{"id":"27cc33b5-b169-4382-bfd7-4ef9ad8268c0","arxiv_id":"2607.16168","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Behaviour-conditioned Attentive Neural Processes with soft clustering labels reduce MAE by 7.9% and CRPS by 6.9% over ANP for residential short-term load forecasting.","lead":"An electricity-load forecasting model that guesses each household's routine from recent history and uses that guess to shape its predictions beat a plain neural-process baseline on 400 Australian smart-meter homes. The gains are modest and largest when little history is available, but the study lacks code, error bars, and a capacity-matched baseline.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The ANP ablation does not isolate behavioural conditioning: FiLM-ANP adds a label-inference network, FiLM/prompt parameters, and a supervised cluster-loss term, so the reported MAE/CRPS gains may be a capacity/auxiliary-task artifact rather than evidence for behaviour-conditioned forecasting.","rationale":"The paper's central contribution is that inferred behavioural structure, embedded via FiLM/prompt conditioning, improves both point accuracy and predictive uncertainty over a label-agnostic ANP. The strongest evidence for this is the ablation in Section 6.1. But the ANP baseline disables the entire label path and FiLM conditioning, so the ablation conflates behavioural information with extra model capacity and an auxiliary supervised loss. This is a more direct threat to the attribution than the reader's recoverability concern: even if the label inference network perfectly recovers the K-means clusters from partial context, the reported gains could still be due to the extra parameters/regularization rather than to the behavioural content of c. The reader does mention 'no capacity-matched ANP baseline' in the rationale, but the stated weakest assumption is about cluster recoverability; hence partial agreement. The proposed null-conditioned test would settle the attribution: if random-c FiLM-ANP matches the soft-label model, the central claim is unsupported; if it clearly underperforms, the behavioural conditioning account is substantially strengthened. A secondary concern is that K-means is fit on complete profiles before the user split, which weakens the user-disjoint guarantee, but the capacity confound is more load-bearing because it undermines the causal interpretation even under an otherwise clean evaluation. The verdict remains CONDITIONAL: the results are plausible and the concern is addressable, but the current evidence does not yet establish that behaviour-conditioning, rather than capacity, drives the gains.","tokens_in":22754,"tokens_out":6431,"duration_ms":64054,"concrete_test":"Train a null-conditioned FiLM-ANP with the same architecture, optimizer, and loss as FiLM-ANP-Soft, but replace each profile's behavioural soft label c with a fixed random/constant vector (or with randomly permuted labels) so that no true behavioural signal is available. Evaluate on the same SGSC user-disjoint splits and the same context/horizon grid. If this null-conditioned model reproduces most of FiLM-ANP-Soft's MAE/CRPS reductions over ANP, the reported gains are a capacity/auxiliary-loss artifact. Also report parameter counts for ANP, FiLM-ANP-Hard, FiLM-ANP-Soft, and the null-conditioned control to confirm the control matches capacity.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that inferred behavioural structure causes the forecast improvements. The ablation supporting this compares ANP with FiLM-ANP-Hard/Soft, but the manipulation changes more than one variable: (i) the decoder gains a label-embedding MLP, a HyperFiLM generator, and a prompt projection — additional parameters; (ii) the training objective adds a supervised cross-entropy term against K-means soft labels; (iii) the model receives a new conditioning input c. The paper states the variants 'differ only in how behavioural information is incorporated,' but the ANP baseline has no label-inference network, no conditioning branch, and no auxiliary cluster-prediction loss. It is therefore not capacity-matched and does not receive the same auxiliary supervision. Without a control that adds matched capacity and supervision while removing the behavioural signal (e.g., random or constant c), the reported 7.9% MAE and 6.9% CRPS improvements cannot be attributed specifically to behaviour-conditioning rather than to a larger, differently regularized model. This is an identification problem in the experimental design, not an internal inconsistency.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a behaviour-conditioned Attentive Neural Process (ANP) for residential short-term load forecasting. The model augments a standard ANP with a discrete behavioural latent variable c, inferred from the observed context by a label-inference network, and uses this variable to condition a Transformer decoder through HyperFiLM modulation and prompt-token injection. K-means soft labels computed from complete daily profiles provide weak supervision during training, while at test time only the context-inferred class distribution is used. Experiments on 400 SGSC households with user-disjoint splits, context lengths C in {48,...,95}, and horizons H=1,...,8 report average MAE/CRPS reductions of 7.9%/6.9% over a label-agnostic ANP baseline, and lower RMSE than fixed-window deterministic baselines at all horizons.","tokens_in":23089,"tokens_out":7306,"duration_ms":63227,"significance":"If the reported improvements are genuinely attributable to behavioural conditioning, this is a useful and timely contribution: it embeds inferred behavioural structure inside a single probabilistic forecasting model, rather than using clusters as an external grouping signal, and it supports uncertainty-aware, context-adaptive forecasting across heterogeneous households. The evaluation has good features: user-disjoint splits, variable context lengths and horizons, CRPS as a proper scoring rule, and explicit attention to avoiding future-label leakage at test time. However, the current experimental design does not isolate the behavioural signal from model capacity, auxiliary supervision, or input information, so the central claim is not yet established.","major_comments":[{"comment":"The ablation is not capacity-matched. The text says the variants 'share the same architecture... and differ only in how behavioural information is incorporated,' but the ANP baseline disables the label-inference network, the HyperFiLM/prompt conditioning branch, and the supervised cross-entropy term. The comparison therefore changes at least three variables: parameter count, auxiliary supervision, and the conditioning input. The 7.9%/6.9% gains could be a capacity/regularization artefact rather than evidence for behavioural conditioning. Please add controls such as (i) ANP with the same added parameters but using a constant or random conditioning vector, (ii) ANP with the auxiliary cluster-prediction loss but no FiLM conditioning, and (iii) test-time shuffling of the inferred class to verify that the behavioural signal itself matters.","section":"Section 6.1"},{"comment":"The comparison with deterministic baselines is confounded by input length. The NP models are evaluated over context lengths C in {48,...,95} and averaged, while the fixed-window baselines receive exactly 48 half-hour observations. The statement that 'all models get the same input information' is therefore incorrect: the average NP context is substantially longer than 48 steps. This alone could explain the RMSE advantage. Please report NP results restricted to C=48 for the baseline comparison, or train/evaluate the deterministic baselines on the same variable context-length distribution (e.g., a direct model with 95-step inputs), so the comparison reflects modelling paradigm rather than information amount.","section":"Section 6.3, Table 4"},{"comment":"There is a serious notation/semantics issue in the conditioning distribution. Eq. (21) defines pcond(x) = alpha pteach(x) + (1-alpha) qψ(c|DC), but qψ was defined as the target-conditioned posterior qψ(c|DT). If the decoder conditioning during training actually uses the target-conditioned posterior, then future target values leak into the conditioning signal and the reported gains may be an artefact of that leak, undermining the causal train/test consistency claim. If, as the text states, the context-conditioned prior pψ(c|DC) is intended, correct Eq. (21) and clarify that qψ(c|DT) is used only for the auxiliary loss in Eq. (16). This point is load-bearing and must be resolved.","section":"Section 5.2, Eq. (21)"},{"comment":"The behavioural classes are K-means clusters of the same z-scored load profiles, with a silhouette value of only 0.042 and no external semantic validation. Since the label-inference network is trained to reconstruct these clusters from the same context that the deterministic attention path already encodes, it is unclear whether the inferred c contributes information beyond simple context statistics (e.g., mean, variance, peak timing). Please include a control that conditions on random labels or on continuous context summaries with matched capacity. Without such a control, the improvements cannot be attributed specifically to recoverable behavioural structure rather than to the extra conditioning mechanism.","section":"Section 7, Table 2"}],"minor_comments":[{"comment":"Beyond the major issue, the notation qψ(c|DC) is undefined because qψ is a target-conditioned posterior; if the prior is intended, use pψ(c|DC) consistently throughout Section 5.2.","section":"Eq. (21)"},{"comment":"The profiles are described as 'two-day load profiles' but W=104 half-hour steps equals 52 hours. Please harmonize the terminology ('two-day plus four hours' or '52-hour profiles').","section":"Section 4"},{"comment":"All results are reported as single point estimates without error bars, confidence intervals, or significance tests. Given the small margins (e.g., FiLM-ANP-Soft vs. SVR at H=5 or vs. LSTM at H=2), multiple seeds or a paired test would materially strengthen the claims.","section":"Tables 3 and 4"},{"comment":"There is a typo: 'ANP–Soft consistently achieves the lowest values' should read 'FiLM-ANP–Soft'.","section":"Section 6.3"},{"comment":"References [1] and [47] appear to be the same Kong et al. paper; please consolidate.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The core idea is plausible and the domain motivation is clear, but the two headline comparisons (ANP ablation and deterministic baselines) are both confounded in ways that are fixable with additional experiments. I would not reject; the manuscript needs a revised evaluation with capacity-matched ablations and an information-matched comparison, plus clarification of the Eq. (21) conditioning signal."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read this one carefully before treating the numbers as evidence. The new element is real: a discrete behavioural latent, weakly supervised by K-means soft labels, that conditions the decoder via FiLM and prompt injection inside an ANP. That combination is novel for STLF, and the implementation is clean. The semi-supervised ELBO is correctly derived, the context-inference design prevents direct label leakage, and the user-disjoint split is the right evaluation for generalization. The gains over ANP are consistent across horizons and context lengths, and the soft-label variant beats the deterministic baselines on RMSE.\n\nThe soft spot is the ablation. The label-agnostic ANP baseline lacks the entire label-inference network, the FiLM/prompt branch, and the supervised cluster-loss term. So the 7.9% MAE / 6.9% CRPS improvement could come from extra capacity and auxiliary supervision rather than from behavioural information per se. The paper says the variants 'differ only in how behavioural information is incorporated,' which is not true when ANP disables all of it. A capacity-matched control with random or constant c would settle this. The stress-test note is right.\n\nThere is also a potential data leak: K-means is fitted on all profiles before the user split, so test-user distribution can influence the cluster centroids used to supervise training. Minor relative to the main issue, but worth fixing. No error bars or significance tests are reported, and code and architecture hyperparameters are not released, so reproducibility is limited.\n\nThe math itself is sound; the paper is honest about the proxy-label limitation. This is not a desk-reject. The authors have a credible proposal and one clear experimental gap. Send to peer review with a request for a matched ablation, pre-split clustering or justification, and uncertainty quantification. A serious referee can make this a much stronger paper.","headline":"A novel ANP extension for load forecasting, but the key ablation doesn't isolate behavioural conditioning—worth a careful referee.","tokens_in":23559,"tokens_out":2983,"would_cite":false,"duration_ms":27245,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Inferred household behavioural classes, embedded as a decoder-conditioning variable in a Neural Process, improve residential short-term load forecasts: a 7.9% average MAE reduction and 6.9% CRPS reduction over a label-agnostic baseline on u","keywords":["residential load forecasting","neural processes","behavioural clustering","feature-wise linear modulation","probabilistic forecasting","smart meter data","consumer heterogeneity","semi-supervised latent variables"],"falsifier":"Retrain the soft-label variant with teacher labels randomly permuted across profiles, breaking any link between cluster and consumption behaviour, under the same splits and hyperparameters; if the MAE/CRPS improvements over the label-agnostic ANP persist, the reported gains are a capacity effect and not behaviour conditioning. Equivalently, measure the label-inference network's agreement with complete-profile K-means assignments on the user-disjoint test set: chance-level agreement would indicate the conditioning signal carries no behavioural information.","tokens_in":22637,"feed_emoji":"⚡","tokens_out":4003,"duration_ms":32439,"temperature":0.7,"pith_summary":"This paper asks whether the behavioural routine behind a household's electricity use can improve near-term load forecasts if that behaviour is inferred from past data and used inside the forecasting model, rather than as an external grouping step. It proposes a behaviour-conditioned Attentive Neural Process: a continuous latent variable captures shared uncertainty across households, while a discrete behavioural latent variable, inferred from the observed context, conditions the decoder through feature-wise modulation and a prompt token. Clustering of complete load profiles supplies weak supervision during training only; at test time the class is inferred from context alone. On user-disjoint SGSC data, the soft-label variant reduces MAE by 7.9% and CRPS by 6.9% on average versus a label-agnostic ANP baseline, with the largest gains when context is short, and achieves lower RMSE than fixed-window deterministic baselines at every horizon.","feed_headline":"Behaviour cues cut household load forecast error 7.9%","feed_subtitle":"A neural process infers each home's consumption style from past data alone, sharpening point and probabilistic forecasts on unseen household","key_machinery":"The central mechanism is the dual-latent Attentive Neural Process: a continuous latent variable z models shared functional uncertainty over plausible load trajectories, while a discrete behavioural latent c, drawn from a context-conditioned categorical prior pψ(c|DC), conditions a Transformer-style decoder through (i) HyperFiLM — a hypernetwork that maps the class embedding to per-layer (γℓ, βℓ) scale/shift parameters applied after LayerNorm, AdaLN(h|c)=(1+γ(c))LN(h)+β(c) — and (ii) a prompt token prepended to the decoder sequence. K-means soft memberships of complete profiles act as the teacher for c during training (with the mixing coefficient annealed from 1 to 0.15), and a semi-supervise","core_discovery":"The paper establishes that behavioural structure discovered offline by K-means clustering of complete two-day load profiles can be distilled into soft class memberships, learned by a label-inference network from partial context, and injected into an Attentive Neural Process decoder via HyperFiLM (per-layer scaling and shifting after layer normalisation) and a prepended prompt token. This conditional decoder, with the continuous latent z kept class-agnostic, yields consistent MAE and CRPS improvements over a label-agnostic ANP across horizons H=1–8 and context lengths C≥48, with the largest improvements (18.3% MAE, 18.8% CRPS at H=1) when context C<60; the soft-label variant outperforms the h","pith_inferences":["The gains suggest behavioural archetypes learned from complete profiles are partially recoverable from a single day of history; a direct test would measure mutual information between the context-inferred class and the complete-profile cluster on held-out contexts.","The method offers a general recipe: cluster the training set on the full target variable, then train a latent-conditioned function approximator to re-infer the cluster from inputs; whether it helps depends on the cluster structure being predictive of future values beyond what a continuous latent captures.","The soft-versus-hard gap hints that a fully Bayesian treatment of the discrete regime, marginalising over c at inference rather than using the prior distribution's mode, could yield further gains.","A clean falsifier: randomise the teacher cluster assignments; if MAE/CRPS gains over the label-agnostic baseline persist, the benefit comes from added model capacity, not from behavioural semantics."],"forward_implications":["A single Neural Process model can serve many households, varying contexts, and horizons without per-household retraining, shifting its predictive distribution with inferred consumption style.","Behavioural conditioning pays off most when history is short: at H=1 and C<60 the soft variant improves over ANP by 18.3% MAE and 18.8% CRPS, which is relevant for cold-start forecasting of new households.","Soft behavioural assignments outperform hard ones, implying that uncertainty in regime membership should be propagated into the forecast rather than collapsed to a single class.","The lower RMSE at all horizons versus deterministic baselines (13.4–18.4% below the strongest baselines) suggests fewer large prediction errors, which matters for operational decisions sensitive to large deviations.","Test-time conditioning relies on context-inferred class distributions rather than complete-profile clusters, avoiding label leakage from future observations."],"fun_headline_variants":["Behaviour-aware model cuts load forecast error 7.9%","Infer home habits to sharpen load forecasts 7.9%","Short-context load forecasts improve 18% with habits","Household behaviour inference boosts forecast accuracy","Load forecast error drops with learned behaviour cues"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The gains depend on the premise that K-means clusters of complete two-day load profiles define a behavioural structure that a label-inference network can reconstruct from a partial day of context, and that this inferred class adds forecasting information beyond what the continuous latent variable and the attention path already provide.","fun_headline_variants_meta":{"raw":{"variants":["Behaviour-aware model cuts load forecast error 7.9%","Infer home habits to sharpen load forecasts 7.9%","Short-context load forecasts improve 18% with habits","Household behaviour inference boosts forecast accuracy","Load forecast error drops with learned behaviour cues"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00017,"raw_usage":{"total_tokens":1151,"prompt_tokens":837,"completion_tokens":314,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":581,"completion_tokens_details":{"reasoning_tokens":237}},"tokens_in":581,"tokens_out":314,"duration_ms":3293,"temperature":1.0,"reasoning_tokens":237,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T21:08:48.909926+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain the soft-label variant with teacher labels randomly permuted across profiles, breaking any link between cluster and consumption behaviour, under the same splits and hyperparameters; if the MAE/CRPS improvements over the label-agnostic ANP persist, the reported gains are a capacity effect and not behaviour conditioning. Equivalently, measure the label-inference network's agreement with complete-profile K-means assignments on the user-disjoint test set: chance-level agreement would indicate the conditioning signal carries no behavioural information.","supporting_citations":[],"review_version":1}