this end, we propose Knowledge-Energy Reinforced Diffusion (KERD) for structured evidence learning from heterogeneous data. KERD organizes peak clusters into a Peak Graph and learns structured relations among band groups through data driven consistency modeling. The learned knowledge energy guides diffusion sampling toward band group combinations with stable structural semantics. Next, we construct class conditional spectral prototypes from peak cluster relations to improve the stability of policy learning. When paired cohort variables are available, KERD optionally refines the structured state through a shared relation graph that strengthens the alignment between peak substructures and auxiliary factor groups. Finally, we evaluate KERD on more than 70,000 measurements from multicenter cohorts. KERD achieves consistent gains across the evaluated heterogeneous datasets covering different sources and biological sample types.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Heterogeneous Data Mining, Structured Evidence Learning, Energy Guided Diffusion, Peak Graph Modeling

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

This paper proposes Knowledge-Energy Reinforced Diffusion (KERD) for structured representation learning from heterogeneous spectral observations. The framework integrates Peak Graph modeling, energy-guided diffusion, and relation-consistent policy optimization to capture dependencies among spectral groups. An optional cohort graph incorporates auxiliary clinical information. Experiments on seven datasets comprising 72,057 spectral measurements demonstrate promising classification performance and robustness under different observation conditions.

Critical strengths

1.The paper is generally well organized, but the overall framework is relatively complex, and some figures are overly dense. The interactions between diffusion guidance and policy optimization could be explained more clearly.

2.The integration of structural relations into diffusion-based evidence selection provides an interesting perspective, although the methodological insights beyond existing graph learning and energy-guided optimization remain limited.

Critical weaknesses

1.Insufficient evidence for generalization under heterogeneous conditions. Although the paper evaluates seven datasets, each dataset is trained and tested independently, which does not directly establish cross-source or cross-center generalization. The robustness experiments mainly consider synthetic perturbations rather than realistic distribution shifts between acquisition devices, institutions, or populations. Furthermore, the additional computational cost of diffusion sampling and policy optimization is not fully evaluated. These limitations weaken the practical significance of the claimed robustness and efficiency.

2.Unclear necessity of diffusion-based reinforcement learning. Although component ablations are provided, the paper does not convincingly demonstrate why the relatively complex diffusion and reinforcement learning framework is necessary for spectral group selection. Comparisons with simpler differentiable feature selection, attention-based masking, or direct energy optimization under matched computational budgets would strengthen this claim.

3.Incomplete computational efficiency analysis. KERD involves graph encoding, iterative diffusion sampling, energy guidance, and policy optimization, potentially introducing substantial computational overhead. Although online inference latency is discussed, comprehensive comparisons of training time, memory consumption, and end-to-end inference cost against competing methods are lacking.

Questions for authors
  1. What are the fundamental methodological innovations of KERD beyond combining existing graph learning, energy-guided diffusion, and policy optimization techniques?

2.Why is diffusion-based policy optimization necessary for spectral group selection? Could the authors compare KERD with simpler differentiable selection or attention-based methods under comparable model capacity and computational budgets?

3.Can the authors provide cross-center, cross-device, or leave-one-domain-out evaluations to demonstrate generalization under realistic distribution shifts?

4 comments
AuthorsOpenReview ↗

Response to Reviewer ELNx [part 1]

First, we would like to express our deepest respect. Thank you for carefully reviewing our work.

We sincerely appreciate your positive evaluation and recognition of our study.

★ We greatly value your suggestions, and we will address your questions one by one below.

We have carefully reviewed and organized your valuable comments.

★ Please rest assured that we have checked all your questions several times to ensure that none were missed. We have grouped related concerns as follows.

[I]. Explanation of Weaknesses and Questions

[1/3] W2&Q1&Q2: Why Use Diffusion Based RL?

• Thank you for your question. Existing feature weighting methods can effectively identify important spectral bands. However, in Raman spectroscopy, preserving the relationships among spectral peaks is essential for extracting stable discriminative information from heterogeneous spectra. This motivated us to reconsider traditional band selection methods.

❖ Notably, a spectral peak may be discriminative on its own, but its combination with other peaks may not be equally effective.

❖ Direct weighting can implicitly learn feature interactions, but it lacks explicit exploration of candidate band combinations.

► Therefore, we considered two complementary mechanisms:

(i) ✓ constructing an energy function from shared relational prototypes to guide diffusion in exploring different band combinations.

(ii) ✓ using prototype consistency and prediction feedback to evaluate candidate combinations and continuously refine the selection policy.

⇒ Thus, diffusion based RL naturally follows from our goal of turning the progressive exploration of band combinations into a sequential decision process that can be continuously improved through feedback.

❖ However, keeping the exploration direction consistent with the optimization objective is not straightforward.

⇒ We carefully analyzed these challenges. We demonstrated that combining relation guided diffusion exploration with prototype driven policy feedback provides a strong baseline. This is the fundamental methodological innovation and core contribution of our work.★

• In our original manuscript, we compared KERD with 15 baselines, including 9 general baselines and 6 domain specific baselines (Section 4.1 and Appendix H). These include several attention based methods, such as PatchTST, Raman-OSDL, and iTransformer.

★ In addition, we introduced four new baselines from two categories: Differentiable Selection (Hide&Seek, ICML'26, and TT-Sparse, ICML'26) and Direct Energy Optimization (SEA, NeurIPS'25, and EBT, ICLR'26).

⇒ Our comparison with 19 baseline methods further demonstrates the effectiveness of KERD. More detailed results are provided in Appendix L.

☆ Table R1. Comparison with Differentiable Selection and Direct Energy Optimization Methods.

❖ CategoryProtocolAcc. (%) ↑F1 (%) ↑AUPRC (%) ↑Training Time (min) ↓
❖ Differentiable SelectionTT-Sparse (ICML'26)88.33±0.9287.96±1.0894.12±1.60≈188
Hide&Seek (ICML'26)75.87±3.3375.18±3.5181.14±4.66≈134
❖ Direct Energy OptimizationSEA (NeurIPS'25)80.40±2.8979.54±2.9787.51±2.32≈246
EBT (ICLR'26)89.00±0.6588.79±0.6393.60±1.50≈285
❖ OursKERD (Ours)★93.37±0.59★93.32±0.61★96.99±0.66≈203.0
AuthorsOpenReview ↗

Response to Reviewer ELNx [part 2]

[2/3] W1&W3: Computational Efficiency

❖ Additional Computational Cost Analysis: To improve the computational efficiency of KERD, we adopted fixed step numerical integration (Fixed-step Integration) in our original method.

✓ This technique reduces the computational overhead of solving continuous state evolution.

✓ We further conducted an efficiency comparison. Reducing the sampling steps from 32 to 8 achieves a speedup of approximately 3.02 times, as shown in Table R2.

✓ In addition, Table R3 compares training time, GPU memory usage, and end to end inference latency.

★ KERD achieves substantial improvements in predictive performance while keeping its additional computational overhead within an acceptable range.

⇒ The results are presented below.

☆ Table R2. Computational Efficiency of Fixed-step Reverse Diffusion Sampling.

Sampling Steps3216842
Relative Computation Steps1×2× fewer4× fewer8× fewer16× fewer
Time per Batch (ms) ↓26.314.28.76.15.2
Speedup ↑1.00×1.85×★3.02×4.31×5.06×
Accuracy (%) ↑93.6193.45★93.3791.7391.54

☆ Table R3. Computational Cost on Gastric-SERS.

❖ MethodF1 (%) ↑Training Time (min) ↓GPU Memory (GB) ↓End-to-end Inference (ms/batch) ↓
❖ Direct Band Weights76.14≈170.0≈4.1≈26.5
❖ Stochastic Mask + RCGRPO87.54≈192.0≈4.7≈30.1
❖ KEDP w/o RCGRPO89.21≈184.0≈4.9≈38.7
►Full KERD (Ours)93.32≈203.0≈5.3≈39.2

☆ Table R4. Cost Comparison with Representative Baselines.

DatasetMethodF1 (%) ↑Training Time (min) ↓Peak GPU Memory (GB) ↓End-to-end Inference (ms/batch) ↓
Gastric-SERSTimeFilter80.34≈182.0≈4.8≈32.1
Gastric-SERSPatchTST82.09≈159.0≈3.7≈25.5
Gastric-SERSMultiPatchFormer82.26≈147.0≈4.2≈24.2
Gastric-SERSiTransformer85.94≈196.0≈4.7≈37.9
Gastric-SERSKERD (Ours)93.32≈203.0≈5.3≈39.2

⇒ We have added further details to the revised manuscript. The corresponding results are provided in Appendix J, Tables R3–R5.

[3/3] W1&Q3: Cross Center Generalization

• Robustness Experiments ✓: The baseline drift, fluorescence interference, and spike noise considered in KERD originate from real Raman spectrometers rather than arbitrarily designed perturbations. Therefore, these perturbations do not undermine the practical relevance of our robustness evaluation.

• Cross Center Generalization ✓: To go beyond evaluating each dataset independently, we conducted transfer experiments from three heterogeneous source domains to HD (See the table below). We compared three settings: training from scratch, using a frozen pretrained encoder, and full fine tuning.

☆ Table R5. Transfer Settings

ComponentTransfer Evaluation Protocol
SourcesGastric-SERS, COVID-Serum, Ovarian Plasma
TargetCross-Center Heart Disease (HD)
PretrainingShared spectral/graph encoders; source-specific heads
ScratchTrain KERD on HD only
FrozenFreeze pretrained encoders; train HD linear probe
Fine-tuningAdapt pretrained KERD to HD
Runs5 seeds (mean±std)

☆ Table R6. HD Transfer Performance

ProtocolAcc. (%) ↑F1 (%) ↑AUPRC (%) ↑
Scratch93.26±4.3893.22±4.3797.07±2.43
Frozen94.65±1.4094.74±1.3598.27±0.60
Fine-tune95.81±1.5895.82±1.6498.34±1.36

The frozen encoder achieves an F1 score of 94.74%. Full fine tuning further improves the F1 score to 95.82%, representing a gain of 2.60 percentage points over training from scratch.

★ These results suggest that the spectral representations learned by KERD have the potential to generalize across centers. The corresponding experiments are provided in Appendix K.

Finally, we sincerely thank you again for your valuable time and encouragement.

We wish you all the best.

With our deepest respect.

Thank you for your reply. The author's rebuttal addressed most of my concerns, and since my original rating was positive, I will maintain it.

AuthorsOpenReview ↗

We sincerely appreciate your positive feedback on our work.

Thank you so much for your positive feedback and continued support! We truly appreciate your professionalism and kindness. Hope everything goes well for you.

Reviewer_8LeeOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Knowledge-Energy Reinforced Diffusion (KERD) for supervised representation learning from Raman spectra. It partitions a spectrum into contiguous band groups, constructs a local Peak Graph, and estimates class-conditional prototypes of normalized band responses and adjacent-band log ratios from training data. These statistics are reused in an energy that guides diffusion over band masks and in the reward for group-relative policy optimization. An optional graph incorporates auxiliary clinical variables. The paper provides an idealized diffusion-control analysis and evaluates the approach on seven independently trained spectral datasets totaling 72,057 measurements, with general sequence baselines, Raman-specific baselines, component and diffusion-strategy ablations, and perturbation experiments.

Critical strengths
  1. The evaluation is broad and includes meaningful controls. Tables 2 and 6 report results over five training seeds on seven heterogeneous datasets, and Appendix H adds six Raman-specific baselines. The main comparisons use shared clinical inputs and the same MCGR interface where needed; diagnosis-defining fields are explicitly excluded (Appendix D, p. 25). Figure 3 includes component removals and several diffusion alternatives, while Tables 4–5 separate spectral-only, concatenation, and graph-based context integration.
  2. The energy and feedback mechanisms share an explicit, deployment-compatible relational representation.
Critical weaknesses
  1. The central sampling and optimization procedure is not sufficiently specified to reproduce or fully assess. These details determine the proposed method, rather than merely facilitating implementation. The substantial ablations are therefore harder to interpret and independently reproduce.
  2. The optimal-path claim does not follow for the fixed random initial distribution used in its derivation. Appendix A.1 fixes the same x0∼ρ0x_0\sim\rho_0 for both uncontrolled and controlled diffusions (Eqs. 15–16). However, the globally tilted measure in Eq. (25) generally has initial marginal =\rho_0(x_0)\, \frac{\mathbb E_{P_{\rm pre}}[e^{R(x_T)}\mid x_0,\mathrm{cond}]} {Z(\mathrm{cond})},$$ which need not equal $\rho_0$. Changing the drift alone cannot realize that measure. Equation (26) remains a valid variational identity, but the asserted attainability and optimum in Eq. (27), and hence Theorem 3.1 as stated, require qualification, for example, conditioning on a deterministic initial state or permitting the initial law to change. This is distinct from the acknowledged local-surrogate approximation and affects the exact theoretical reference used to motivate the method.
Questions for authors
  1. Could the authors provide a fully specified sampling and training update: the discrete guided transition and the log probability used by RCGRPO; the construction of m~,ν\tilde m,\nu, including the sigmoid/mask transformation; the actual policy and state-updater objective and gradient paths; and the values of GG and Tmax⁡T_{\max}? Please also distinguish the number of reverse-diffusion steps from the number of sequential evidence-selection steps, if these are different.
  2. Is Theorem 3.1 intended to optimize over arbitrary path measures, or over drift controls with the fixed initial distribution of Eqs. (15)–(16)? In the latter case, should the path tilt be normalized conditionally on x0x_0, with value Eρ0log⁡E[eR(xT)∣x0,cond]\mathbb E_{\rho_0}\log\mathbb E[e^{R(x_T)}\mid x_0,\mathrm{cond}]? Please state the applicable conditions and explain whether the correction changes any implemented guidance or claimed guarantee.
4 comments
AuthorsOpenReview ↗

Response to Reviewer 8Lee [part 1]

Thank you for recognizing our study as innovative and comprehensive in its evaluation. We extend our highest respect and sincere appreciation.

We have devoted great care to this work. We sincerely hope you could spare 10 minutes to reopen the manuscript PDF and continue our discussion.

★ We have carefully reviewed and organized your valuable comments.

★ Please rest assured that we have checked all your questions several times to ensure that none were missed. We have grouped related concerns as follows.

[I]. Explanation of Weaknesses and Questions

[1/2] W1&Q1: Clarification of Sampling Details

Thank you for your careful review! We would like to clarify that the relevant details were already provided in the original manuscript, although some were included in the appendix.

• In fact, Section 3.2, Appendix A.2.3, and Algorithm 1 already provide the basic definitions of sampling and policy updates. These details were not added only during the rebuttal.

❖ Specifically, KERD uses the guided reverse transition probability πη(at∣st)\pi_{\eta}(a_t\mid s_t) with energy correction. It updates the policy using the following loss:

⇒ Lpol=−1G∑i=1Gsg(A^(i))∑t=1Tmax⁡log⁡πη(at(i)∣st(i))\mathcal{L}_{\mathrm{pol}}=-\frac{1}{G}\sum_{i=1}^{G}\mathrm{sg}(\hat{A}^{(i)})\sum_{t=1}^{T_{\max}}\log\pi_{\eta}(a_t^{(i)}\mid s_t^{(i)})

The energy network remains fixed during policy updates.

❖ Importantly, GG denotes the number of independent trajectories sampled from the same state, whereas Tmax⁡T_{\max} denotes the number of reverse diffusion steps and sequential selection steps per trajectory. These are different quantities. Each step generates a continuous mask and updates the sequential state as follows:

⇒ ht=Aξ(ht−1,vˉt)h_t=A_{\xi}(h_{t-1},\bar{v}_t)

Mask uncertainty is estimated using the mean and sample variance of MM conditional samples.

✓ We agree that these details could be presented more clearly. Therefore, we have further clarified the implementation in Appendix D (Lines 1311 to 1316). Appendix B (Algorithm 1) provides the complete training and inference procedure.

AuthorsOpenReview ↗

Response to Reviewer 8Lee [part 2]

[2/2] W2&Q2: Theoretical Clarification

Thank you for your careful review. ❖ In Raman spectroscopy, preserving relationships among spectral peaks is essential for extracting stable discriminative information from heterogeneous spectra. This motivated us to reconsider existing spectral feature analysis methods. Notably, a spectral peak may be discriminative on its own, but it may not remain effective when combined with other peaks.

⇒ Previous studies have often overlooked this aspect. This is the main motivation for our work.

☆ However, providing a clear theoretical basis for this exploration is not straightforward. Traditional KL regularized stochastic control can adjust path distributions through rewards. However, globally tilting the entire path distribution may also change the initial distribution.

⇒ Therefore, we considered two related questions:

(i) ★ how to identify the optimal attainable path distribution through conditional KL regularized path optimization under a fixed initial distribution. (ii) ★ how to convert the theoretical guidance direction into practical stepwise diffusion updates using a local energy surrogate.

► In this way, relational energy can evaluate different band combinations. It can also guide the model toward more informative candidate combinations.

★ We carefully analyzed these limitations and demonstrated that combining progressive energy guidance with policy optimization provides a strong baseline. ✓ Specifically, we introduced the conditional normalization term Z0(x0)Z_0(x_0) in Section 3.2 (Lines 208 to 213) and Appendix A.1.2 (Lines 859 to 882). We corrected the optimal path distribution. Under assumptions including drift realizability, we obtained:

⇒ sup⁡gJ(g)=Ex0∼ρ0[log⁡Z0(x0)]≤log⁡Z(cond)\sup_g J(g)=\mathbb{E}_{x_0\sim\rho_0}[\log Z_0(x_0)]\leq\log Z(\mathrm{cond}).

Notably, KERD does not need to solve this optimal path distribution exactly in practice. Instead, it uses the local energy surrogate in Eqs. (8) and (9) for stepwise guidance. It also uses relational consistency and prediction changes to provide feedback for policy optimization.

☆ Furthermore, to verify whether these mechanisms are truly necessary, we separately examined the contributions of direct band weight learning, diffusion exploration, and policy feedback. The experimental settings and results are provided in Appendix I (Tables R1 and R2).

☆ Table R1. Necessity of Diffusion and RL

❖ VariantWhat Is RetainedWhat Is ReplacedQuestion Addressed
❖ Direct Band WeightsBackbone, MCGR, Eq. (5)Directly optimize continuous masks, removing diffusion and RL ✓⇒ Is direct band weight learning sufficient?
❖ Stochastic Mask + RCGRPOBackbone, MCGR, relational energy, RCGRPOReplace KEDP with a stochastic mask policy without diffusion ✓⇒ Is policy feedback alone sufficient without diffusion?
❖ KEDP w/o RCGRPOBackbone, MCGR, KEDPReplace policy feedback with direct differentiable training ✓⇒ Is diffusion based exploration alone sufficient?
► Full KERD (Ours)All componentsNone⇒ Can structured exploration and feedback based on relational consistency provide additional gains?

☆ Table R2. Results on Gastric-SERS

❖ VariantF1 (%) ↑Pair Retention (%) ↑Mask Stability ↑
Direct Band Weights76.1469.350.62
Stochastic Mask + RCGRPO87.5473.540.78
KEDP w/o RCGRPO89.2181.350.76
Full KERD (Ours)★93.32★89.16★0.85

Metric Descriptions:

  • F1: Final performance.
  • Pair Retention: Proportion of important spectral peak pairs jointly retained.
  • Mask Stability: Consistency of selected band subsets.

► In contrast, the full KERD further improves performance to 93.32% F1, 89.16% Pair Retention, and 0.85 Mask Stability. These results show that relation guided diffusion exploration and prototype driven policy feedback play complementary roles.

★ Our theoretical contribution is to connect structured relational energy, conditional path optimization under a fixed initial distribution, and practical local diffusion guidance. This provides a theoretical basis for the stepwise exploration of band combinations.

Ethical Details: Thank you for raising this important point. We have already stated in the Ethics Statement (Page 10) of the original manuscript that our study follows institutional ethical requirements. We also explained that all clinical data were deidentified.

We spent more than a year collecting and organizing extensive medical data across several hospitals. We have devoted great care to this study. We sincerely hope you could spare just 10 minutes of your valuable time to take another quick look at our manuscript.

We believe that the aforementioned issues can be effectively addressed.

We sincerely hope you could reconsider your score based on our clarifications.

We wish you all the best. With our deepest respect.

Most of my concerns have been addressed, and I would like to raise my score.

AuthorsOpenReview ↗

Thank you very much for your positive feedback on our work.

We sincerely appreciate your thoughtful reconsideration and your willingness to raise the score.

We deeply appreciate your valuable guidance and encouragement, and we will carefully incorporate your suggestions to make our manuscript stronger and more complete!

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

The paper proposes KERD for Raman spectral classification, combining graph-based spectral representations, energy-guided diffusion for band selection, and reinforcement learning with prototype-based rewards. Experiments on seven datasets report improvements over general and domain-specific baselines.

Critical strengths

The reported improvements across several datasets are promising. Using shared spectral relation prototypes for both selection guidance and training feedback is an interesting design.

Critical weaknesses

The necessity of diffusion and reinforcement learning is not established. The complete spectrum is available before selection, the band masks are continuous, and the main objectives are differentiable. Directly learning band weights is therefore a natural, simpler alternative. However, the paper does not compare against this alternative under the same backbone and relational constraints. Removing entire modules does not establish that diffusion and reinforcement learning outperform simpler replacements. Consequently, the framework appears to stack several complex components without sufficiently demonstrating their additional value.

The theoretical justification requires correction, and implementation details remain unclear. Appendix A.1 fixes the initial distribution, whereas the globally tilted path measure in Eq. (25) generally changes that distribution. The derivation therefore does not establish that the claimed optimum is attainable through drift modification alone. The actual policy loss, the relationship between diffusion steps and sequential selection, and the computation of mask uncertainty also require clearer specification.

Questions for authors

Under the same backbone, inputs, and relational constraints, how does KERD compare with directly learning continuous band weights in predictive performance and computational cost? How is the globally tilted path measure reconciled with the fixed initial distribution? What are the exact policy objective, sampling procedure, and mask-uncertainty computation used in the experiments?

4 comments
AuthorsOpenReview ↗

Response to Reviewer wqSJ [part 1]

★ First, we would like to express our deepest respect. Thank you for your careful review! We hold you in the highest regard.

We have devoted considerable effort to this work. We sincerely hope you could spare 10 minutes to reopen the manuscript PDF and continue the discussion.

We have carefully reviewed and organized your valuable comments.

★ Please rest assured that we have checked all your questions several times to ensure that none were missed. We have grouped related concerns as follows.

[I]. Explanation of Weaknesses

[1/2] W1: How can we demonstrate the necessity of RL-based diffusion?

This is an important question! The reviewer raises a key issue in this field: why did we not simply learn band weights? The main reason is that learning band importance is not the same as learning relationships among band combinations.

  • ✓ In Raman spectroscopy, a band may be informative on its own, but this does not mean it remains useful when combined with other bands. In contrast, some less prominent peaks may provide discriminative information through their relative intensities and coordinated changes.

  • ✓ The 2025 study Classification of Alzheimer's Disease in a Mixed Clinical Cohort Using Biofluid Raman Spectroscopy reported correlations among different Raman peaks.

This directly motivated the development of KERD.

◈ In fact, we have already compared KERD with methods that adaptively learn feature weights, including Raman OSDL and Objectosphere (Appendix H, Fig. 6). Our method consistently outperforms these baselines, which agrees with the trends in Table 2.

❖ Moreover, during spectral data collection in hospitals, we observed that preserving relationships among spectral peaks can help distinguish stable molecular fingerprints from measurement variations unrelated to biological factors. This motivated us to reconsider methods that rely only on feature weighting. Notably, direct weighting can highlight important bands, but it cannot explicitly explore changes in their combinations. Such changes are often overlooked in spectral representation learning.

◈ Since diffusion models can explore complex combination spaces, we first considered using spectral peak relationships to guide the exploration process instead of optimizing individual band weights. We then identified a new challenge: how can we ensure that relational constraints remain consistent throughout combination exploration and policy optimization? This led us to consider two complementary mechanisms: (i) using relational energy to guide the exploration of different band combinations, and (ii) using prototype consistency and prediction changes to provide trajectory feedback.

◈ However, making these mechanisms work together is not straightforward. We carefully analyzed their limitations. We then demonstrated that combining progressive energy guidance with policy optimization provides a strong baseline. This is the main motivation and core contribution of our work. Due to space limitations, Appendix D of the original manuscript explains the roles of these components and their added value. Specifically, Appendix D.1 (Page 25, Lines 1318 to 1349) discusses the three components in Lines 1320 to 1333. Lines 1335 to 1340 analyze robustness to perturbations.

◈ Finally, to better understand these differences, we separately examined the contributions of direct band weight learning, diffusion exploration, and policy feedback. The experimental settings and results are presented below. We have also added them to Section I (Lines 1620 to 1660) of the revised manuscript.

☆Table R1. Necessity of Diffusion and RL.

❖ VariantWhat Is RetainedWhat Is ReplacedQuestion Addressed
❖ Direct Band WeightsBackbone, MCGR, Eq. (5)Directly optimize continuous masks, removing diffusion and RL ✓⇒ Is direct band weight learning sufficient?
❖ Stochastic Mask + RCGRPOBackbone, MCGR, relational energy, RCGRPOReplace KEDP with a stochastic mask policy without diffusion ✓⇒ Is policy feedback alone sufficient without diffusion?
❖ KEDP w/o RCGRPOBackbone, MCGR, KEDPReplace policy feedback with direct differentiable training ✓⇒ Is diffusion based exploration alone sufficient?
➤ Full KERD (Ours)All componentsNone⇒ Can structured exploration and feedback based on relational consistency provide additional gains?

☆Table R2. Results on Gastric SERS.

❖ VariantF1 ↑Pair Retention ↑Mask Stability ↑
Direct Band Weights76.1469.350.62
Stochastic Mask + RCGRPO87.5473.540.78
KEDP w/o RCGRPO89.2181.350.76
➤ Full KERD (Ours)★93.32★89.16★0.85

-Metric Descriptions:

  • F1: Final performance.
  • Pair Retention: Proportion of important spectral peak pairs jointly retained.
  • Mask Stability: Consistency of selected band subsets.
AuthorsOpenReview ↗

Response to Reviewer wqSJ [part 2]

[2/2] W2: Clarification of Theoretical and Implementation Details

We thank the reviewer for the thoughtful observation regarding the theoretical derivation and welcome the opportunity to provide further clarification.

We address the concerns as follows.

◈ Theoretical Derivation: The key distinction is that changing the probability of an entire trajectory is not the same as changing only its evolution after a fixed initial state. The former may also change the initial distribution. Therefore, we have clarified this point in Lines 208 to 213 of the revised manuscript.

◈ This correction does not change the conclusions of the local energy guidance used by KERD. In practice, our method performs progressive guidance through the local energy surrogate in Eqs. (8) and (9). We made the following revisions:

☆ Table R3. Summary of Theoretical Revisions.

ConcernDescriptionRevisionExact Position in Revised Paper
W2Fixed initial distribution★ Clarified the fixed initial state and conditional KL optimization objective⇒ Section 3.2, Lines 208 to 213 ✓
W2Optimal path measure★ Introduced conditional normalization to correct the optimal path distribution⇒ Appendix A.1.2, Lines 859 to 863 ✓
W2Theoretical proof★ Clarified the global upper bound and established the conditional optimality result⇒ Appendix A.1.2, Lines 880 to 882 ✓

• ✓ Implementation Details:We have already described the policy loss, the relationship between diffusion steps and sequential selection, and the computation of mask uncertainty in the original manuscript. Please refer to Section 3.2 (Energy Guided Decision and Policy Optimization), Appendix A.1.4, and Appendix A.2.3.

★ To make the implementation clearer, we have added further details in Appendix D (Page 25, Lines 1311 to 1316). These details clarify the relationship between the actual policy loss and sequential state updates.

AuthorsOpenReview ↗

Response to Reviewer wqSJ [part 3]

[II]. Explanation of Questions

[1/1] Q1: Additional Computational Cost Analysis and Two Implementation Details (1. Consistency with the Initial Distribution 2. Policy Objective, Sampling, and Computation)

◈ Additional Computational Cost Analysis: To improve the computational efficiency of KERD, we adopted fixed step numerical integration in our original method.

❖ This technique reduces the computational overhead of solving continuous state evolution. We further conducted an efficiency comparison.

❖ Reducing the sampling steps from 32 to 8 achieves approximately 3.02 times speedup, as shown in Table R4.

❖ In addition, Table R5 compares training time, GPU memory usage, and end to end inference latency.

✓ KERD achieves substantial improvements in predictive performance while keeping its additional computational overhead within an acceptable range. The results are presented below.

☆ Table R4. Computational Efficiency of Fixed Step Reverse Diffusion Sampling.

Sampling Steps3216842
Relative Computation Steps1×2× fewer4× fewer8× fewer16× fewer
Time per Batch (ms) ↓26.314.2★ 8.76.15.2
Speedup ↑1.00×1.85×★ 3.02×4.31×5.06×
★ Accuracy (%) ↑93.6193.45★ 93.37*91.7391.54

☆ Table R5. KERD achieves substantial performance gains at a manageable computational cost.

MethodAccuracy (%) ↑F1 (%) ↑Training Time (min) ↓GPU Memory (GB) ↓End to End Inference (ms/batch) ↓
Direct Band Weights78.2576.14≈170.0≈4.1≈26.5
KERD (Ours)★93.37★93.32≈203.0≈5.3≈39.2

◈ The corresponding revisions are provided in Section J, Tables R3 to R5, of the revised manuscript.

• In addition, we conducted further evaluations using four new baselines with comparable model capacities, covering two categories:

✓ - Differentiable Selection (Hide&Seek, ICML'26, and TT-Sparse, ICML'26)

✓ - Direct Energy Optimization (SEA, NeurIPS'25, and EBT, ICLR'26).

★ KERD significantly improves predictive performance with acceptable computational overhead. The results are shown in the table below.

☆ Table R6. Comparison with Differentiable Selection and Direct Energy Optimization Methods.

CategoryProtocolAcc. (%) ↑F1 (%) ↑AUPRC (%) ↑Training Time (min) ↓
Differentiable SelectionTT-Sparse (ICML'26)88.33±0.9287.96±1.0894.12±1.60≈188
Hide&Seek (ICML'26)75.87±3.3375.18±3.5181.14±4.66≈134
Direct Energy OptimizationSEA (NeurIPS'25)80.40±2.8979.54±2.9787.51±2.32≈246
EBT (ICLR'26)89.00±0.6588.79±0.6393.60±1.50≈285
OursKERD (Ours)★93.37±0.59★93.32±0.61★96.99±0.66≈203

• Consistency with the Initial Distribution: ✓ As clarified in W2, global path tilting may change the initial distribution.

➤ Therefore, we introduced conditional normalization and revised the optimality claim to reflect what is attainable. Please see Table R3.

➤ This correction does not require any changes to the actual local energy guidance mechanism or its conclusions.

• Policy Objective, Sampling, and Computation: ✓ As explained in W2, Section 3.2 and the related appendices of the original manuscript provide the basic definitions. Appendix D (Lines 1311 to 1316) of the revised manuscript further clarifies the actual computation process.

◉ We welcome further discussion if any concerns remain.

We spent more than a year collecting and organizing extensive medical data across several hospitals. We have devoted great care to this study. We sincerely hope you could spare just 10 minutes of your valuable time to take another quick look at our manuscript.

We believe that the aforementioned issues can be effectively addressed.

We sincerely hope you could reconsider your score based on our clarifications.

We extend our highest respect and sincere appreciation. We hope everything goes well for you.

AuthorsOpenReview ↗

A Kind Follow-up and Thank You

Dear Reviewer wqSJ,

We are writing to ask whether our rebuttal has addressed your concerns. We would be happy to discuss any remaining issues.

We truly value your feedback and thoughtful suggestions. Within the limited time available, we have made every effort to conduct additional experiments and address your concerns as thoroughly as possible. We will remain actively engaged in the discussion and ensure that we respond promptly to any further questions you may have.

With our sincere appreciation and deepest respect! We hope you are doing well.

Best regards, The authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

Brief introduction to the revised PDF

We sincerely thank all reviewers for their valuable time and constructive comments.

We really hope you could take just 10 minutes to reopen the manuscript PDF, briefly review it, and continue the discussion with us. We extend our deepest respect to you.

◈ We manually added bullet points to our rebuttal to make our responses clearer and easier to follow. These symbols were not generated by AI or LLMs. We simply wanted to organize our responses better, make them easier to read, and save reviewers some valuable time. We hope this also makes the review and discussion process more efficient.

☆ We added new content only after Line 1619 to preserve the original line references. The only exceptions are revisions to Lines 208 to 213, 859 to 882, and 1311 to 1316, following the suggestions from reviewers wqSJ and 8Lee.

★ In addition, we added eight tables (Tables R1 to R8) to the revised manuscript to further address the reviewers' main concerns:

  1. Tables R1 to R2 ✓: We added controlled ablation experiments on direct band weight learning, diffusion exploration, and policy feedback to validate the necessity of each mechanism.
  2. Tables R3 to R5 ✓: We added analyses of diffusion sampling efficiency, training time, GPU memory usage, and end to end inference cost.
  3. Tables R6 to R8 ✓: We added cross dataset transfer experiments, along with comparisons against recent differentiable feature selection and direct energy optimization methods.

We sincerely thank all reviewers again for their valuable suggestions. We welcome further discussion.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.