acceptodds
Under review as a conference paper at ICLR 2027

LHC-Net: Bridging LLM Priors and Hypergraph Structure via Channel Retrieval for Biochemical Forecasting

Abstract

Multivariate Time Series Forecasting (MTSF) serves as a core instrument in AI for Science (AI4S) for analyzing complex biochemical processes. However, practical applications face a fundamental dilemma: balancing the physical deficiency caused by Channel Independence (CI) against the noise overfitting induced by Channel Dependence (CD). While CI mitigates high-dimensional noise, it severs essential physical coupling; conversely, CD models global dependencies but introduces spurious correlations, leading to redundancy and overfitting. To address this, we propose LHC-Net (LLM-initialized Hypergraph Channel-Retrieval Network), a novel paradigm integrating scientific domain knowledge with deep learning. Unlike methods employing Large Language Models (LLMs) as direct numerical predictors, we leverage their general scientific knowledge to extract sparse hypergraph structural priors, characterizing high-order mechanistic dependencies among variables. To mitigate the risk of LLM hallucinations, we define this topology as learnable parameter matrices. This allows dynamic correction of priors via gradient backpropagation from observational data, achieving robust integration of mechanistic priors and empirical evidence. Furthermore, we design a temporal channel retrieval strategy that enables the model to access critical variable information from the mechanistic hypergraph on-demand, guided by future temporal contexts. This structured channel fusion strategy restores the necessary physical coupling of dynamical systems while effectively shielding against irrelevant interference. Extensive experiments on a million-scale real-world biopharmaceutical dataset demonstrate that LHC-Net significantly outperforms state-of-the-art models in accuracy. Moreover, it successfully recovers reaction paths consistent with biochemical principles. By grounding LLM priors in observational data, LHC-Net effectively mitigates hallucination risks, enhancing the application potential of AI4S. Code is available at https://anonymous.4open.science/r/LHC-Net-8B41

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Multivariate Time Series Forecasting, Large Language Models, Hypergraph Structural Priors, Channel Retrieval Strategy, Biochemical Processes

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_oK6gOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper focuses on Multivariate Time Series Forecasting (MTSF) for analyzing complex biochemical processes. The authors stress the challenges of balancing Channel Independence (CI) and Channel Dependence (CD) strategies, and therefore propose LLM-initialized Hypergraph Channel-Retrieval Network (LHC-Net). They suggest using LLM to initialize a hypergraph as a structural prior using its general scientific knowledge instead of using LLM for prediction directly. The LLM-initialized hypergraph then gets updated and refined through observational data, where a hypergraph encoder(HIE) captures high-order interactions , and a channel retrieval-based decoder (CRD) is used to select relevant information for dynamic forecasting. The authors benchmark their method on proprietary and public fermentation data and demonstrate improved performance against state-of-the-art baselines.

Critical strengths
  • The problem formulation is reasonable. The design choice of building a structured hypergraph strikes a balance between CI and CD. The LLM-initialization as prior followed by data-driven refinement for correction is an effective way to incorporate different types of knowledge including expert’s notes and measured experimental data. It’s also intuitive that the LLM prior gets the model close to the solution first, then the observational data continues corrects and the structure.
  • The experimental evaluation is extensive and supports the paper’s design choice well. The authors comprehensively consider alternative options throughout their architectural design and perform benchmarking accordingly. They evaluate the method across multiple forecasting horizons against different categories of baseline approaches. They perform ablation studies on individual components, and assess the model robustness using different structural priors.
Critical weaknesses
  • The biological validation is primarily qualitative. A more systematic and quantifiable assessment of the learned hypergraph against independent biological evidence (e.g., a reference knowledge graph for known biological relationships) would strengthen the paper, and help determine whether the learned relationships reveal true biological mechanisms or mainly statistical correction.
  • The curation of knowledge base by experts is critical and will directly impact the LLM initialized prior. The current manuscript does not seem to cover how important the knowledge engineering aspect, for example how much the knowledge base size, variability, comprehensiveness, and correctness could affect the final performance. There’s also a lack of validation how LLM initiated hypergraph faithfully aligns with the knowledge base. That to be me is a good anti-hallucination check.
Questions for authors
  • What if LLM is not used for constructing the structural prior, but for retrieval of relevant information from the expert curated knowledge base? How does such RAG system compare against LHC-Net, where LLM is for direct prediction but context enriched at inference time?
  • Is it possible to construct an expert-derived hypergraph for the same domain knowledge and used as a reference to evaluation the LLM-derived prior? If not, could the LLM-derived prior be thoroughly evaluated against experts judgements?
11 comments
AuthorsOpenReview ↗

1. Response to W1: Quantitative Assessment and Biological Validation

Thank you for emphasizing the importance of validating learned hypergraph structures against biological evidence. We agree that topological convergence alone is insufficient to establish mechanistic fidelity. We would like to clarify that our original submission already included quantitative expert-based validation, which we have now complemented with a new event-driven physical consistency analysis.

1. Expert-Referenced Quantitative Validation (Existing Evidence):

As detailed in Appendix J (LSI Robustness and Feasibility Testing, Table 8), we have systematically evaluated LLM-generated structural priors against expert-annotated biochemical relationships using two complementary metrics: Quadratic Weighted Kappa (QWK) for scoring agreement and Normalized Discounted Cumulative Gain (NDCG) for the ranking consistency of mechanistic drivers.

These evaluations demonstrate strong alignment between LLM-generated priors and expert judgments, providing quantitative evidence beyond qualitative topology visualizations. The detailed expert–LLM comparison process is further illustrated in Appendix K (LLM Reasoning Data Flow and Expert Alignment, Figure 10).

2. Event-Driven Physical Consistency Analysis (New Evidence):

To further examine whether the learned relationships reflect meaningful biochemical dynamics rather than merely statistical corrections, we have added Section 4.5 (Event-Driven Physical Consistency Analysis, Figure 3), with additional explanations in Appendix P (Detailed Biochemical Interpretation of Dynamic Retrieval).

By jointly examining the converged hypergraph, CRD attention patterns, and actual biochemical sensor trajectories, we observe that:

  • Structural Consistency: The learned hypergraph captures established biochemical coupling between substrate availability and microbial biomass.
  • Temporal Consistency: CRD dynamically adjusts its attention to substrate availability in alignment with the observed biochemical transition.

Together, these evaluations provide complementary evidence of expert-aligned structural initialization and physically interpretable learned dynamics, strengthening the biological validation beyond topology convergence alone.

We have also made the new event-driven analysis materials available in our anonymous repository for further inspection.

AuthorsOpenReview ↗

2. Response to W2: Knowledge Base Curation and Alignment Validation

Thank you for highlighting the importance of knowledge engineering and its potential influence on LLM-generated structural priors. We agree that the quality and coverage of the knowledge base are important factors in ensuring reliable initialization.

Knowledge Base Construction and Transparency:

As detailed in Appendix O (Knowledge Base Construction and Prompt Design), our industrial knowledge base integrates scientific literature and factory Standard Operating Procedures (SOPs). Rather than directly feeding unstructured documents into the LLM, we employ a structured RAG pipeline featuring:

  • Hybrid Retrieval: BM25-based sparse retrieval and vector-based semantic retrieval to balance terminology matching and contextual relevance.
  • Information Filtering: ID-based deduplication and structured context organization to reduce redundant information and prompt interference.
  • Reproducibility: Detailed retrieval configurations, prompt templates, and representative knowledge-base examples are provided in our anonymous repository.

Expert Alignment and Anti-Hallucination Validation:

Importantly, we have already quantitatively evaluated the alignment between LLM-generated priors and independent expert judgments in Appendix J (LSI Robustness and Feasibility Testing, Table 8).

The QWK and NDCG evaluations demonstrate strong agreement in both mechanistic association scores and the identification of key biochemical drivers. Furthermore, Appendix K (LLM Reasoning Data Flow and Expert Alignment, Figure 10) provides a concrete comparison of expert annotations and LLM reasoning outputs, illustrating how retrieved domain knowledge is translated into structured hypergraph priors.

Together, these evaluations provide an expert-referenced consistency check against potentially misleading LLM-generated associations, complementing the data-driven correction mechanism discussed in Section 4.4 (Analysis of Anti-Hallucination Capabilities).

AuthorsOpenReview ↗

3. Response to Q1: Comparison with Direct LLM Prediction via RAG

Thank you for this insightful suggestion. To directly investigate the alternative use of retrieved expert knowledge for LLM-based forecasting, we conducted a new comparative experiment using Time-LLM, a representative LLM-based time-series forecasting framework.

Specifically, we incorporated RAG-retrieved domain knowledge into the prompt prefix and evaluated two settings:

  • TimeLLM-RAG: Retrieved expert knowledge is incorporated during both training and inference.
  • TimeLLM-Infer-RAG: The model is trained without retrieved knowledge, which is introduced only at inference time to evaluate inference-time context enrichment.

The complete results are provided in Appendix R (Comparison with Direct LLM Prediction via RAG, Table 10), with the key comparisons summarized below.

[Insert experimental results table here]

Key Findings:

  • Prediction Performance: Both RAG-augmented variants exhibit higher forecasting errors than the original Time-LLM. In particular, TimeLLM-Infer-RAG performs worse, suggesting that the inconsistency between training and inference contexts may further impair numerical forecasting. These results indicate that directly incorporating qualitative domain knowledge into prediction prompts does not necessarily improve numerical modeling.

  • Computational Overhead: Incorporating lengthy retrieved contexts substantially increases memory requirements. Notably, TimeLLM-RAG encounters Out-of-Memory (OOM) failures for prediction horizons of 192 and above under our experimental settings. This further compounds the computational overhead of Time-LLM, as previously analyzed in Appendix M (Computational Efficiency and Industrial Scalability).

These findings further support our architectural choice: Rather than relying on knowledge-enriched LLM inference for numerical prediction, LHC-Net leverages LLMs offline to construct mechanism-informed hypergraph priors, which are subsequently refined and utilized by a lightweight forecasting network. This design effectively separates knowledge extraction from numerical forecasting, avoiding repeated LLM inference and the associated context overhead while maintaining superior predictive performance in our experiments.

Table R3: Performance comparison of LHC-Net, TimeLLM, TimeLLM-RAG, and TimeLLM-Infer-RAG at H=96 and H=192. Best results are highlighted in bold. "-" indicates unavailable results due to Out-Of-Memory (OOM).

ModelDatasetH=96 (MSE / MAE)H=192 (MSE / MAE)
LHC-NetSub10.267 / 0.3000.318 / 0.345
Sub23.753 / 0.5074.201 / 0.545
Sub33.743 / 0.4604.140 / 0.494
Exhaust Gas0.794 / 0.3140.899 / 0.374
Vessel Control0.533 / 0.3490.662 / 0.426
TimeLLMSub10.371 / 0.3640.424 / 0.397
Sub24.318 / 0.5646.048 / 0.647
Sub34.166 / 0.4995.857 / 0.576
Exhaust Gas1.251 / 0.3763.005 / 0.499
Vessel Control0.592 / 0.3550.727 / 0.431
TimeLLM-RAGSub10.888 / 0.691- / -
Sub25.723 / 0.889- / -
Sub35.664 / 0.842- / -
Exhaust Gas1.294 / 0.586- / -
Vessel Control0.705 / 0.475- / -
TimeLLM-Infer-RAGSub10.989 / 0.7491.026 / 0.771
Sub27.354 / 0.9248.205 / 1.001
Sub37.212 / 0.8417.965 / 0.882
Exhaust Gas1.915 / 0.7922.401 / 0.964
Vessel Control1.131 / 0.6631.797 / 0.923
Reviewer_KuG9OpenReview ↗
Rating
2
Confidence
2(−1)
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes LHC-Net, a framework for multivariate biochemical time-series forecasting that integrates LLM-derived structural priors with hypergraph-based feature learning and temporal channel retrieval. The method consists of three components: LLM-based Structural Initialization (LSI), Hypergraph Interaction Encoder (HIE), and Channel-Retrieval Decoder (CRD). LSI initializes variable-to-hyperedge associations using domain knowledge, HIE models high-order interactions among biochemical variables, and CRD employs temporal queries to retrieve variable representations for forecasting. Experiments on a proprietary erythromycin fermentation dataset and the public IndPenSim benchmark demonstrate competitive predictive performance.

Critical strengths
  1. Well-motivated problem formulation.The paper addresses a meaningful challenge in biochemical time-series forecasting by considering the trade-off between capturing cross-variable physical dependencies and reducing interference from irrelevant correlations. The motivation is clearly presented and relevant to real-world biochemical processes.
  2. Integration of domain knowledge and data-driven learning.The proposed framework incorporates LLM-derived structural priors into a learnable hypergraph, allowing the initial mechanistic relationships to be refined using observational data. This provides a practical approach to combining external scientific knowledge with data-driven forecasting.
  3. Comprehensive experimental evaluation. The authors evaluate LHC-Net against a diverse set of baselines on both proprietary industrial data and a public benchmark. The experiments include component ablations, alternative structural initializations, statistical significance tests, and visualizations of learned hypergraph structures, providing multiple perspectives on the method's effectiveness.
Critical weaknesses
  1. Limited methodological novelty. The HIE mainly relies on conventional hypergraph message passing and self-attention, while the CRD employs standard cross-attention. The methodological distinctions from existing hypergraph-based forecasting and adaptive channel interaction methods remain unclear.
  2. Inconsistent sparsity formulation. Although the authors claim that assigning -Smax enforces structural isolation, the Softmax normalization in Eq. (6) produces strictly positive weights. This allows soft suppression but does not guarantee a strictly sparse hypergraph.
  3. Insufficient validation of hallucination correction. The learned topology converging toward that of standard LHC-Net does not necessarily establish physical correctness. Independent biochemical evidence is needed to substantiate the claimed mechanistic fidelity.
Questions for authors
  1. What distinguishes the proposed HIE and CRD from existing hypergraph-based forecasting methods and conventional cross-attention mechanisms?
  2. How is structural sparsity enforced given the strictly positive Softmax weights in Eq. (6)? Is additional masking or sparsification applied?
  3. Can the authors quantitatively validate the recovered hypergraph against independently established biochemical relationships?
  4. In Section 3.4.2, the causal mask is defined as (Mij=0) for (i>j) and (-∞) otherwise. This definition excludes self-attention at the current time step and leaves the first query position with no valid attention entries, potentially resulting in undefined Softmax outputs. Could the authors clarify whether this is a typographical error or whether a special handling mechanism is implemented?
9 comments
AuthorsOpenReview ↗

1. Response to W1 & Q1: Methodological Novelty of HIE and CRD

Thank you for this important observation. We acknowledge that HIE and CRD build upon established message-passing and attention operations. However, the methodological novelty of LHC-Net lies in the integration of mechanistically guided hypergraph encoding and dynamic temporal channel retrieval, which jointly addresses the trade-off between physical dependency preservation and noise interference in MTSF.

Methodological Distinctions:

  • Hypergraph Interaction Encoder (HIE): Unlike conventional data-driven hypergraph learning, HIE incorporates LLM-derived biochemical priors to initialize learnable variable-to-mechanism associations. Through node-to-hyperedge aggregation, mechanistic interaction, and hyperedge-to-node reconstruction, HIE captures higher-order physical dependencies while allowing the topology to be refined by observational data.

  • Channel-Retrieval Decoder (CRD): While employing standard cross-attention operations, CRD introduces a mechanism-aware retrieval formulation: known future temporal features serve as queries, while HIE-generated mechanistic channel representations provide keys and values. This enables time-dependent retrieval of relevant physical dependencies, rather than relying on static channel fusion.

Experimental Evidence: The component ablations in Appendix G.2 (Ablation Study Result) demonstrate the contributions of both HIE and CRD. Furthermore, our newly added Section 4.5 (Event-Driven Physical Consistency Analysis, Figure 3) illustrates their synergy: HIE associates substrate and biomass variables with a shared mechanistic group, while CRD dynamically increases attention to substrate availability during the observed depletion phase. This provides interpretable evidence of event-aligned channel retrieval.

We have also expanded the baseline comparisons in Section 4.3 (Main Results) and Appendix E (Complete Main Results under Different Horizons). For ease of reference, the key additional results are also summarized in our response to R-yoYD (W1 & Q1).

AuthorsOpenReview ↗

2. Response to W2 & Q2: Clarification of Structural Sparsity

Thank you for identifying this important mathematical inconsistency. We agree that Softmax normalization in Eq. (6) produces strictly positive weights, resulting in soft structural suppression rather than mathematically exact sparsity, consistent with the distinction discussed by Martins & Astudillo [1] and Peters et al.[2]

Mechanism Clarification: Assigning (−Smax⁡)(-S_{\max}) to unrelated variable–hyperedge pairs reduces their relative weights after Softmax normalization, thereby attenuating their contributions during HIE message passing. We deliberately adopt this continuous, differentiable formulation without additional hard masking or thresholding, allowing the hypergraph topology to be refined through gradient-based optimization. However, we acknowledge that this mechanism does not guarantee exact zero connections.

Manuscript Revisions: To thoroughly address this concern, we have systematically revised the terminology across the manuscript, rather than merely correcting the local description:

  • Abstract & Conclusion: Replaced "sparse hypergraph priors" with "mechanism-informed hypergraph priors" to accurately characterize our contribution without implying strict sparsity.
  • Section 3.2.1 (Structural Prior Elicitation): Replaced the original sparsity claims with "mechanistically structured hypergraph topology with selective connection weights", and clarified that −Smax⁡-S_{\max} induces soft structural suppression rather than hard isolation.

These revisions ensure that our methodological claims consistently reflect the actual mathematical formulation while preserving the intended mechanism-guided channel interaction design.

[1] Martins, A., & Astudillo, R. (2016, June). From softmax to sparsemax: A sparse model of attention and multi-label classification. In International conference on machine learning (pp. 1614-1623). PMLR.

[2] Peters, B., Niculae, V., & Martins, A. F. (2019, July). Sparse sequence-to-sequence models. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics (pp. 1504-1519).

AuthorsOpenReview ↗

3. Response to W3 & Q3: Biochemical Validation of the Learned Hypergraph

Thank you for this important observation. We agree that topological convergence alone does not establish mechanistic correctness. To address this concern, we highlight two complementary levels of biochemical validation, combining our existing quantitative expert evaluation with a newly added event-driven analysis.

1. Expert-Referenced Quantitative Validation (Existing Evidence):

As previously detailed in Appendix J (LSI Robustness and Feasibility Testing, Table 8), we have quantitatively evaluated the LLM-generated structural priors against independently annotated variable–mechanism associations provided by domain experts.

Specifically, we employ Quadratic Weighted Kappa (QWK) to measure scoring agreement and Normalized Discounted Cumulative Gain (NDCG) to assess the ranking consistency of key mechanistic drivers. The results demonstrate strong agreement between LLM-generated priors and expert knowledge. The detailed expert–LLM alignment process is further illustrated in Appendix K (LLM Reasoning Data Flow and Expert Alignment, Figure 10).

2. Event-Driven Physical Consistency Analysis (New Evidence):

To further assess the physical plausibility of the learned topology, we have added Section 4.5 (Event-Driven Physical Consistency Analysis, Figure 3), with detailed biochemical interpretations in Appendix P (Detailed Biochemical Interpretation of Dynamic Retrieval).

Focusing on substrate limitation during the stationary phase (Sub3), we examine the relationship between Total Sugar (σ) and Bacterial Density (ϕ):

  • Structural Consistency (Figure 3a): The learned HIE assigns both variables strong membership to the same mechanistic hyperedge, consistent with their established biochemical coupling.
  • Temporal Consistency (Figure 3b–c): The CRD attention toward substrate availability increases during the observed depletion phase and subsequently decreases as the process stabilizes, exhibiting temporal alignment with the underlying biochemical transition.

Together, these findings provide complementary evidence of expert-aligned initialization and physically interpretable learned dynamics, extending our validation beyond topology convergence alone.

For transparency, we have also uploaded the new event-driven analysis materials to our anonymous repository to facilitate independent inspection and reproducibility.

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

This paper proposes LHC-Net, a framework for multivariate time series forecasting in biochemical processes. It uses an LLM augmented with domain knowledge to initialize a learnable hypergraph linking process variables to expert-defined functional groups. A hypergraph encoder captures group-level dependencies, while a channel-retrieval decoder uses future temporal context to selectively aggregate variable information. Experiments on industrial erythromycin fermentation data and public penicillin simulation data show improved forecasting performance over competitive baselines. Ablation studies and structural visualizations further examine the contributions of individual components and the model’s ability to refine inaccurate structural priors.

Critical strengths

This paper addresses an important problem in AI for Science: incorporating domain knowledge into multivariate forecasting of complex biochemical processes while retaining the flexibility to learn from observational data. The method is carefully developed, combining LLM-based structural initialization, hypergraph modeling, and temporal channel retrieval. Extensive experiments on industrial and simulated fermentation datasets, complemented by ablation studies, robustness analyses, and structural visualizations, provide substantial empirical support for the framework.

Critical weaknesses
  1. Insufficient discussion of closely related work. Several core ideas have clear precedents that are not cited or adequately discussed: LLM-initialized Differentiable Causal Discovery (2024) initializes learnable graph structures with LLM priors before data-driven optimization; RealTCD (2024) introduces LLM-guided initialization for temporal causal discovery in industrial systems; and HyperMixer (AAAI 2025) uses channel-to-group and group-to-channel hypergraph message passing to model higher-order, time-varying channel relationships. The authors should clarify the methodological advances beyond these approaches and include direct comparisons where applicable.
  2. The specific contribution of the LLM is difficult to isolate. Experts define the functional groups, and the knowledge base incorporates domain knowledge and industrial experience. Consequently, the benefits of LLM-generated initialization are difficult to separate from those of expert guidance. Comparisons with uniform, random, and reversed initialization do not establish whether an LLM is necessary. A comparison against expert-assigned weights under the same architecture would help quantify its added value.
  3. Limited evidence of generalizability and reproducibility. Although the paper targets biochemical forecasting broadly, both data sources concern fermentation processes. The experiments therefore provide limited evidence of applicability to other biochemical systems. Moreover, the primary industrial dataset cannot be fully released, restricting independent verification.
Questions for authors
  1. How does LHC-Net advance beyond LLM-guided structure initialization methods such as LLM-DCD and RealTCD, and hypergraph channel modeling methods such as HyperMixer? Which component constitutes the main methodological novelty?
  2. How much of the improvement can be attributed specifically to the LLM, rather than expert-defined functional groups and the domain knowledge base? Could the authors compare LLM-generated initialization with expert-assigned weights under the same architecture and training settings?
6 comments
AuthorsOpenReview ↗

1. Response to W1 & Q1: Closely Related Work, Methodological Distinctions, and Direct Comparisons

Thank you for highlighting these relevant works. We have expanded Section 2.1 (Related Work) to clarify our methodological distinctions and incorporated HyperMixer as a new baseline.

Methodological Distinctions:

  • RealTCD / LLM-DCD: These methods leverage LLM-derived knowledge to guide causal structure learning, with RealTCD specifically addressing temporal causal discovery. In contrast, LHC-Net targets multivariate time series forecasting (MTSF) rather than causal graph identification. Beyond structural initialization, our Channel-Retrieval Decoder (CRD) uses future temporal contexts to dynamically retrieve HIE-encoded channel representations, integrating mechanism-informed structural priors with time-dependent forecasting requirements.

  • HyperMixer: While HyperMixer captures higher-order channel interactions through data-driven routing, LHC-Net introduces LLM-derived biochemical priors to initialize a learnable hypergraph, providing mechanistic guidance for channel interactions and helping mitigate cold-start challenges and spurious correlations in noisy biochemical data.

Direct Comparisons:

We have additionally implemented and evaluated two HyperMixer variants to ensure a comprehensive and fair comparison:

  • HyperMixer-D (DLinear backbone): The best-performing configuration reported in the original AAAI 2025 paper.
  • HyperMixer (PatchTST backbone): The variant achieving superior performance on our datasets.

Both variants follow the original paper's hyperparameter configurations. We have incorporated HyperMixer's average performance in Section 4.3 (Main Results), detailed implementation settings in Appendix D.2 (Baseline Implementations), and complete results for both variants in Appendix E (Complete Main Results under Different Horizons).

Since the official implementations of HyperMixer and DHSL are not publicly available, we have carefully reproduced both methods and released their implementations in our anonymous repository to facilitate reproducibility. For ease of reference, we also present the key experimental results in the table below, providing a direct comparison with our proposed method.

Table R1: Performance comparison of LHC-Net, HyperMixer, and HyperMixer-D

DatasetModelH=96H=192H=336H=720Avg
Sub1LHC-Net0.267 / 0.3000.318 / 0.3450.369 / 0.3840.432 / 0.4420.347 / 0.368
HyperMixer0.309 / 0.3320.344 / 0.3580.392 / 0.4010.455 / 0.4470.375 / 0.385
HyperMixer-D0.400 / 0.4160.414 / 0.4260.435 / 0.4430.527 / 0.5150.444 / 0.450
Sub2LHC-Net3.753 / 0.5074.201 / 0.5454.386 / 0.5834.658 / 0.6494.249 / 0.571
HyperMixer3.872 / 0.5214.372 / 0.5714.512 / 0.6074.618 / 0.6514.344 / 0.588
HyperMixer-D7.001 / 0.7246.725 / 0.7256.390 / 0.7266.226 / 0.7666.585 / 0.735
Sub3LHC-Net3.743 / 0.4604.140 / 0.4944.268 / 0.5264.518 / 0.5744.167 / 0.513
HyperMixer3.837 / 0.4684.283 / 0.5034.416 / 0.5324.608 / 0.5724.286 / 0.519
HyperMixer-D5.044 / 0.5975.364 / 0.6105.872 / 0.6165.457 / 0.6555.427 / 0.608
Exhaust GasLHC-Net0.794 / 0.3140.899 / 0.3740.982 / 0.4481.130 / 0.5270.951 / 0.416
HyperMixer0.857 / 0.3660.988 / 0.4451.067 / 0.4961.209 / 0.5721.030 / 0.470
HyperMixer-D0.932 / 0.4091.169 / 0.5321.331 / 0.5871.384 / 0.6441.204 / 0.543
Vessel ControlLHC-Net0.533 / 0.3490.662 / 0.4260.743 / 0.4760.884 / 0.5620.706 / 0.453
HyperMixer0.633 / 0.4290.715 / 0.4760.854 / 0.5610.888 / 0.5800.773 / 0.512
HyperMixer-D0.564 / 0.3980.744 / 0.4930.893 / 0.5781.051 / 0.6720.813 / 0.535
AuthorsOpenReview ↗

2. Response to W2 & Q2: Isolating the Contribution of LLM-Generated Priors

Thank you for this insightful question regarding the specific contribution of LLMs. Following your suggestion, we conducted an additional ablation study (Expert-Init), replacing the LLM-generated initialization weights with those manually assigned by our collaborating domain experts, while keeping the architecture, functional groups, and training settings identical.

The complete experimental results and analysis have been added to Appendix Q (Assessing the Role of LLMs).

Key Findings:

  • Proprietary Datasets (Sub1–3): LHC-Net achieves comparable performance to Expert-Init, demonstrating that LLM-generated priors can effectively approximate expert-assigned initialization. This observation is consistent with our previous evaluation in Table 8 (Appendix J: LSI Robustness and Feasibility Testing), where high QWK scores indicate strong agreement between LLM-generated assessments and human expert judgments.

  • Public Dataset (IndPenSim): LHC-Net outperforms Expert-Init. One possible explanation is that our collaborating experts specialize in erythromycin fermentation, while IndPenSim involves penicillin production with different process characteristics. Consequently, their manually assigned weights may be less transferable, whereas LLMs can draw upon broader scientific knowledge to provide more adaptable initialization.

These findings highlight the practical value of LLMs beyond expert-guided initialization: rather than necessarily outperforming specialists within their own domains, LLMs offer a scalable alternative that reduces reliance on scenario-specific expert intervention while maintaining competitive or superior forecasting performance.

For clarity, we also provide the key experimental results below for direct comparison.

Table R2: Performance comparison of LHC-Net and Expert-Init across five sub-datasets. Best results are highlighted in bold.

ModelDatasetH=96 (MSE / MAE)H=192 (MSE / MAE)H=336 (MSE / MAE)H=720 (MSE / MAE)
LHC-NetSub10.267 / 0.3000.318 / 0.3450.369 / 0.3840.432 / 0.442
Sub23.753 / 0.5074.201 / 0.5454.386 / 0.5834.658 / 0.649
Sub33.743 / 0.4604.140 / 0.4944.268 / 0.5264.518 / 0.574
Exhaust Gas0.794 / 0.3140.899 / 0.3740.982 / 0.4481.130 / 0.527
Vessel Control0.533 / 0.3490.662 / 0.4260.743 / 0.4760.884 / 0.562
Expert-InitSub10.262 / 0.2980.323 / 0.3470.364 / 0.3820.438 / 0.444
Sub23.830 / 0.5044.326 / 0.5534.436 / 0.5804.658 / 0.644
Sub33.742 / 0.4504.121 / 0.4934.346 / 0.5204.547 / 0.574
Exhaust Gas0.806 / 0.3670.938 / 0.4321.056 / 0.4901.143 / 0.550
Vessel Control0.607 / 0.3940.735 / 0.4670.808 / 0.5190.912 / 0.584
AuthorsOpenReview ↗

3. Response to W3: Generalizability and Reproducibility

Thank you for raising these important concerns regarding generalizability and reproducibility.

Generalizability: Although our datasets originate from fermentation-related processes, they cover distinct physical, chemical, and biological mechanisms. As detailed in Appendix C.2 (Detailed IndPenSim Dataset Description), the public IndPenSim benchmark includes two subsystems beyond core fermentation dynamics:

  • Vessel Control: Focuses on thermodynamic and fluid-dynamic processes, including heating/cooling water flow rates, vessel weight, and heat generation.
  • Exhaust Gas: Captures gas-exchange dynamics, including CO₂ emission and oxygen uptake rates.

LHC-Net's consistent performance across these distinct subsystems provides evidence of its adaptability to heterogeneous process dynamics, although broader validation across other biochemical industries remains valuable future work.

Reproducibility: While the proprietary EPD dataset cannot be publicly released due to industrial confidentiality agreements, we have made extensive efforts to maximize transparency:

-Public Companion Dataset: As detailed in Appendix C.1 (Detailed EPD Description, Paragraph 3), our industrial partner has released a companion dataset on Zenodo. Both datasets originate from the same factory and share the same process variables, but differ in collection period and scale: the companion dataset is research-oriented, whereas EPD contains records from actual large-scale industrial operations. Although the companion dataset cannot fully replace EPD, its public availability provides a valuable resource for examining representative process characteristics and supporting independent validation.

  • Data and Code Availability: We have uploaded all publicly shareable datasets as complete CSV files, together with our implementation and reproduced baselines (including HyperMixer and DHSL), to our anonymous GitHub repository.

These resources facilitate independent evaluation and reproduction of our methodology on publicly available data, while respecting industrial confidentiality constraints.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.