Pith. sign in

REVIEW 4 major objections 4 minor 35 references

LLM-Guided Task-Semantic Field Factorization for Industrial Process Forecasting

T0 review · 4 major / 4 minor · reviewed 2026-08-02 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read TSF converts process documents into a frozen semantic input prior that reduces forecasting MAE by 3.6% on average.

desk verdict Useful empirical adapter, but the 'semantic constraint' is mathematically vacuous at k=128, and the paper's interpretation needs to change. read the letter →

arxiv 2607.06623 v2 pith:SEA2UDIG submitted 2026-07-07 cs.LG cs.AIcs.SYeess.SY

Youcheng Zong , Runda Jia , Mingxuan Ren , Dakuo He This is my paper
classification cs.LGcs.AIcs.SYeess.SY
keywords task-semanticfieldfactorizationLLM-guidedsemanticpriorindustrialprocessforecastingsoftsensingvariablesemanticsinputtime-seriesbackboneslow-rankadapter
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that static process documents—variable names, units, physical meanings, and process roles—can be turned into a lightweight input prior that improves time-series forecasting and soft sensing across standard backbones. The proposed TSF method uses an LLM only once, before training, to write a semantic card for each input variable, embeds those cards into frozen direction vectors, and lets the current numerical window multiply by those directions to form a task-semantic field. A constrained adapter then feeds both the raw values and the semantic field into the backbone. The paper reports a 3.6% average MAE reduction across four industrial tasks, with gains largest when test operating conditions shift, and online overhead under 8 microseconds per sample. If true, existing plant documents become measurable forecasting gains without any online LLM use.

What carries the argument

The carrying mechanism is the variable-semantic direction matrix V and the constrained input factorization Z = X̃(D + V B) + 1_L b^⊤. Before training, an LLM writes a semantic card for each input variable, a validation step checks the cards against data dictionaries, and an embedding model turns the cards into normalized rows of V. During training and inference, the normalized window X̃ multiplies V to form the task-semantic field S = X̃V, so the current values of variables activate their semantics. The adapter preserves a per-variable diagonal path D and mixes semantics through a learned projection B, producing a structured low-rank map that any window-based backbone can consume.

What would settle it

Run the same TSF adapter on the datasets with operating shifts using semantic, random, and learnable direction matrices while keeping all else fixed; if random or learnable directions reproduce the MAE gains of semantic directions on most dataset–backbone pairs, the claim that LLM-derived variable semantics drive the improvement is falsified.

Watch

Extended reading notes

Core claim

On its own terms, the paper claims that semantic-logical relations between input variables and the prediction target, extracted once from process documents by an LLM, can be made visible to a numerical forecaster inside every prediction window. The concrete discovery is that representing those relations as frozen normalized direction vectors and inserting them through the constrained product V B yields consistent MAE reductions—3.6% on average across four industrial tasks, 2.9% macro-average across 32 dataset–backbone pairs, and a maximum reduction of 24.9%—while adding only about 0.7–4.3k parameters and under 8 microseconds per sample of online inference time. Gains are most consistent on d

Load-bearing premise

The improvement depends on the frozen semantic direction matrix V carrying task-relevant variable–target relations that are better than what a generic learned direction matrix would provide; if random or learnable directions match the semantic ones on most tasks, the 'semantic' part of the claim collapses and only the low-rank input adapter remains.

Editorial extensions

If this is right

  • Any standard recurrent, attention, state-space, or convolutional forecasting backbone can be improved simply by replacing its input window with the factorized representation; no backbone changes are required.
  • The gains concentrate on delayed soft-sensing and regime-shifted test sets, so TSF offers a way to make models generalize across operating shifts using static documents.
  • Because the LLM and embedding run only offline, deployment cost stays low—hundreds to thousands of extra parameters and microseconds per sample.
  • The method depends only on variable tables, units, and process descriptions that many plants already maintain, so it can be applied without new sensors or paired text corpora.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: if the LLM's main contribution is a stable direction geometry rather than deep domain reasoning, then cheaper deterministic sources—hand-curated ontologies or structured data dictionaries—might capture most of the gain on tasks with small distribution shifts; the paper's Ladle ablation already shows learnable directions matching semantic ones there.
  • Editorial inference: the same factorization should transfer to other tabular regression settings with variable metadata, such as energy-load forecasting, quality control, or soft sensors; a natural test is to apply TSF to a non-industrial dataset with rich column descriptions.
  • Editorial inference: the sensitivity to the embedding model suggests that geometry among variable directions is what matters, not the raw text generator; comparing embeddings built from domain ontologies against generic LLM embeddings on the same semantic cards would be a direct test.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper proposes Task-Semantic Field Factorization (TSF), an offline LLM-based framework for industrial time-series forecasting and soft sensing. Before training, an LLM converts task protocols and variable documents into semantic cards, which are embedded into a frozen direction matrix V (Eq. (5)). During training and inference, the normalized window X̃ is transformed into Z = X̃(D + V B) + 1_L b^⊤ (Eq. (10)), where D is a learnable diagonal scaling, B is a learnable projection, and b a bias; Z is then fed to a standard backbone. The authors claim that the frozen semantic directions constrain the input map and that the current numerical window 'activates' variable semantics. Experiments cover four industrial datasets (ladle preheating, thickener dewatering, IndPenSim, Tennessee Eastman Process) across eight backbones, reporting MAE reductions (3.6% average, 2.9% macro), low parameter overhead, and small online latency. The central claim is that the semantic factorization, not merely extra input capacity, drives the gains.

Significance. If the semantic-factorization claim held, TSF would be a practical, backbone-agnostic way to inject process documents into forecasting at negligible online cost. The experimental protocol is genuinely controlled: fixed train/validation/test splits at process or scenario level, shared backbones and budgets, 32 dataset–backbone pairs, ablations, sensitivity analysis, and source code. The paper also correctly keeps V frozen and uses the test set only for evaluation, so there is no direct label-fitting circularity. However, the mathematical core of the paper — that Eq. (10) is a semantically constrained input map — is false at the reported hyperparameters (k = 128, d ≤ 33). The same function class as a free full-matrix input layer is obtained, so the ablation results cannot establish that semantic content, rather than optimization/initialization effects, is responsible for the gains. This is a load-bearing issue that must be resolved before the paper's central contribution can be accepted.

major comments (4)
  1. [§III-B, Eq. (10); Fig. 3; Table X] The paper calls D+VB a 'constrained' map and contrasts it with an arbitrary dense W. But with k=128 and d≤33, V is d×k and has more columns than rows. Unless rank(V) < d, V has a right inverse, so for any W ∈ R^{d×d} there exists B with VB = W. Hence D+VB spans all d×d matrices, and the TSF input map is exactly the same function class as the 'free full matrix' baseline. The 'free full matrix' row in Table X (IndPenSim–ModernTCN: 2.429 vs 2.028) therefore cannot demonstrate a representational advantage. Differences among semantic, random, learnable, and full-matrix variants in Table X are optimization-trajectory or initialization effects unless the experiment is run in a genuinely rank-constrained regime (k < d). The paper does not report rank(V) or its singular-value spectrum, and it never tests k < d. This invalidates the central 'semantic constraint' interpretation of Eq. (10) and Fig.
  2. [§IV-E, Table X] Even if TSF is reframed as an initialization or parameterization prior, the evidence for semantic content is weak and uneven. On IndPenSim–ModernTCN, random directions (2.340) and learnable directions (2.288) are close to each other and to the raw backbone (2.320), while the semantic direction reaches 2.028. On Ladle, all direction-source variants are within noise (0.017 vs 0.016/0.017). Only one of four ablation cells clearly favors semantic directions. To support the claim that the LLM's semantic content — rather than the factorization's optimization bias — drives the gains, the authors should report (i) rank(V) and the effective dimension of the V B term, (ii) results with k < d where the factorization is actually constrained, and (iii) multiple random draws of V to characterize the optimization prior and its variance.
  3. [§IV-C, Tables V–VIII] The headline numbers need a precise definition and a more careful statistical account. The abstract states a 3.6% average reduction and a 2.9% macro-average across 32 pairs; the relation between these two numbers is not explained (is the 3.6% sample-weighted or dataset-averaged?). More importantly, the asterisks in Tables V–VIII denote p<0.05 but the direction is not controlled: in Table V, LSTM +TSF (0.018±0.0008) is marked * even though Base is better (0.017±0.0009), and iTransformer shows the same pattern (0.023 vs 0.020). The text says TSF reduces MAE for seven of eight IndPenSim backbones and all eight TEP backbones, but it does not report how many of the 32 paired comparisons were significant improvements versus significant degradations. Please provide paired effect sizes, counts of significant wins/losses, and a discussion of the one very large gain (IndPenSim–iTransformer: 24.9%)
  4. [§IV-B, §IV-E, and offline construction] For the two public benchmarks (TEP and IndPenSim), the LLM and the text-embedding model may have memorized widely available process descriptions, including variable names, units, and fault/control scenarios. The paper does not test for this contamination, so the semantic directions V could encode target-related information that would not be available for a genuinely new private process. This is not an equation-level circularity, but it is a correctness risk for the claim that the gains come from 'task semantics' rather than from memorized benchmark knowledge. A concrete test would be to rebuild V from deliberately incomplete or renamed variable descriptions (e.g., replacing variable tags with generic identifiers) and check whether the TSF gains persist on the public datasets; the private datasets already provide a partial check, but the public results are the ones with the largest and mo
minor comments (4)
  1. [General notation] The asterisk in Tables V–VIII is defined only as 'p<0.05 for MAE'; it should state explicitly that the difference can be in either direction, or better, use separate symbols for significant improvement and significant degradation.
  2. [§IV-G, Fig. 8] The sensitivity curves are reported as relative MAE with respect to the default setting. Please state the normalization base explicitly for each panel and include absolute MAE values for at least the default k=128 and the neighboring k values, since the non-monotonic pattern is central to the choice k=128.
  3. [Table II and Fig. 2] There are minor typos and formatting issues: the author name 'Y oucheng Zong' in the header, 'JUL Y 2026' in the preprint footer, and stray 'R' glyphs in Fig. 2. These should be cleaned up.
  4. [§V and Table IV] The conclusion says TSF 'supports adaptation to different prediction targets,' but every dataset has a fixed target; the paper does not demonstrate adaptation across targets within a dataset. Either soften this phrasing or add an experiment where the same backbone and semantic pipeline are retargeted.

Circularity Check

1 steps flagged · score 6.0 of 10

Vacuous semantic constraint: Eq. (10) equals free input mixing when V is full-row-rank, so the semantic factorization is a reparameterization of the dense adapter.

  1. renaming known result [Section III-B, Eq. (10), Fig. 3; Table X ablation; Table III settings]
    "This equivalent form shows that cross-variable mixing comes from the product V B of the frozen variable-semantic directions V and the learnable projection B, while the diagonal term D preserves each variable’s numerical channel. Fig. 3 compares this constrained map with free input mixing. A free layer learns an arbitrary matrix W∈R^{d×d}, whereas TSF restricts the map to the diagonal residual D and the semantically constrained product V B."

    With k=128 (Table III) and d≤33 (Table IV), a generic V∈R^{d×k} has full row rank and hence a right inverse. For any W∈R^{d×d}, setting D=diag(diag(W)) and B=V^T(VV^T)^{-1}(W−D) yields D+VB=W. Thus Eq. (10) is exactly the 'free input mixing' map X̃W+1_L b^⊤ of Fig. 3(a); the claimed semantic restriction is empty. The semantic-vs-random and free-matrix ablations in Table X therefore compare optimization trajectories, not representable functions, so the paper's conclusion that semantic directions 'guide cross-variable mixing' reduces to a reparameterization of the unconstrained adapter.

full rationale

The empirical MAE improvements are not circular in the usual sense: V is frozen before training, test sets are only used for evaluation, and the self-citations [2],[8],[16],[32],[33] are background/dataset sources, not load-bearing derivations. However, the central architectural claim—that V semantically constrains the input map—is void by construction when V has full row rank. Since k=128 and d≤33, the paper's Eq. (10) is mathematically equivalent to a free dense linear layer, so the 'constrained factorization' is a reparameterization of the unconstrained adapter. The reported differences between semantic, random, learnable, and free-matrix variants therefore cannot be attributed to semantic content without controlling for optimization effects; the paper never reports rank(V) or checks this equivalence. This is a construction-level circularity in the explanatory claim, while the benchmark numbers themselves remain independent empirical results.

Assumptions & free parameters 1 free parameters · 4 assumptions · 1 invented entities

The method's stated gains rest on the assumptions that LLM-generated variable semantics are accurate, that a linear value-weighted sum of semantic vectors captures task relations, and that comparisons to baselines are unconfounded. These are reasonable but not established outside the paper.

free parameters (1)
  • Semantic dimension k = 128
    Dimension of the frozen semantic-direction matrix V; chosen as default in Table III. Fig. 8 shows performance is non-monotonic in k and k=256 can raise relative MAE to ~1.17 on IndPenSim-ModernTCN.
assumptions (4)
  • domain assumption The embedding model Eψ produces a vector space in which value-weighted sums of variable-semantic directions are semantically meaningful (Eqs. 6–7)
    The paper defines S = X̃V and claims it 'activates' semantics, but this linear additive structure is asserted, not derived from process physics.
  • domain assumption LLM-generated semantic cards, after pre-freezing validation, correctly describe variable–target relations and process roles
    Section III-A: Eq. (3) and Table II. Errors in semantic cards would misdirect the frozen V and degrade performance; validation relies on structured records that may be incomplete.
  • domain assumption No test-set information enters semantic construction or normalization
    Section III-C explicitly excludes test statistics, but LLM pretraining may have memorized the public TEP and IndPenSim process descriptions, an unaddressed leakage channel.
  • domain assumption The experimental protocol equally favors Base and +TSF
    Section IV-A: same split, optimizer, budget; however TSF adds parameters and the model-selection rule is shared, so any benefit of extra capacity is not penalized.
invented entities (1)
  • Task-semantic field S_{n,t} = X̃_{n,t} V
    purpose: Represent the current numerical window in a semantic direction space before the backbone
    Defined by Eq. (7); internal latent quantity with no external observable handle; Fig. 7 visualizations are post-hoc analyses, not independent evidence.

how reviews work

0 comments
Cite this review

Pith. "Pith review of LLM-Guided Task-Semantic Field Factorization for Industrial Process Forecasting." pith.science (2026). https://pith.science/paper/SEA2UDIG

@misc{pith2026260706623,
  author       = {Pith},
  title        = {Pith review of: LLM-Guided Task-Semantic Field Factorization for Industrial Process Forecasting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SEA2UDIG}},
  note         = {Machine review of arXiv:2607.06623}
}
abstract

Process industries rely on time-series forecasting and soft sensing to estimate quality variables that are hard to measure online. Labeled data are scarce, operating regimes change frequently, and retraining models or rebuilding alignment pipelines for each scenario is costly. Such settings often provide variable tables and process documents that record variable names, units, physical meanings, and process roles. However, standard time-series backbones usually treat inputs as anonymous numerical columns. Existing text-enhanced methods also rarely make the semantic-logical relations between input variables and the prediction target available to the model within each numerical window. To address this problem, this article proposes Task-Semantic Field Factorization (TSF), a large language model (LLM)-guided framework. TSF builds a task-semantic field from task protocols and variable documents before training and uses the LLM only for offline semantic construction. Online training and inference are handled by conventional time-series backbones. During training and inference, the current numerical window activates variable semantics, so semantic information participates in each prediction and supports adaptation to different prediction targets and operating shifts. Across multiple complex industrial forecasting and delayed soft-sensing tasks, TSF reduces MAE by 3.6\% on average. Across all dataset--backbone pairs, the macro-average reduction is 2.9\%, with a maximum reduction of 24.9\%. It adds only about 0.7--4.3k parameters, with less than 8\,$\mu$s/sample of additional online inference overhead. These results show that TSF turns existing process documents into measurable forecasting gains across backbones and semantic generators while remaining lightweight for deployment.

Figures

Figures reproduced from arXiv: 2607.06623 by the authors.

Figure 1
Figure 1. Paradigm transition for variable semantics in industrial forecasting. (a) Conventional numerical forecasting only processes sensor-value windows, leaving variable documents outside the model. (b) LLM-based and text-augmented forecasting introduces semantic guidance, but the link to the current numerical window can still remain loose. (c) TSF first turns task protocols and variable descriptions into task-aware semant… view at source ↗
Figure 2
Figure 2. TSF workflow. Offline construction maps the task protocol and ordered variable metadata to semantic cards, rendered phrases, embeddings, and the frozen variable-semantic direction matrix V . During training and inference, the normalized window activates the task-semantic field S = XV˜ . The constrained dual-path adapter combines the semantic field, raw-value path, and shared bias into the backbone input Z; the inset… view at source ↗
Figure 3
Figure 3. Free input mixing versus TSF-constrained input factorization. (a) A conventional pre-backbone layer learns an arbitrary dense matrix W and forms XW˜ . (b) TSF uses Z = X˜(D + V B) + 1Lb⊤, where D preserves per-variable numerical channels and V B constrains cross-variable mixing through the frozen variable-semantic directions V and the learnable projection B. C. Training and Inference After the input adapter forms Zn… view at source ↗
Figures from the paper (9 more)
Figure 3
Figure 3. Figure 3: Free input mixing versus TSF-constrained input factorization. (a) A conventional pre-backbone layer learns an arbitrary dense matrix W and forms XW˜ . (b) TSF uses Z = X˜(D + V B) + 1Lb⊤, where D preserves per-variable numerical channels and V B constrains cross-variab…
Figure 4
Figure 4. Figure 4: On-site views of the two plant processes. (a) Ladle preheating at a steel plant in Maanshan, China. (b) Tailings thickening dewatering at a plant in Nanjing, China. TABLE III QUANTITATIVE COMPARISON ON THE LADLE PREHEATING DATASET. ACCURACY METRICS REPORT THE MEAN AND …
Figure 4
Figure 4. Figure 4: On-site views of the two plant processes. (a) Ladle preheating at a steel plant in Maanshan, China. (b) Tailings thickening dewatering at a plant in Nanjing, China. TABLE VI QUANTITATIVE COMPARISON ON THE THICKENER DEWATERING DATASET. ∗ DENOTES p < 0.05 FOR MAE. Backbo…
Figure 5
Figure 5. Figure 5: Qualitative comparison on ladle preheating and IndPenSim. The top row gives fifth-step ladle-temperature forecasts, 30 min ahead, for three ladle processes; insets enlarge high-temperature regions. The bottom row gives assay-time penicillin-concentration estimates for …
Figure 5
Figure 5. Figure 5: Qualitative comparison on ladle preheating and IndPenSim. The top row gives fifth-step ladle-temperature forecasts, 25 min ahead, for three ladle processes; insets enlarge high-temperature regions. The bottom row gives assay-time penicillin-concentration estimates for …
Figure 6
Figure 6. Figure 6: Runtime task-semantic field analysis on IndPenSim-ModernTCN. (a) PCA state atlas of raw normalized input windows. (b) Runtime task￾semantic field atlas formed by S = XV˜ . (c) Semantic intervention intensity ρ = ∥SB∥/(∥XD˜ ∥ + ∥SB∥), measuring the semantic projection r…
Figure 7
Figure 7. Figure 7: Sensitivity analysis of TSF across semantic dimension, semantic-card generator, and embedding model. Values are relative MAE with respect to the default setting in each panel. Shaded bands and error bars denote standard deviations over three runs; gray dashed lines mar…
Figure 7
Figure 7. Figure 7: Runtime task-semantic field analysis on IndPenSim-ModernTCN. (a) PCA state atlas of raw normalized input windows. (b) Runtime task-semantic field atlas formed by S = XV˜ . (c) Semantic intervention intensity ρ = ∥SB∥F /(∥XD˜ ∥F + ∥SB∥F ), measuring the relative Frobeni…
Figure 8
Figure 8. Figure 8: Sensitivity analysis of TSF across semantic dimension, semantic-card generator, and embedding model. Values are relative MAE with respect to the default setting in each panel. Shaded bands and error bars denote standard deviations over five runs; gray dashed lines mark…

Review history (2 revisions) →

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

35 extracted references

  1. [1]

    A survey on deep learning for data-driven soft sensors,

    Q. Sun and Z. Ge, “A survey on deep learning for data-driven soft sensors,”IEEE Transactions on Industrial Informatics, vol. 17, no. 9, pp. 5853–5866, 2021

    2021

  2. [2]

    LLM-driven human-AI collaborative decision support system for complex industrial processes: A case study in metallurgy,

    Y . Zong, R. Jia, K. Li, D. Xue, L. Zhang, and D. He, “LLM-driven human-AI collaborative decision support system for complex industrial processes: A case study in metallurgy,”Neural Networks, vol. 202, p. 109055, 2026

    2026

  3. [3]

    Large language models are zero-shot time series forecasters,

    N. Gruver, M. Finzi, S. Qiu, and A. G. Wilson, “Large language models are zero-shot time series forecasters,” inAdvances in Neural Information Processing Systems 36, 2023

    2023

  4. [4]

    Time-LLM: Time series forecasting by reprogramming large language models,

    M. Jin, S. Wang, L. Ma, Z. Chu, J. Y . Zhang, X. Shi, P.-Y . Chen, Y . Liang, Y .-F. Li, S. Pan, and Q. Wen, “Time-LLM: Time series forecasting by reprogramming large language models,” inProceedings of the International Conference on Learning Representations (ICLR), 2024

    2024

  5. [5]

    AutoTimes: Autoregressive time series forecasters via large language models,

    Y . Liu, G. Qin, X. Huang, J. Wang, and M. Long, “AutoTimes: Autoregressive time series forecasters via large language models,” in Advances in Neural Information Processing Systems 37, 2024

    2024

  6. [6]

    Nonlinear dynamic soft sensor modeling with supervised long short-term memory network,

    X. Yuan, L. Li, and Y . Wang, “Nonlinear dynamic soft sensor modeling with supervised long short-term memory network,”IEEE Transactions on Industrial Informatics, vol. 16, no. 5, pp. 3168–3176, 2020

    2020

  7. [7]

    Novel transformer based on gated convolutional neural network for dynamic soft sensor modeling of industrial processes,

    Z. Geng, Z. Chen, Q. Meng, and Y . Han, “Novel transformer based on gated convolutional neural network for dynamic soft sensor modeling of industrial processes,”IEEE Transactions on Industrial Informatics, vol. 18, no. 3, pp. 1521–1529, 2022

    2022

  8. [8]

    Hybrid grid search and Bayesian optimization-based random forest regression for predicting material compression pressure in manufacturing processes,

    Y . Zong, Y . Nian, C. Zhang, X. Tang, L. Wang, and L. Zhang, “Hybrid grid search and Bayesian optimization-based random forest regression for predicting material compression pressure in manufacturing processes,”Engineering Applications of Artificial Intelligence, vol. 141, p. 109580, 2025

    2025

Show all 35 references
  1. [9]

    Deep learning framework for collaborative variable time delay estimation and uncertainty quantifi- cation in industrial quality prediction,

    L. Yu, W. Yu, Y . Jia, and T. Chai, “Deep learning framework for collaborative variable time delay estimation and uncertainty quantifi- cation in industrial quality prediction,”IEEE Transactions on Industrial Informatics, vol. 21, no. 2, pp. 1960–1969, 2025. 12 PREPRINT, JUL Y 2026

    1960

  2. [10]

    From complexity to clarity: Structural process knowledge-informed neural network for alumina concentration distribution prediction,

    Z. Wang, C. Yang, K. Huang, D. Wu, and W. Gui, “From complexity to clarity: Structural process knowledge-informed neural network for alumina concentration distribution prediction,”IEEE Transactions on Industrial Informatics, vol. 21, no. 9, pp. 7198–7208, 2025

    2025

  3. [11]

    Performance-driven distillation and confident pseudo labeling for semi-supervised industrial soft-sensor application,

    B. Yue, K. Wang, H. Zhu, C. Yang, and W. Gui, “Performance-driven distillation and confident pseudo labeling for semi-supervised industrial soft-sensor application,”IEEE Transactions on Cybernetics, vol. 55, no. 9, pp. 4414–4426, 2025

    2025

  4. [12]

    PromptCast: A new prompt-based learning paradigm for time series forecasting,

    H. Xue and F. D. Salim, “PromptCast: A new prompt-based learning paradigm for time series forecasting,”IEEE Transactions on Knowledge and Data Engineering, vol. 36, no. 11, pp. 6851–6864, 2024

    2024

  5. [13]

    One fits all: Power general time series analysis by pretrained LM,

    T. Zhou, P. Niu, X. Wang, L. Sun, and R. Jin, “One fits all: Power general time series analysis by pretrained LM,” inAdvances in Neural Information Processing Systems 36, 2023, pp. 43 322–43 355

    2023

  6. [14]

    TimeCMA: Towards LLM-empowered multivariate time se- ries forecasting via cross-modality alignment,

    C. Liu, Q. Xu, H. Miao, S. Yang, L. Zhang, C. Long, Z. Li, and R. Zhao, “TimeCMA: Towards LLM-empowered multivariate time se- ries forecasting via cross-modality alignment,” inProceedings of the AAAI Conference on Artificial Intelligence, 2025, pp. 18 780–18 788

    2025

  7. [15]

    TEST: Text prototype aligned embedding to activate LLM’s ability for time series,

    C. Sun, H. Li, Y . Li, and S. Hong, “TEST: Text prototype aligned embedding to activate LLM’s ability for time series,” inProceedings of the International Conference on Learning Representations (ICLR), 2024

    2024

  8. [16]

    Zero-shot capillary segmentation in dermoscopy images via SAM2: A case study on oral mucosa,

    W. Su, Y . Zong, R. Jia, J. Qin, and M. Li, “Zero-shot capillary segmentation in dermoscopy images via SAM2: A case study on oral mucosa,”IEEE Journal of Biomedical and Health Informatics, vol. 30, no. 5, pp. 4376–4387, 2026

    2026

  9. [17]

    Are language models actually useful for time series forecasting?

    M. Tan, M. A. Merrill, V . Gupta, T. Althoff, and T. Hartvigsen, “Are language models actually useful for time series forecasting?” in Advances in Neural Information Processing Systems 37, vol. 37, 2024, pp. 60 162–60 191

    2024

  10. [18]

    A domain knowledge- guided industrial large model framework: A case study in battery health estimation and recycling,

    B. Chen, H. Shao, Y . Qin, Y . Jin, and X. Hu, “A domain knowledge- guided industrial large model framework: A case study in battery health estimation and recycling,”IEEE Transactions on Industrial Informatics, vol. 21, no. 10, pp. 8080–8090, 2025

    2025

  11. [19]

    Socially aware load forecasting utilizing large language models,

    W. Chen, X. Zhang, L. Zhu, J. Shi, Z. Chang, Z. Han, and Y . Zhang, “Socially aware load forecasting utilizing large language models,”IEEE Transactions on Industrial Informatics, vol. 22, no. 1, pp. 231–242, 2026

    2026

  12. [20]

    A 2RA-NSMTSllm: Adversarially aligning retrieval-augmented LLMs for nonstationary multivariate time series forecasting,

    J. Chu, C. Liu, X. Bai, and J. Tan, “A 2RA-NSMTSllm: Adversarially aligning retrieval-augmented LLMs for nonstationary multivariate time series forecasting,”IEEE Transactions on Industrial Informatics, vol. 22, no. 3, pp. 1805–1816, 2026

    2026

  13. [21]

    Causality-aware LLM-enhanced graph representation learning for adaptive power system control,

    F. Yao, J. Liu, Y . Tao, J. Qiu, H. H.-C. Iu, G. Chen, and Z. Y . Dong, “Causality-aware LLM-enhanced graph representation learning for adaptive power system control,”IEEE Transactions on Industrial Informatics, vol. 22, no. 5, pp. 3681–3692, 2026

    2026

  14. [22]

    Zero-shot fault diagnosis via LLM-guided complexity-aware fuzzy boundary learning,

    H. Han, X. Gao, H. Han, H. Gao, and Y . Qi, “Zero-shot fault diagnosis via LLM-guided complexity-aware fuzzy boundary learning,”IEEE Transactions on Industrial Informatics, vol. 22, no. 6, pp. 4705–4716, 2026

    2026

  15. [23]

    Joint knowledge graph and large language model for fault diagnosis and its application in aviation assembly,

    P. Liu, L. Qian, X. Zhao, and B. Tao, “Joint knowledge graph and large language model for fault diagnosis and its application in aviation assembly,”IEEE Transactions on Industrial Informatics, vol. 20, no. 6, pp. 8160–8169, 2024

    2024

  16. [24]

    Learning phrase representations using RNN encoder–decoder for statistical machine translation,

    K. Cho, B. van Merri ¨enboer, C. Gulcehre, D. Bahdanau, F. Bougares, H. Schwenk, and Y . Bengio, “Learning phrase representations using RNN encoder–decoder for statistical machine translation,” inProceed- ings of the 2014 Conference on Empirical Methods in Natural Language Pro...

    2014

  17. [25]

    Long short-term memory,

    S. Hochreiter and J. Schmidhuber, “Long short-term memory,”Neural Computation, vol. 9, no. 8, pp. 1735–1780, 1997

    1997

  18. [26]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,” inAdvances in Neural Information Processing Systems 30 (NeurIPS 2017), 2017, pp. 5998–6008

    2017

  19. [27]

    Informer: Beyond efficient transformer for long sequence time-series forecasting,

    H. Zhou, S. Zhang, J. Peng, S. Zhang, J. Li, H. Xiong, and W. Zhang, “Informer: Beyond efficient transformer for long sequence time-series forecasting,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 35, no. 12, 2021, pp. 11 106–11 115

    2021

  20. [28]

    Mamba: Linear-time sequence modeling with selective state spaces,

    A. Gu and T. Dao, “Mamba: Linear-time sequence modeling with selective state spaces,” inThe First Conference on Language Modeling (COLM), 2024

    2024

  21. [29]

    iTransformer: Inverted transformers are effective for time series fore- casting,

    Y . Liu, T. Hu, H. Zhang, H. Wu, S. Wang, L. Ma, and M. Long, “iTransformer: Inverted transformers are effective for time series fore- casting,” inProceedings of the International Conference on Learning Representations (ICLR), 2024

    2024

  22. [30]

    A time series is worth 64 words: Long-term forecasting with transformers,

    Y . Nie, N. H. Nguyen, P. Sinthong, and J. Kalagnanam, “A time series is worth 64 words: Long-term forecasting with transformers,” inPro- ceedings of the International Conference on Learning Representations (ICLR), 2023

    2023

  23. [31]

    ModernTCN: A modern pure convolution struc- ture for general time series analysis,

    D. Luo and X. Wang, “ModernTCN: A modern pure convolution struc- ture for general time series analysis,” inProceedings of the International Conference on Learning Representations (ICLR), 2024

    2024

  24. [32]

    A meta-contrastive learn- ing hybrid model for adaptive temperature trend prediction in variable ladle preheating,

    Y . Zong, R. Jia, S. Wu, L. Zhang, and D. He, “A meta-contrastive learn- ing hybrid model for adaptive temperature trend prediction in variable ladle preheating,”Engineering Applications of Artificial Intelligence, vol. 162, p. 112750, 2025

    2025

  25. [33]

    Expert-augmented dual-stage reinforcement learning for coordinated optimization of the thickening-dewatering process,

    R. Lin, R. Jia, F. Jiang, D. He, and F. Wang, “Expert-augmented dual-stage reinforcement learning for coordinated optimization of the thickening-dewatering process,”MetaResource, vol. 2, no. 3, pp. 182– 196, 2025

    2025

  26. [34]

    The development of an industrial-scale fed-batch fermentation simulation,

    S. Goldrick, A. S ¸tefan, D. Lovett, G. Montague, and B. Lennox, “The development of an industrial-scale fed-batch fermentation simulation,” Journal of Biotechnology, vol. 193, pp. 70–82, 2015

    2015

  27. [35]

    Additional Tennessee Eastman Process simulation data for anomaly detection evaluation,

    C. A. Rieth, B. D. Amsel, R. Tran, and M. B. Cook, “Additional Tennessee Eastman Process simulation data for anomaly detection evaluation,” Harvard Dataverse, 2017

    2017

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.