Pith. sign in

REVIEW 4 major objections 5 minor 25 references

Pairing an LLM planner with a frozen time-series foundation model as a physics sentinel lets industrial control systems reject unsafe AI actions before they ever actuate.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-13 00:30 UTC pith:SRLY3O3S

load-bearing objection Useful safety pattern (LLM proposes, frozen TimesFM vetoes via linear counterfactuals), but the SWaT numbers are thin and “physics-grounded” is oversold. the 4 major comments →

arxiv 2607.09076 v1 pith:SRLY3O3S submitted 2026-07-10 cs.AI

Neuro-Agentic Control: A Deep Learning-based LLM-Powered Agentic AI Framework for Controlling Security Controls

Saroj Gopali , Bipin Chhetri , Deepika Giri , Sima Siami-Namini , Akbar Siami Namin This is my paper
classification cs.AI
keywords Agentic AITime-series foundation modelsCounterfactual Physics InjectionIndustrial IoT securityHallucination rejectionSecure Water TreatmentLLM-powered controlPhysics-grounded planning
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Large language models can reason about industrial process manuals and invent drain strategies when a tank is about to overflow, yet their tendency to hallucinate makes direct closed-loop control unsafe. This paper shows that the risk can be contained by routing every LLM proposal through a pre-trained time-series foundation model that forecasts the physical consequence of the action before any valve moves. The mechanism, called Counterfactual Physics Injection, simply adds a cumulative linear drain to the foundation model’s forecast so that unsafe or ineffective strategies are scored and discarded. On the Secure Water Treatment dataset under fifteen stochastic attack patterns the resulting Neuro-Agentic Loop prevented five breaches—more than LSTM or TCN baselines—and executed zero physically invalid actions. The claim matters because cyber-physical downtime is measured in millions of dollars per hour and pure agentic AI is currently too unreliable for safety-critical actuation.

Core claim

A pre-trained time-series foundation model can act as a deterministic “Sentinel” that grounds LLM-generated control actions in simulated physics, enabling autonomous defense of industrial processes while automatically rejecting hallucinated or unsafe interventions before they are executed.

What carries the argument

Counterfactual Physics Injection: a closed-form linear perturbation that adds a cumulative drain (magnitude × elapsed steps, capped after duration) to the foundation model’s forecast window so each LLM strategy can be ranked by its worst-case projected peak risk without fine-tuning the model.

Load-bearing premise

The method treats a simple cumulative linear drain added to a foundation-model forecast as faithful enough to real valve and pump physics that ranking actions by the simulated peak correctly predicts plant safety.

What would settle it

Replace the linear injection with a high-fidelity digital twin or physical SWaT plant under the same fifteen attack trials; if the Sentinel then selects actions that cause overflows or rejects actions that would have prevented them, the physics-grounding claim fails.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Agentic AI can be placed in closed-loop industrial control provided every candidate action is first simulated by a frozen time-series foundation model.
  • Physically invalid (hallucinated) control actions can be filtered to zero executions without fine-tuning the foundation model.
  • Breach-prevention rates on tank-level processes can exceed classical LSTM and TCN baselines under sudden-spike attacks.
  • A 1.5–2.5 s decision cycle remains acceptable for fluid processes whose time constants are measured in minutes.
  • The same Sentinel pattern can be reused in any cyber-physical domain that already possesses a reliable time-series forecaster.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The linear-injection approximation will likely need replacement by a calibrated plant simulator before the method scales to multi-actuator systems with hydraulic coupling and sensor lag.
  • The architecture could protect LLM planners in neighboring safety-critical domains—power grids, chemical plants—that already maintain strong forecasting models.
  • Adversarial attacks that deliberately bias the foundation model’s forecast horizon remain an open robustness question the paper itself flags for future work.
  • Quantization or distillation of the foundation model could shrink the decision loop enough for faster electromechanical processes.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes a Neuro-Agentic Control framework that couples an LLM planner (Gemini 2.5 Flash-Lite) with a frozen TimesFM time-series foundation model as a “Sentinel.” When TimesFM’s forecast of SWaT tank level LIT301 exceeds a critical threshold, the LLM proposes structured drain actions; each candidate is scored by Counterfactual Physics Injection—a closed-form linear cumulative offset applied to the TimesFM forecast (Eq. 1)—and the system selects the action minimizing simulated peak risk (Eq. 2), defaulting to Monitor if none improves the baseline. On 15 hand-cycled synthetic attack trials (sudden spike / gradual drift / high noise), the loop reports 5/15 breaches prevented (33.3%) versus LSTM 4/15 and TCN 2/15, higher mean risk reduction (48.51 vs 16.77 / 20.29), and zero executed physically invalid (hallucinated) actions. The authors position TimesFM as a deterministic physics-grounding layer that filters LLM hallucinations before actuation.

Significance. If validated more rigorously, the Architect–Sentinel separation is a useful architectural pattern for safety-constrained agentic control in industrial IoT: it keeps the LLM in a structured planning role and uses a frozen foundation model as a pre-actuation filter rather than as an opaque closed-loop controller. Explicit feasibility checks, structured JSON outputs, RAG from the plant manual, and an honest limitations section (linear injection vs digital twin; latency; scalability) are genuine strengths. The claim of zero executed invalid actions is operationally meaningful within the defined rejection loop. However, current significance is bounded by (i) an unvalidated first-order linear world model and (ii) a small, uncontrolled 15-trial comparison whose breach edge over LSTM is one trial. The work is a promising systems contribution that needs stronger empirical grounding before it can support the stronger “physics-grounded autonomous defense” claim.

major comments (4)
  1. [§IV-C Eq. (1), §IV-D Eq. (2), §VII-A] §IV-C Eq. (1) and §IV-D Eq. (2): The central “physics-grounded” and “Sentinel” claims rest on ranking actions by max of a linearly adjusted TimesFM forecast, ŷ'_t = ŷ_t + μ·min(t+1, δ). This is a closed-form offset on the forecast, not a dynamical plant model. The paper itself states (§VII-A) that it omits inflow coupling, hydraulic transients, and sensor lag. Consequently “breach prevented” and “hallucination rejected” are defined relative to this proxy, not relative to verified plant trajectories under the same actuator command. Without at least a digital-twin or residual-error study showing that argmin max F(Φ) preserves ranking of real outcomes, the safety and superiority claims are not yet load-bearing.
  2. [Tables II–III, §V-B, §VI] Tables II–III: Initial peaks are systematically unmatched across models (e.g., Trial 1 Sudden Spike: Neuro-Agentic 1183.07 vs LSTM 1059.40 vs TCN 1173.60). Breach prevention and risk-reduction comparisons therefore mix different threat trajectories with different controllers. A fair comparison requires identical injected attack windows (or identical base forecasts) for all three systems, or an ablation that holds the forecast fixed and varies only the planner/injection layer. The reported 5/15 vs 4/15 edge is not interpretable as a controlled superiority result under the current design.
  3. [§VI, Table IV, Abstract] §VI / Table IV: The headline result is 5 vs 4 vs 2 breaches on n=15 hand-cycled trials, with no confidence intervals, no significance tests, and no multiple-comparison control. Mean risk reduction for Neuro-Agentic has large variance (std 64.54). A one-trial difference over LSTM is too thin to support the abstract’s claim of better performance. Either expand the trial set with proper randomization and report uncertainty, or reframe claims as a case-study demonstration rather than a comparative superiority result.
  4. [§IV-D, Table III, Abstract] §IV-D and Table III “Hallucination Rejected” column: “Zero physically invalid (hallucinated) actions executed” is partly definitional. A hallucinated action is defined as positive magnitude, out-of-bounds, malformed, or R_sim ≥ R_base; the loop never executes such candidates by construction. That is a useful safety property of the architecture, but it should not be presented as independent empirical evidence that TimesFM correctly identified plant-unsafe actions. Separate (i) definitional rejection rate from (ii) any external check that rejected actions would indeed have been harmful on the plant.
minor comments (5)
  1. [Title, Abstract] Title and abstract use “Controlling Security Controls” and “physics-grounded” somewhat loosely; align terminology with what is actually implemented (linear forecast offset + feasibility filter).
  2. [Fig. 1, §V-B] Figure 1 is referenced as the system flowchart but the manuscript text does not fully specify how anomaly detection thresholds interact with the three attack-injection modes in §V-B; a short caption or algorithm cross-reference would help.
  3. [§II, §VII-A] Related work discusses TimeXer, DBLoss, medical hallucination benchmarks, and TS-Agent, but does not engage closely with existing CPS/ICS LLM-defense or digital-twin control literature beyond a brief L2M-AID mention in limitations; a tighter positioning would strengthen novelty claims.
  4. [Throughout, Table I] Typographical and consistency issues: “Neuro-Agentic Controlframework” missing space; mixed use of TimesFM / TimesFM2.5; arXiv date line “10 Jul 2026” looks like a placeholder; ensure Table I dropout/activation details match the experimental code description.
  5. [§VII] Latency (1.5–2.5 s) is discussed honestly in §VII; consider reporting per-component breakdown (LLM generation vs TimesFM counterfactuals) so readers can judge real-time fit for tank-level processes.

Circularity Check

1 steps flagged

"Zero hallucinated actions executed" is true by construction of the rejection rule that defines hallucination as any candidate failing feasibility or R_sim >= R_base.

specific steps
  1. self definitional [Abstract; §IV-D Hallucination Rejection Algorithm; Table III column "Hallucination Rejected" / "Hallucinated Action Executed"]
    "A hallucinated action is an LLM-proposed control vector that is physically invalid or unsafe in direction. This includes positive-magnitude (filling) under an overflow scenario, magnitude or duration outside actuator bounds, malformed JSON, or any vector whose counterfactual forecast exceeds the no-action baseline Rbase (i.e., it would actively worsen the state). ... No hallucinated action was ever executed ... with zero physically invalid (hallucinated) actions executed."

    The algorithm never actuates any candidate that meets the paper's own definition of "hallucinated." Consequently the claim that zero such actions were executed is true by construction of the filter, not an independent empirical observation that the LLM produced no unsafe proposals that slipped through. The result is definitionally guaranteed once the rejection rule is written.

full rationale

The paper's core empirical claims (5/15 breach prevention, mean risk reduction 48.51, superiority to LSTM/TCN) are obtained by running the closed loop on 15 injected-attack windows and measuring the same counterfactual peaks that the selection rule optimizes; TimesFM weights are frozen and never fit to the attack outcomes, so those numbers are not tautological. The linear injection Φ (Eq. 1) is an explicit modeling choice, not a fitted parameter renamed as a prediction, and the self-citation to the authors' prior TimesFM RMSE result [10] is only background justification for the sentinel, not a uniqueness theorem that forces the control result. The sole clear circularity is definitional: a "hallucinated" action is defined (IV-D) as any vector that is infeasible, positive-magnitude, or whose Φ-forecast exceeds the no-action baseline, and the loop is written to reject exactly those vectors before actuation; therefore the reported count of zero executed hallucinations cannot fail. That single definitional step raises the score from 0 to 3; nothing else in the derivation chain reduces by construction to its own inputs.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 2 invented entities

The central claim depends on treating a frozen TimesFM forecast plus a hand-designed linear drain as a safety oracle, on fixed actuator bounds and thresholds taken from the SWaT manual, and on a small set of synthetic attack amplitudes. No new physical constants are fit, but several design knobs (horizon, k, magnitude/duration ranges, critical level) and the linear Φ form are paper-chosen and load-bearing for the reported safety guarantees.

free parameters (5)
  • critical threshold θ / L_crit = ~1000 mm (CRITICAL); HIGH ALARM >800 mm
    Breach is defined relative to a fixed tank level (paper uses ~1000 mm CRITICAL from manual); all prevented-breach counts depend on this cutoff.
  • lookback W and forecast horizon H = W=720, H=24
    Window of 720 steps and horizon of 24 steps fix what TimesFM/LSTM/TCN see and optimize; not derived from first principles.
  • candidate set size k and actuator bounds = k=3; |μ|≤60; 5≤δ≤50; μ<0 for drain
    LLM asked for 3 strategies; magnitude in emergency/moderate/slow bands within [-60,-10] mm/timestep and duration in [5,50]; feasibility filter uses these hand limits.
  • synthetic attack amplitudes = +150/+100/+80 mm patterns cycled mod 3
    Sudden +150 mm / 20 steps, gradual +100 mm / 100 steps, high-noise +80 mm / 50 steps define the 15-trial test distribution.
  • linear injection form Φ (μ·min(t+1,δ)) = ŷ'_t = ŷ_t + μ·min(t+1, δ)
    Closed-form cumulative drain is a modeling choice, not identified from plant data; selection and 'physics grounding' rest on this functional form.
axioms (4)
  • ad hoc to paper A frozen pre-trained TimesFM forecast plus linear valve offset is an adequate world model for ranking control safety on SWaT tank dynamics without fine-tuning or a digital twin.
    Stated as Sentinel role and Eq. 1; limitations §VII-A admit missing nonlinear hydraulics.
  • domain assumption LLM-proposed JSON control vectors constrained by RAG manual text are a valid action space for closed-loop ICS defense when filtered by counterfactual risk.
    Architect stage §IV-B and prompt box; standard agentic-AI assumption for this paper's setting.
  • domain assumption Worst-case peak risk R = max_t forecast is the right scalar objective for selecting among drain strategies (Eq. 2).
    Hallucination rejection / optimization §IV-D; alternative objectives (integral risk, wear, multi-sensor) not used.
  • standard math Standard sequence-model training and evaluation splits (70/15/15) and causal dilated TCN / gated LSTM architectures are appropriate baselines for this control task.
    §V-A Table I; conventional DL practice.
invented entities (2)
  • Counterfactual Physics Injection Φ no independent evidence
    purpose: Translate LLM semantic drain commands into numerical forecast perturbations without modifying TimesFM weights, enabling pre-actuation simulation.
    Core named contribution §IV-C Eq. 1; independent_evidence false because validation is only under the same synthetic SWaT trials, not external plant identification.
  • Neuro-Agentic Loop (Architect–Sentinel) no independent evidence
    purpose: Separate semantic planning (Gemini) from deterministic numerical veto (TimesFM) for closed-loop ICS defense.
    Architectural framing Fig. 1 / Algorithm 1; pattern is compositional rather than a new physical entity, but is the paper's named system.

pith-pipeline@v1.1.0-grok45 · 17075 in / 4032 out tokens · 51358 ms · 2026-07-13T00:30:40.143485+00:00 · methodology

0 comments
read the original abstract

Cyberattacks on operational technology are increasingly causing costly downtime and physical damage, exposing the limitations of traditional rule-based monitoring in industrial IoT environments. While Large Language Models (LLMs) have strong semantic reasoning abilities to assist in decision support, their hallucinatory nature presents unacceptable safety liabilities for closed-loop control. This paper introduces a neuro-agentic control framework, a novel architecture that couples an LLM-based planner (i.e., such as Gemini 2.5 Flash-Lite) with a pre-trained Time-Series Foundation Model (TimesFM), to achieve physics-grounded autonomous defense. The paper introduces a ``Counterfactual Physics Injection'' mechanism that simulates the impact of LLM-proposed interventions within the numerical latent space of the foundation model before actuation, while allowing the system to reject hallucinatory or unsafe actions. Evaluated on an industrial dataset (e.g., the Secure Water Treatment (SWaT)) in the context of stochastic attack scenarios, the framework exhibited better performance compared to LSTM and TCN baselines. The Neuro-Agentic Loop prevented five breaches (33.3%) below the threshold versus LSTM (26.7%) and TCN (13.3%), with zero physically invalid (hallucinated) actions executed. These results demonstrate the efficacy of using foundation models as deterministic ``Sentinels'' to safeguard agentic AI in critical infrastructure.

Figures

Figures reproduced from arXiv: 2607.09076 by Akbar Siami Namin, Bipin Chhetri, Deepika Giri, Saroj Gopali, Sima Siami-Namini.

Figure 1
Figure 1. Figure 1: Neuro-Agentic System Flow Chart The proposed architecture Neuro-Agentic Loop Flow chart 1 operates as a two-stage control system. It integrates a semantic reasoner (i.e., The Architect) with a numerical physics engine (i.e., The Sentinel) to perform closed-loop controls on systems with multivariate time-series data. Algorithm 1 introduces the Neuro-Agentic Counterfactual Control Loop. It combines time￾seri… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

25 extracted references · 4 linked inside Pith

  1. [1]

    Structured agentic workflows for financial time-series modelling with llms and reflective feedback

    Yihao Ang, Yifan Bao, Lei Jiang, Jiajie Tao, Anthony KH Tung, Lukasz Szpurch, and Hao Ni. Structured agentic workflows for financial time-series modelling with llms and reflective feedback. InProceedings of the 6th ACM International Conference on AI in Finance, pages 924–932, 2025

    2025

  2. [2]

    An empirical evaluation of generic convolutional and recurrent networks for sequence modeling.arXiv preprint arXiv:1803.01271, 2018

    Shaojie Bai. An empirical evaluation of generic convolutional and recurrent networks for sequence modeling.arXiv preprint arXiv:1803.01271, 2018

    Pith/arXiv arXiv 2018

  3. [3]

    Agentic ai for autonomous anomaly management in complex systems.arXiv preprint arXiv:2507.15676, 2025

    Reza Vatankhah Barenji and Sina Khoshgoftar. Agentic ai for autonomous anomaly management in complex systems.arXiv preprint arXiv:2507.15676, 2025

    Pith/arXiv arXiv 2025

  4. [4]

    Zero-shot llm-guided counterfactual generation: A case study on nlp model evaluation

    Amrita Bhattacharjee, Raha Moraffah, Joshua Garland, and Huan Liu. Zero-shot llm-guided counterfactual generation: A case study on nlp model evaluation. In2024 IEEE International Conference on Big Data (BigData), pages 1243–1248. IEEE, 2024

    2024

  5. [5]

    Hallucination detection in foundation models for decision-making: A flexible definition and review of the state of the art.ACM Computing Surveys, 57(7):1–35, 2025

    Neeloy Chakraborty, Melkior Ornik, and Katherine Driggs-Campbell. Hallucination detection in foundation models for decision-making: A flexible definition and review of the state of the art.ACM Computing Surveys, 57(7):1–35, 2025

    2025

  6. [6]

    A decoder-only foundation model for time-series forecasting

    Abhimanyu Das, Weihao Kong, Rajat Sen, and Yichen Zhou. A decoder-only foundation model for time-series forecasting. InForty-first International Conference on Machine Learning, 2024

    2024

  7. [7]

    Gemini 2.5 updates: Flash/pro ga, sft, flash-lite on vertex ai, 2025

    Jason Gelman. Gemini 2.5 updates: Flash/pro ga, sft, flash-lite on vertex ai, 2025. Google Cloud Blog, accessed October 25, 2025

    2025

  8. [8]

    Gemini 2.5 flash-lite model card

    Google DeepMind. Gemini 2.5 flash-lite model card. https://storage.googleapis.com/deepmind-media/Model-Cards/Gemini-2-5-Flash-Lite-Model-Card. pdf, 2025. Accessed: 2026-01-08

    2025

  9. [9]

    A comparison of tcn and lstm models in detecting anomalies in time series data

    Saroj Gopali, Faranak Abri, Sima Siami-Namini, and Akbar Siami Namin. A comparison of tcn and lstm models in detecting anomalies in time series data. In2021 IEEE International Conference on Big Data (Big Data), pages 2415–2420, 2021

    2021

  10. [10]

    In-context and few-shots learning for forecasting time series data based on large language models.arXiv preprint arXiv:2512.07705, 2025

    Saroj Gopali, Bipin Chhetri, Deepika Giri, Sima Siami-Namini, and Akbar Siami Namin. In-context and few-shots learning for forecasting time series data based on large language models.arXiv preprint arXiv:2512.07705, 2025

    arXiv 2025

  11. [11]

    Long short-term memory.Neural computation, 9(8):1735–1780, 1997

    Sepp Hochreiter and J ¨urgen Schmidhuber. Long short-term memory.Neural computation, 9(8):1735–1780, 1997

    1997

  12. [12]

    Datasets

    iTrust. Datasets. https://itrust.sutd.edu.sg/itrust-labs .datasets/

  13. [13]

    H1 2023 – a Brief Overview of Main Incidents in Industrial Cybersecurity

    Kaspersky ICS CERT. H1 2023 – a Brief Overview of Main Incidents in Industrial Cybersecurity. Technical report, Kaspersky Lab, 2023. Industrial incident report describing major ICS cyber incidents and associated operational impacts, including downtime and production disruption

    2023

  14. [14]

    Jonathan Marshell Kevin and Naufan Raharya. Anomaly detection in imbalance secure water treatment dataset using lstm-dc-wasserstein generative adversarial network with gradient penalty.International Journal of Electrical, Computer, and Biomedical Engineering, 2(3):284–301, 2024

    2024

  15. [15]

    Medical hallucinations in foundation models and their impact on healthcare.arXiv preprint arXiv:2503.05777, 2025

    Yubin Kim, Hyewon Jeong, Shan Chen, Shuyue Stella Li, Chanwoo Park, Mingyu Lu, Kumail Alhamoud, Jimin Mun, Cristina Grau, Minseok Jung, et al. Medical hallucinations in foundation models and their impact on healthcare.arXiv preprint arXiv:2503.05777, 2025

    arXiv 2025

  16. [16]

    Foundation models for time series analysis: A tutorial and survey

    Yuxuan Liang, Haomin Wen, Yuqi Nie, Yushan Jiang, Ming Jin, Dongjin Song, Shirui Pan, and Qingsong Wen. Foundation models for time series analysis: A tutorial and survey. InProceedings of the 30th ACM SIGKDD conference on knowledge discovery and data mining, pages 6555–6565, 2024

    2024

  17. [17]

    Calf: Aligning llms for time series forecasting via cross-modal fine-tuning

    Peiyuan Liu, Hang Guo, Tao Dai, Naiqi Li, Jigang Bao, Xudong Ren, Yong Jiang, and Shu-Tao Xia. Calf: Aligning llms for time series forecasting via cross-modal fine-tuning. InProceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 18915–18923, 2025

    2025

  18. [18]

    itransformer: Inverted transformers are effective for time series forecasting.arXiv preprint arXiv:2310.06625, 2023

    Yong Liu, Tengge Hu, Haoran Zhang, Haixu Wu, Shiyu Wang, Lintao Ma, and Mingsheng Long. itransformer: Inverted transformers are effective for time series forecasting.arXiv preprint arXiv:2310.06625, 2023

    Pith/arXiv arXiv 2023

  19. [19]

    Swat: A water treatment testbed for research and training on ics security

    Aditya P Mathur and Nils Ole Tippenhauer. Swat: A water treatment testbed for research and training on ics security. In2016 international workshop on cyber-physical systems for smart water networks (CySWater), pages 31–36. IEEE, 2016

    2016

  20. [20]

    A time series is worth 64words: Long-term forecasting with transformers.arXiv preprint arXiv:2211.14730, 2022

    Y Nie. A time series is worth 64words: Long-term forecasting with transformers.arXiv preprint arXiv:2211.14730, 2022

    Pith/arXiv arXiv 2022

  21. [21]

    Dbloss: Decomposition-based loss function for time series forecasting.arXiv preprint arXiv:2510.23672, 2025

    Xiangfei Qiu, Xingjian Wu, Hanyin Cheng, Xvyuan Liu, Chenjuan Guo, Jilin Hu, and Bin Yang. Dbloss: Decomposition-based loss function for time series forecasting.arXiv preprint arXiv:2510.23672, 2025

    arXiv 2025

  22. [22]

    The true cost of downtime

    SMC Electric. The true cost of downtime. https://www.smcelectric.com/company/industry-insights/industry-insights/2024/06/17/ the-true-cost-of-downtime/, June 2024. Accessed: 2025-12-26

    2024

  23. [23]

    Timexer: Empowering transformers for time series forecasting with exogenous variables.Advances in Neural Information Processing Systems, 37:469–498, 2024

    Yuxuan Wang, Haixu Wu, Jiaxiang Dong, Guo Qin, Haoran Zhang, Yong Liu, Yunzhong Qiu, Jianmin Wang, and Mingsheng Long. Timexer: Empowering transformers for time series forecasting with exogenous variables.Advances in Neural Information Processing Systems, 37:469–498, 2024

    2024

  24. [24]

    Tianxiang Xu, Zhichao Wen, Xinyu Zhao, Jun Wang, Yan Li, and Chang Liu. L2m-aid: Autonomous cyber-physical defense by fusing semantic reasoning of large language models with multi-agent reinforcement learning (preprint).arXiv preprint arXiv:2510.07363, 2025

    arXiv 2025

  25. [25]

    One fits all: Power general time series analysis by pretrained lm.Advances in neural information processing systems, 36:43322–43355, 2023

    Tian Zhou, Peisong Niu, Liang Sun, Rong Jin, et al. One fits all: Power general time series analysis by pretrained lm.Advances in neural information processing systems, 36:43322–43355, 2023

    2023