{"id":"da512c4d-f677-4d79-9470-dc9618e0c52a","arxiv_id":"2607.04374","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A modular pipeline combining MLP prediction, LIME/SHAP explanations, fidelity/stability metrics, and LLM natural-language summaries is demonstrated on power-system fault and building-energy datasets.","lead":"The paper assembles a pipeline that trains neural nets on energy data, runs LIME and SHAP, scores explanation fidelity/stability, and feeds structured outputs to an LLM for plain-language summaries. It is a practical packaging of known tools for power-system fault detection and building energy prediction rather than a new theory of explainability.","discovery_kind":"new_application","skeptic_critique":{"model":"grok-4.5","headline":"LLM natural-language layer is claimed to improve interpretability for non-technical stakeholders without any human evaluation of comprehension, trust, or factual accuracy.","rationale":"The reader correctly isolates the human-evaluation gap as the weakest assumption supporting the strongest claim. Predictive performance and fidelity/stability numbers are reported and a random-forest baseline is included, so the technical XAI pipeline is credible; the novelty is modest and the LLM layer is the only component that elevates the work beyond a standard LIME+SHAP engineering assembly. Because that layer is never validated with humans, the claim that the framework produces explanations ‘which can be easily understood by the non-technical stakeholders’ is unsupported. The concrete user study above would settle the issue cleanly. No stronger internal inconsistency or methodological flaw is present; therefore the reader’s CONDITIONAL verdict (usable once the human-evaluation gap and missing artifacts are addressed) remains appropriate and needs no change.","tokens_in":15381,"tokens_out":627,"duration_ms":7193,"concrete_test":"Recruit 20–30 non-technical participants; give half only the structured SHAP/LIME numbers + fidelity/stability scores and half the LLM five-bullet text for the same 10 held-out instances (5 fault, 5 energy). Measure (a) accuracy on 3 multiple-choice questions about which features drove the prediction and in which direction, and (b) 5-point trust/comprehension Likert scores. If the LLM arm does not improve accuracy by ≥15 % or trust by ≥0.5 points, the central ‘improves interpretability for everyone’ claim fails.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The strongest claim rests on two parts: (1) LIME/SHAP explanations are stable and faithful (supported by reported fidelity ≈ 1.0 / stability 0.998 on fault detection and 0.998 / 0.719 on energy regression), and (2) the LLM conversion of those structured scores into five-bullet natural-language text improves interpretability for non-technical stakeholders. Part (2) is load-bearing for the paper’s title, abstract, and contributions yet is never measured. Section 2.3 feeds only model name, prediction, ranked SHAP/LIME coefficients, direction of influence, and fidelity/stability scores into a fixed prompt; no raw plots, no domain context beyond a short string, and no human study. The paper never reports comprehension accuracy, trust ratings, factual faithfulness of the generated bullets, or comparison against raw SHAP/LIME plots. Without that evidence the claim that the pipeline ‘improves interpretability … to everyone with the help of LLMs’ remains an untested assertion, even though the XAI metrics themselves are solid.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The manuscript proposes a reusable end-to-end XAI pipeline for energy systems that (i) trains multilayer perceptrons for classification and regression, (ii) produces local and global post-hoc explanations with LIME and KernelSHAP, (iii) scores those explanations with fidelity and stability metrics, and (iv) feeds structured feature attributions into a fixed LLM prompt to generate five-bullet natural-language explanations for non-technical stakeholders. The pipeline is demonstrated on a power-system fault-detection dataset (≈99% accuracy, ROC-AUC 1.00) and a building energy-labels regression dataset (test R² ≈ 0.67), with a random-forest baseline for comparison. The authors conclude that the framework yields faithful, stable explanations and that the LLM layer improves interpretability of black-box models for everyone.","tokens_in":15631,"tokens_out":1493,"duration_ms":23047,"significance":"If the full claim holds, the work would offer a practical, model-agnostic template for deploying explainable neural networks in safety-critical and sustainability-oriented energy applications, where both quantitative explanation quality and stakeholder-facing language matter. Strengths include dual-task validation (classification and regression), explicit fidelity/stability reporting, and a side-by-side neural-network vs. random-forest comparison (Tables 2–3). The distinctive contribution relative to prior energy-XAI work is the automated LLM natural-language layer; that contribution, however, currently rests on an untested assumption about human comprehension and trust, so the paper’s significance is conditional on either adding human evaluation or substantially narrowing the claim.","major_comments":[{"comment":"The title, abstract, listed contributions, and §2.3 assert that the LLM converts LIME/SHAP outputs into natural-language explanations that non-technical stakeholders can understand and that thereby improve interpretability “to everyone.” Section 2.3 only specifies a fixed five-bullet prompt fed with model name, prediction, ranked feature scores, influence direction, and fidelity/stability numbers; no human study of comprehension, trust, decision usefulness, or factual faithfulness of the generated text is reported, nor is any comparison against raw SHAP/LIME plots. This is load-bearing for the paper’s central claim and must be addressed either by a user study (or at least expert rating of faithfulness/clarity) or by reframing the contribution as automated NL generation without claiming improved stakeholder interpretability.","section":null},{"comment":"§2.1 is internally inconsistent on the activation function. The prose states that both the classifier and regressor use ReLU, yet Eq. (2) is written as a tanh activation and the surrounding text then re-describes ReLU. The feed-forward equations, gradient expressions (Eqs. 4–5), and the claimed architecture must be made consistent; as written, a reader cannot reproduce the trained models.","section":null},{"comment":"§2.2 / Eq. (9): SHAP fidelity is reported as 1.00 (fault detection) and 0.998 (energy). Exact Shapley values satisfy local accuracy by construction (∑φᵢ + φ₀ = f(x)), so a fidelity of 1.0 is expected for exact SHAP and is only informative for the KernelSHAP approximation residual. The manuscript never states the numerical definition used in Eq. (9) (the equation body is incomplete in the text) nor how many background samples / coalitions were used. Without that definition and without reporting the approximation residual distribution, the “nearly perfect fidelity” claim is difficult to interpret and may be partly tautological.","section":null},{"comment":"Abstract and §5 claim the approach produces “stable” explanations, yet §4.3.4 reports SHAP stability of only 0.719 on the energy regression task (vs. 0.998 on fault detection; RF reaches 0.982). The paper correctly notes sensitivity to building-type features, but the unqualified stability language in the abstract and conclusion should be qualified, and the authors should discuss whether explanations with stability ≈0.72 are reliable enough for the intended urban-planning use case.","section":null},{"comment":"Reproducibility of the LLM layer is incomplete. §2.3 never names the LLM (model family, version, temperature, or whether a local vs. API model was used), nor does it release the exact structured context strings or example outputs beyond a high-level description. Given that the LLM step is a listed contribution, the model identity and generation settings are free parameters that must be fixed for the pipeline to be reusable.","section":null}],"minor_comments":[{"comment":"Throughout: numerous grammatical and typographical issues (“block-box”, “aa well as”, “ans 0.8040”, “ProprtyGFATotal”, repeated sentence in §2.4, “Fig 4” referenced without a corresponding figure in the provided text). A careful copy-edit is needed.","section":null},{"comment":"§2.1: hidden-layer widths (192 / 448) and iteration caps appear without ablation or selection rationale; a short sensitivity note would help readers judge robustness of the free parameters listed in the architecture.","section":null},{"comment":"Figures 5–14 are described but several captions and in-text references are incomplete or inconsistent (e.g., “Fig 4: LLM Integration” appears after the methodology without a visible figure body in the manuscript text). Ensure every figure is present, labeled, and referenced once.","section":null},{"comment":"Table 1 and §3: building-type class imbalance (Educational 4%, Institutional 2.3%) is noted but not discussed as a possible driver of the low stability or of SHAP dominance of Residential/Commercial indicators; a brief remark would strengthen the regression analysis.","section":null},{"comment":"References: several energy-XAI and LLM-for-XAI works are cited, but TalkToModel and related interactive NL explanation systems are mentioned only briefly; a clearer positioning of what is new relative to Slack et al. (2023) would help.","section":null},{"comment":"§4.1.3 waterfall discussion of the misclassified instance is hard to follow (“contributions seemed to point to the fault class the model made an incorrect prediction”); clarify predicted probability vs. true label for both instances.","section":null}],"recommendation":"major_revision","confidential_remarks":"The engineering integration is competent and the dual energy-system case studies are appropriate for eess.SY, but novelty is largely compositional (standard MLP + LIME/SHAP + off-the-shelf LLM prompt). The decisive gap is the untested stakeholder-interpretability claim that drives the title. I would accept a revised version that either adds a small human evaluation or renames/reframes the contribution around automated NL generation and quantitative XAI metrics alone. No concerns about misconduct; the manuscript simply overclaims relative to the evidence provided."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"This is a clean engineering assembly, not a new algorithm. They wire MLP training, KernelSHAP, LIME, fidelity/stability scores, and a fixed five-bullet LLM prompt into one reusable pipeline and run it on power-system fault detection (99% accuracy, ROC-AUC 1.0) and building energy regression (R^{2} 0.67). That is exactly what the contribution list says, and they deliver it with clear figures, RF baselines, and quantitative explanation metrics.\n\nWhat works: the XAI numbers are reported honestly. Fault-detection SHAP fidelity is 1.0 and stability 0.998; energy regression fidelity stays high (0.998) while stability drops to 0.719, which they correctly attribute to categorical building-type sensitivity. The RF comparisons are useful—they show the NN is competitive but not always superior, and that explanation stability can differ by model class. The math is standard (Shapley, LIME local R^{2}, L2-regularized MLP) and the citation trail is appropriate (Ribeiro, Lundberg, TalkToModel, energy XAI surveys). No circularity; metrics are computed post-hoc on held-out data.\n\nThe soft spot is real but proportional: the title, abstract, and contribution list claim the LLM layer “improves interpretability for everyone / non-technical stakeholders.” Section 2.3 only feeds structured scores into a prompt template; there is no human study of comprehension, trust, or factual accuracy of the generated bullets. That claim is therefore an assertion, not a result. Missing code and unspecified LLM also hurt reproducibility. These are fixable gaps, not load-bearing math failures.\n\nWho benefits: practitioners who want a documented, end-to-end recipe for energy tabular models and who already accept LIME/SHAP. It will not change XAI theory. I would send it to peer review at an applied energy or systems journal; referees can demand the human evaluation and artifacts. Worth a look if you are building similar tooling; not a must-read for the core XAI literature.","headline":"Solid engineering pipeline of LIME/SHAP + fidelity/stability + LLM text on two energy tasks; the human-interpretability claim is untested but the rest is usable.","tokens_in":16270,"tokens_out":531,"would_cite":false,"duration_ms":5469,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"An end-to-end pipeline turns neural-network predictions for energy systems into LIME/SHAP scores, fidelity and stability checks, then LLM natural-language explanations.","keywords":["Explainable AI","Large language model","LIME","SHAP","Neural networks","Power system fault detection","Building energy labels","Fidelity and stability"],"falsifier":"A controlled user study in which non-technical readers of the LLM bullets are tested on factual accuracy of the explanations and on measured trust or decision quality; if comprehension or trust scores stay near chance, the natural-language claim fails.","tokens_in":16242,"feed_emoji":"⚡","tokens_out":604,"duration_ms":5746,"temperature":0.7,"pith_summary":"Black-box neural networks can predict power-system faults and building energy use accurately, but their decisions are hard for non-technical people to trust. This paper claims a reusable pipeline solves that: it trains a multilayer perceptron on tabular energy data, generates local and global feature attributions with LIME and SHAP, scores those attributions for fidelity and stability, then feeds the structured numbers into a fixed LLM prompt that returns five short plain-language bullets. On a balanced fault-detection set the network reaches 99 percent accuracy and ROC-AUC 1.0 with near-perfect explanation scores; on building energy labels it reaches R² ≈ 0.67 with high fidelity but lower stability. The authors argue that packaging prediction, quantitative explanation quality, and automatic natural-language translation into one model-agnostic workflow makes high-stakes energy decisions transparent enough for engineers and non-experts alike.","feed_headline":"XAI pipeline turns energy neural nets into plain-language explanations","feed_subtitle":"LIME, SHAP, fidelity scores and a fixed LLM prompt make black-box fault and energy predictions readable by non-experts.","key_machinery":"The integrated pipeline itself: standardized preprocessing → MLP training → LIME/SHAP attribution → fidelity/stability evaluation → fixed five-bullet LLM prompt that turns numerical attributions into plain-language plot overview, key findings, feature importance, model behaviour and potential issues.","core_discovery":"A single end-to-end, model-agnostic XAI framework can train neural networks for both classification and regression energy tasks, produce LIME local and SHAP global explanations, quantify those explanations with fidelity and stability, and convert the resulting structured feature scores into concise natural-language summaries via a fixed LLM prompt, thereby rendering black-box decisions understandable to non-technical stakeholders while preserving high predictive performance.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["End-to-end XAI turns energy neural nets into LLM plain-language summaries","LIME and SHAP scores feed fixed LLM prompts for energy model explanations","Framework trains energy NNs then converts XAI outputs to readable text","Fidelity-stable LIME/SHAP pipeline yields natural-language energy insights","Model-agnostic XAI plus LLM makes energy black-box decisions clear"],"cache_read_input_tokens":128,"weakest_assumption_plain":"That a fixed five-bullet LLM prompt fed only structured feature scores and fidelity/stability numbers (never raw plots or human ratings) produces text that non-technical stakeholders will correctly understand and trust.","fun_headline_variants_meta":{"raw":{"variants":["End-to-end XAI turns energy neural nets into LLM plain-language summaries","LIME and SHAP scores feed fixed LLM prompts for energy model explanations","Framework trains energy NNs then converts XAI outputs to readable text","Fidelity-stable LIME/SHAP pipeline yields natural-language energy insights","Model-agnostic XAI plus LLM makes energy black-box decisions clear"]},"model":"grok-4.5","effort":"low","cost_usd":0.00535,"raw_usage":{"total_tokens":1495,"prompt_tokens":857,"num_sources_used":0,"completion_tokens":102,"cost_in_usd_ticks":53500000,"prompt_tokens_details":{"text_tokens":857,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":536,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":857,"tokens_out":102,"duration_ms":4767,"temperature":1.0,"reasoning_tokens":536,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-11T19:38:11.333371+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"A controlled user study in which non-technical readers of the LLM bullets are tested on factual accuracy of the explanations and on measured trust or decision quality; if comprehension or trust scores stay near chance, the natural-language claim fails.","supporting_citations":[],"review_version":1}