Pith. sign in

REVIEW 4 major objections 4 minor 16 references

InsightBuild: LLM-Powered Causal Reasoning in Smart Building Systems

T0 review · 4 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read InsightBuild combines Granger causality tests with a fine-tuned language model to turn sensor data into plain-language explanations of building energy anomalies, reporting top-1 accuracy of 84.7 percent on Google buildings and 80.0…

desk verdict A genuinely new pipeline that grounds LLM explanations in Granger-based causal ranking, but the evaluation's expert ground truth is underspecified and the baseline adaptations are unclear. read the letter →

arxiv 2507.08235 v1 pith:RMPFCIGV submitted 2025-07-11 cs.LG cs.AI

classification cs.LGcs.AI
keywords causalinferenceGrangercausalitylargelanguagemodelssmartbuildingsenergyanomalyexplanationtime-seriesLLMfine-tuning
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

InsightBuild aims to establish that explanations of building energy anomalies are trustworthy only when they are anchored by an explicit causal ranking, and that a fine-tuned language model can turn that ranking into readable, actionable text. The pipeline runs pairwise Granger causality tests on sensor and control data, prunes indirect edges, selects the top three causes, and passes that ranked set to a LLaMA 2 7B model fine-tuned on about 2,500 expert-written (causes, explanation) pairs. On held-out anomalies from the Google and Berkeley datasets, it reports top-1 explanation accuracy of 84.7 percent and 80.0 percent, outperforming DeepAR, Causal Transformer, and a vanilla LLM baseline, while earning higher satisfaction ratings from facility managers. If accurate, this gives building dashboards a concrete path from raw charts and generic alerts to causal, human-readable diagnoses.

What carries the argument

The load-bearing object is the ranked causal set $\mathrm{Pa}(z)$ of top-3 parent variables of total energy, produced by pairwise Granger causality tests with $p=3$ lags, a structural pruning rule that deletes a direct edge unless its F-statistic exceeds the indirect path's by a factor of 1.5, and ranking by F-statistic. This ranked set, not the raw time series, is the only input the LLM sees, so the generator cannot invent drivers outside it. The fine-tuned LLaMA 2 7B model, trained on roughly 2,500 aligned (cause-set, explanation) pairs, supplies the natural-language layer that turns the ranked causes into a concise paragraph with a suggested action.

What would settle it

If the pipeline is tested on faults with physically known causes—for example, deliberately closing a damper or changing an occupancy schedule and measuring the resulting energy response—its top-1 accuracy should stay near the reported 80 to 85 percent; if it drops toward random agreement on such seeded faults, the expert labels were measuring consistency with the same deviation heuristic the causal module uses.

Watch

Extended reading notes

Core claim

The central claim is that combining explicit statistical causal discovery with LLM-based natural language generation produces explanations that both match expert judgments and help facility managers act. The paper demonstrates this with a two-stage system: a causal module applies Granger tests with three lags, prunes spurious edges using a structural causal model criterion, and ranks the top three parent variables of total energy by F-statistic; a fine-tuned LLaMA 2 7B then converts that cause set into a short paragraph naming the driver and suggesting a remedy. Against expert-annotated ground-truth causes, the full system reaches 84.7 percent top-1 accuracy on Google and 80.0 percent on Berkeley, with mean expert satisfaction of 4.2 and 4.0 on a five-point scale, while ablations show that removing causal ranking drops accuracy to 65.3 percent and removing language generation drops satisfaction to 2.1.

Load-bearing premise

The load-bearing premise is that the expert-annotated causes recorded for each test anomaly are the true causes, and that Granger causality with three lags plus the 1.5x pruning rule recovers those causes from the sensor windows.

Editorial extensions

If this is right

  • Building dashboards could move from generic alerts to text such as 'occupancy rose in Zone 3, so the chiller ran harder,' including a suggested mitigation.
  • A facility manager's first response can be guided by the reported top cause, which matched expert labels in 84.7 percent of Google and 80.0 percent of Berkeley test anomalies.
  • Both stages are necessary: causal ranking without language generation yields low satisfaction (2.1/5), while language generation without causal ranking yields 65.3 percent accuracy.
  • The same two-stage design is intended to transfer to other cyber-physical systems, such as data centers and manufacturing plants, where sensor streams and control actions are plentiful.

Reading between the lines

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

  • Editorial inference: because the evaluation labels were curated from maintenance logs and operational records, the reported accuracy may partly reflect how well the Granger heuristic matches the labeling heuristic; a cleaner test would inject faults with known physical causes and check whether the pipeline recovers them.
  • Editorial inference: the 1.5x pruning threshold and three-lag choice are empirical knobs; the paper reports no sensitivity analysis, so the method's behavior under lag misspecification or denser causal graphs remains untested.
  • Editorial inference: a practical extension would run the causal module online and regenerate the LLM explanation only when the top cause set changes, reducing cost while keeping the dashboard current.
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 InsightBuild, a two-stage system that first applies Granger causality tests with structural pruning to a sliding window of building telemetry to rank candidate causes of an energy anomaly, then fine-tunes LLaMA 2 7B on approximately 2,500 annotated (cause set, explanation) pairs to produce natural-language explanations. Evaluations on the Google Smart Buildings and Berkeley Office datasets report top-1 explanation accuracy of 84.7% and 80.0%, outperforming DeepAR, Causal Transformer, and a vanilla LLM baseline, together with higher expert satisfaction. The central claim is that combining explicit causal discovery with LLM-based natural language generation yields clear, precise explanations that assist facility managers in diagnosing and mitigating energy inefficiencies.

Significance. If the accuracy claims held, the work would be a useful application of LLM-based explanation generation to building energy management, with a concrete and re-implementable pipeline. The paper deserves credit for specifying fine-tuning details, providing an ablation that separates the causal-ranking component from the LLM component, and evaluating on two real-world datasets. However, the evaluation foundation is currently too weak to establish the claimed causal accuracy: the expert annotation protocol is unspecified, the causal module's validity is not demonstrated, and the baseline comparisons are not defined at the level of output types. The contribution is promising but not yet substantiated.

major comments (4)
  1. [Section 5.1, Table 1] The expert-annotated ground-truth causes are the sole yardstick for the headline Acc@1 numbers, but the paper does not specify who the annotators were, what instructions or evidence they received, how many anomalies were multiply annotated, or what inter-annotator agreement was. Because the causal module ranks variables by Granger F-statistics and the LLM input encodes increase/decrease relative to a moving mean, and because the illustrative example in Section 4 selects the variable with the largest deviation, the reported accuracy may reflect agreement with the same deviation-based heuristic that produced the labels rather than independent causal identification. The paper must provide a blinded annotation protocol, show that the approximately 2,500 fine-tuning examples are disjoint from the 200 plus 60 test anomalies, and report inter-annotator agreement such as Cohen's kappa.
  2. [Section 3.2, Eqs. (1)-(2)] Granger causality with p=3 lags and the 1.5x F-ratio pruning rule are introduced without justification or sensitivity analysis. The three lags are said to be chosen by BIC, but no BIC values are reported, and the 1.5 threshold is presented as an axiom rather than a tuned or validated choice. Since the entire explanation pipeline consumes only the top-3 ranked causes produced by this module, the causal module's false-positive rate directly bounds the achievable explanation accuracy. At minimum, the authors need a synthetic time-series experiment with known ground-truth causes, or a sensitivity analysis over p, w, k, and the pruning threshold, to show that the reported results are not artifacts of these free parameters.
  3. [Table 1, Section 5.3] DeepAR and Causal Transformer are forecasting and attention models; the paper does not describe how their outputs are converted into ranked lists of cause variables for comparison on Acc@1, P@3, and R@3. Without this adaptation, the baseline numbers are not well-defined. In addition, no confidence intervals or significance tests are reported for any difference, despite test sets of only 200 and 60 anomalies; a 5-10 percentage point gap may not be statistically meaningful at these sample sizes.
  4. [Section 5.2, Expert Satisfaction] The satisfaction score is reported to one decimal place, but the paper gives no information about the number of facility managers who rated the explanations, their experience, the rating instrument, or inter-rater reliability. A mean Likert score without this context is not interpretable as evidence that InsightBuild 'assists facility managers' as claimed in the abstract.
minor comments (4)
  1. [Table 1, Section 5.1] RBE is used in Table 1 but never expanded; it should be defined at first use in Section 5.
  2. [Section 3, Figure 2] Figure 2 is referenced in Section 3 but no figure content appears in the manuscript text; either include the figure or remove the reference.
  3. [Section 3.1 and 3.2] The z-score threshold of |z|>3 and the missing-data cutoff of >20% are stated without justification or citation; a brief rationale or reference would improve reproducibility.
  4. [References] The Google Smart Buildings dataset URL and the Berkeley Dryad repository identifier are missing from the references; include them so the data can be located.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the causal module is computed from telemetry statistics rather than fitted to labels, the held-out evaluation is standard, and the self-citations are background only.

full rationale

The derivation chain is self-contained. InsightBuild's causal module estimates Granger F-statistics from sensor telemetry (Section 3.2) and ranks top-k parents of energy; these F-statistics are computed from the time series, not fitted to expert labels. The fine-tuned LLM (Section 3.3) maps a cause set to text; its training pairs are annotated examples, and the test anomalies are explicitly held out (10% Google, 20% Berkeley) in Section 5.1. The headline Acc@1 for InsightBuild is the causal module's top cause against expert labels, which is a standard evaluation of a discovery method against an external ground truth. The authors' self-citations (Mohammadshirazi et al., 2022–2024; Neogi, 2019; Guha Neogi et al., 2020; etc.) are background references on sensors and prior ML work and do not carry the central argument; no uniqueness theorem or ansatz is imported. The unspecified annotation protocol and the possibility that expert labels may encode deviation-based heuristics are validity concerns, but the paper provides no equations or construction showing that labels are generated by the same deviation rule as the causal module. Without such a reduction, the concern is speculative and does not constitute circularity.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The central claim rests on several hand-set hyperparameters (lag order, window length, k, pruning factor, z-score threshold) and on domain assumptions that Granger-style tests and expert labels capture true causation. None of these are validated with sensitivity analysis, and the evaluation may share annotation heuristics with the fine-tuning data.

free parameters (6)
  • Granger lag order p = 3
    Chosen empirically via BIC (Section 3.2). All causal edges depend on this choice; no sensitivity analysis is reported.
  • Sliding window length w = 24h (Google), 6h (Berkeley)
    Set by hand in Section 3.2. The anomaly explanation depends on the window containing the relevant pre-history.
  • Top-k cause list length k = 3
    Selects the number of causes passed to the LLM (Section 3.2). Fixed with no ablation over k.
  • Structural pruning F-ratio threshold = 1.5
    Edges are removed unless the direct F-statistic is 1.5x the indirect path (Section 3.2, Step 2). The threshold is ad hoc and untested.
  • Anomaly z-score threshold = 3
    Defines which events are anomalies and therefore which explanations are evaluated (Section 3.1). Manual threshold with no validation.
  • Missing-data exclusion cutoff = 20%
    Variables with more than 20% missing values are dropped (Section 3.1). This may bias which sensors are available for causal discovery.
assumptions (4)
  • domain assumption Granger causality with lag p=3 approximates true causal structure among building telemetry
    The causal module's output is called causal, but Granger tests establish predictive precedence, not structural causation. This is the central interpretive leap in Section 3.2.
  • ad hoc to paper The 1.5x F-ratio pruning rule removes spurious indirect edges without removing genuine direct causes
    The pruning criterion is introduced for this paper (Section 3.2, Step 2) and has no theoretical guarantee or empirical calibration.
  • domain assumption Expert-annotated causes are correct and independent of the model outputs
    Ground truth for accuracy and satisfaction (Section 5.1). Protocol, inter-annotator agreement, and independence from the fine-tuning corpus are not reported.
  • domain assumption Z-score anomalies correspond to events whose causes can be reliably annotated
    The anomaly definition is a threshold rule (Section 3.1); no verification that these events have identifiable, annotatable causes.

how reviews work

0 comments
Cite this review

Pith. "Pith review of InsightBuild: LLM-Powered Causal Reasoning in Smart Building Systems." pith.science (2026). https://pith.science/paper/RMPFCIGV

@misc{pith2026250708235,
  author       = {Pith},
  title        = {Pith review of: InsightBuild: LLM-Powered Causal Reasoning in Smart Building Systems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RMPFCIGV}},
  note         = {Machine review of arXiv:2507.08235}
}
read the original abstract

Smart buildings generate vast streams of sensor and control data, but facility managers often lack clear explanations for anomalous energy usage. We propose InsightBuild, a two-stage framework that integrates causality analysis with a fine-tuned large language model (LLM) to provide human-readable, causal explanations of energy consumption patterns. First, a lightweight causal inference module applies Granger causality tests and structural causal discovery on building telemetry (e.g., temperature, HVAC settings, occupancy) drawn from Google Smart Buildings and Berkeley Office datasets. Next, an LLM, fine-tuned on aligned pairs of sensor-level causes and textual explanations, receives as input the detected causal relations and generates concise, actionable explanations. We evaluate InsightBuild on two real-world datasets (Google: 2017-2022; Berkeley: 2018-2020), using expert-annotated ground-truth causes for a held-out set of anomalies. Our results demonstrate that combining explicit causal discovery with LLM-based natural language generation yields clear, precise explanations that assist facility managers in diagnosing and mitigating energy inefficiencies.

Figures

Figures reproduced from arXiv: 2507.08235 by the authors.

Figure 1
Figure 1. Illustrative example in Google Building B: raw total energy (top), detected anomaly at 14:00 (red), occupancy in Zone 3 (middle), and zone temperature in Zone 3 (bottom). InsightBuild correctly identifies occupancy↑ as primary driver. This enforces sparsity and approximate faithfulness (adapted from Hyvärinen & Smith, 2020). 3. Cause Ranking. For a given target z (e.g., total en￾ergy), we collect all immediate paren… view at source ↗
Figure 2
Figure 2. Overview of InsightBuild framework. Time-series data is first preprocessed and fed into a causal discovery module. Detected causal relations for a target anomaly are passed to a fine-tuned LLM, which generates a human-readable explanation. increased or decreased relative to its moving mean in the window. Target Outputs. Each target is a short paragraph (2–3 sen￾tences) that states the causal chain and suggests corre… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

16 extracted references · 10 canonical work pages

  1. [1]

    B., Mann, B., Ryder, N., et al

    Brown, T. B., Mann, B., Ryder, N., et al. Language models are few-shot learners. InAdvances in Neural Information Processing Systems (NeurIPS), pp. 1877–1901,

  2. [9]

    Novel Physics-Based Machine-Learning Models for Indoor Air Quality Approximations

    Mohammadshirazi, A., Nadafian, A., Monsefi, A. K., Rafiei, M. H., and Ramnath, R. Novel physics-based machine- learning models for indoor air quality approximations. arXiv preprint arXiv:2308.01438,

  3. [10]

    PIAD-SRNN: Physics-Informed Adaptive Decomposition in State-Space RNN

    Mohammadshirazi, A., Nosratifiroozsalari, A., and Ram- nath, R. Dssrnn: Decomposition-enhanced state-space recurrent neural network for time-series analysis.arXiv preprint arXiv:2412.00994,

  4. [12]

    doi: 10.1109/ ICCCI.2019.8821993. OpenAI. Gpt-4 technical report. Technical report,

  5. [13]

    Salinas, D., Flunkert, V ., Gasthaus, J., and Tim, J

    https://arxiv.org/abs/2303.08774. Salinas, D., Flunkert, V ., Gasthaus, J., and Tim, J. Deepar: Probabilistic forecasting with autoregressive recurrent networks.International Journal of Forecasting, 36(3): 1181–1191,

  6. [14]

    Llama 2: Open foundation and fine-tuned chat models.arXiv preprint arXiv:2307.09288,

    Touvron, H., Martin, L., Stone, K., et al. Llama 2: Open foundation and fine-tuned chat models.arXiv preprint arXiv:2307.09288,

  7. [15]

    Caiformer: A causal informed transformer for multivariate time series forecasting.arXiv preprint arXiv:2505.16308,

    Zhang, X., Qiang, W., Zhao, S., Guo, H., Li, J., Sun, C., and Zheng, C. Caiformer: A causal informed transformer for multivariate time series forecasting.arXiv preprint arXiv:2505.16308,

  8. [16]

    doi: https://doi.org/10.1016/j.ifacol.2024.08.317

    ISSN 2405-8963. doi: https://doi.org/10.1016/j.ifacol.2024.08.317. URL https://www.sciencedirect.com/ science/article/pii/S2405896324010607. 12th IFAC Symposium on Advanced Control of Chemical Processes ADCHEM

Show all 16 references
  1. [2000]

    Liu, H., Li, C., Wu, Q., and Lee, Y

    Morgan Kaufmann. Liu, H., Li, C., Wu, Q., and Lee, Y . J. Visual instruction tuning.arXiv preprint arXiv:2304.08485,

  2. [2019]

    doi: 10.3390/ make1010019

    ISSN 2504-4990. doi: 10.3390/ make1010019. URL https://www.mdpi.com/ 2504-4990/1/1/19. Naveed, H., Khan, A. U., Qiu, S., Saqib, M., Anwar, S., Usman, M., Akhtar, N., Barnes, N., and Mian, A. A comprehensive overview of large language models.arXiv preprint arXiv:2307.06435,

  3. [2020]

    Karbasforoushha, M

    doi: 10.1007/s10489-019-01545-x. Karbasforoushha, M. A., Khajehzadeh, M., Jearsiripongkul, T., Keawsawasvong, S., and Eslami, M. A com- prehensive review of building energy optimization 5 InsightBuild using metaheuristic algorithms.Journal of Building Engineering, 98:111377,

  4. [2021]

    ISBN 978- 981-15-5616-6

    Springer Singapore. ISBN 978- 981-15-5616-6. Guha Neogi, P. P., Goswami, S., and Mustafi, J. Intelli- gent water drops-based image steganography. In Mandal, J. K. and Mukhopadhyay, S. (eds.),Proceedings of the Global AI Congress 2019, pp. 363–375, Singapore,

  5. [2022]

    Crafting papers on machine learning

    Langley, P. Crafting papers on machine learning. In Langley, P. (ed.),Proceedings of the 17th International Conference on Machine Learning (ICML 2000), pp. 1207–1216, Stan- ford, CA,

  6. [2023]

    and Torlone, R

    Matarazzo, A. and Torlone, R. A survey on large language models with some insights on their capabilities and limi- tations.arXiv preprint arXiv:2501.04040,

  7. [2024]

    doi: https://doi.org/10.1016/j.jobe.2024.111377

    ISSN 2352-7102. doi: https://doi.org/10.1016/j.jobe.2024.111377. URL https://www.sciencedirect.com/ science/article/pii/S2352710224029450. Kojima, T., Gu, S. S., Reid, M., Matsuo, Y ., and Iwasawa, Y . Large language models are zero-shot reasoners.arXiv preprint arXiv:2205.11916,

  8. [2025]

    Accessed: 2025-05-31

    URLhttps://ai.meta.com/ blog/llama-4-multimodal-intelligence/ . Accessed: 2025-05-31. Moens, P., V olckaert, B., and Van Hoecke, S. Viscars: Knowledge graph-based context-aware recommender sys- tem for time-series data visualization and monitoring dashboards.IEEE Transactions ...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.