{"id":"7b4129e3-2f97-4fac-830a-ba22e04fe5d9","arxiv_id":"2512.04223","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A conditional variational autoencoder generates diverse, realistic daily activity schedules from demographic labels, outperforming both a most-likely-schedule baseline and a label-blind generative baseline on density estimation.","lead":"The authors build ActVAE, a machine-learning model that generates realistic 24-hour activity schedules (what activities a person does and for how long) conditioned on personal and household characteristics such as age, income, and car access. The work is a step toward cheaper, more flexible activity-based transport demand models that can simulate how schedules might change as populations age or policies shift.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Latent-label entanglement violates the independence assumption behind controllable generation; the paper's own MI estimates show ~37% label leakage and a ~10% conditional bias, but the impact on the central claim is untested.","rationale":"The reader's weakest_assumption identifies the same load-bearing concern: the theoretical requirement of z ⊥ y for controllable generation is violated, and the paper's own evidence (MI and conditional bias) confirms the violation. I agree that this is the most important point because it directly affects the correctness of the generation procedure when sampling from the prior, and thus the validity of counterfactual predictions. The paper is transparent about the limitation but does not quantify its influence on conditional generation fidelity. The proposed concrete test would directly measure whether the entanglement degrades conditional accuracy, thereby settling whether the concern actually undermines the central claim. I keep the verdict as CONDITIONAL because the concern is real but potentially addressable; a direct test could show the model is still practically useful, or could show the conditional capability is overstated. The reader's verdict already accounts for this as a structural weakness, so no change is needed.","tokens_in":34486,"tokens_out":5152,"duration_ms":49153,"concrete_test":"Train ActVAE on NTS data. For each label category y in a held-out set, compute the empirical encoder distribution q_φ(z | x, y) over real schedules x with that label, and aggregate to q(z | y). Compute the average KL divergence D_KL(q(z | y) || p(z)) across y. Then generate synthetic schedules for each y using (i) z ~ p(z) (current method) and (ii) z ~ q(z | y) (ideal posterior). Compare both synthetic sets to the real p(x | y) using the same EMD metrics as in Table 16. If method (ii) substantially outperforms (i), the independence violation is material and the reported conditional performance is limited by entanglement. If (i) and (ii) are comparable, the decoder has learned to compensate and the concern is less severe.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that ActVAE provides controllable conditional generation rests on the independence assumption in Eq. (5)–(7): p(x|y) = ∫ pθ(x|y,z)p(z)dz requires z ⊥ y. The paper's own mutual information analysis (Sec 7.7, Table 18) shows I(z_φ; y) = 0.170 vs I(x; y) = 0.458, i.e., about 37% of the available label information leaks into the latent. Section 7.6 confirms a systematic ~10% under-estimate of label effects (e.g., the employed–unemployed work-participation difference is 0.442 modelled vs 0.481 real, an 8% under-estimate), which the paper attributes to this incomplete disentanglement. This is load-bearing because it directly affects the model's ability to generate schedules for new label distributions (counterfactual scenarios), the primary motivation of the paper. The paper acknowledges the issue but does not quantify its impact on the fidelity of conditional generation; it only reports aggregate bias. Consequently, the claimed conditional capability is not yet demonstrated to be robust under the intended operating procedure (sampling z ~ p(z)).","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ActVAE, a conditional variational autoencoder for generating 24-hour human activity schedules conditioned on socio-demographic labels. The architecture combines an LSTM-based continuous schedule encoding with a structured Gaussian latent and learns a joint conditional-generative model. ActVAE is compared against two baselines: ConditionalRNN, a purely conditional model producing the most-likely schedule, and GenerativeRNN, a non-conditional VAE. Evaluation uses reconstruction losses, feasibility and creativity metrics, a joint density estimation framework based on first-order label marginals, mutual information estimates of latent-label entanglement, case studies on forecasting, data/label availability, and a foundational model, plus runtime reporting. The main claims are that ActVAE produces schedules whose joint distribution with labels is closer to the real distribution than both baselines, and that explicit random variation is essential for schedule realism.","tokens_in":34775,"tokens_out":3082,"duration_ms":32100,"significance":"If the claims hold, ActVAE offers a practical, computationally efficient, open-source generative component for activity-based demand models, replacing compositional sub-models with an end-to-end conditional-generative approach. The paper's strengths include a reproducible implementation in Caveat, five-run reporting with variances, honest documentation of limitations (underestimation of label effects, imperfect disentanglement), ablations, and additional case studies. The mutual information analysis is a useful diagnostic for conditional VAEs in this domain. The work also provides evidence that a large fraction of schedule variance is not explained by available labels, which is a substantive empirical finding for activity-scheduling research.","major_comments":[{"comment":"The controlled-generation guarantee depends on the independence assumption z⊥y in Eq. (5)–(7). The paper's own measurements, however, show substantial label leakage: Table 18 reports I(z_φ;y)=0.170 against an upper bound I(x;y)=0.458, i.e. about 37% of available label information remains in the latent, and §7.6 documents a systematic ~10% underestimation of label effects. Since the stated motivation is counterfactual generation under novel label distributions, the present evaluation (aggregate conditional means on observed labels) does not demonstrate robustness of the central claim under the intended operating procedure. The authors should either quantify the impact of leakage on scenario-level outputs (e.g., by comparing p(x|y) from sampling z~p(z) against an estimate using p(z|y)) or substantially weaken the controllability claim.","section":"§4.3, §7.6, §7.7"},{"comment":"The paper repeatedly refers to 'joint density estimation' and matching the 'joint distribution of schedules and labels', but the evaluation in Eq. (17)–(19) conditions only on first-order label marginals. Higher-order interactions among labels (e.g., age × car access × income) are not tested, even though such interactions are likely relevant for activity-based demand models. This is acknowledged in §6.14, but it means the central 'joint' claim is only supported at the marginal level. Please qualify the claims accordingly or add a small number of second-order conditional evaluations.","section":"§6.14, §7.4–7.6"},{"comment":"No significance tests are reported. Several ActVAE improvements over GenerativeRNN are small relative to the reported run-to-run variance — for example, label-level participations for age group are 0.050±0.006 for ActVAE versus 0.061±0.005 for GenerativeRNN, and timing distances overlap substantially. Given that the paper's framing is comparative ('ActVAE best matches the joint distribution'), the reader cannot distinguish genuine improvement from noise. Paired tests across the five runs or confidence intervals on the EMD differences should be provided, at least for the summary tables.","section":"Table 14, §7.5, Table 16"},{"comment":"The computation 'label variation is responsible for only 16% of schedule variation in MI terms' via (2.3705−1.9897)/2.3705 assumes that the ActVAE latent captures only random information and the GenerativeRNN latent captures all label information; neither assumption is exactly true, as the paper states. Since this number feeds into a broader conclusion about the relative importance of labels, the authors should report the uncertainty in this derived quantity and perhaps show a sensitivity analysis over the MI estimation model.","section":"§7.7"}],"minor_comments":[{"comment":"The hyperparameter notation '4.256.64' is cryptic; please define the tuple (N, S, H) explicitly in the caption or in Section 6.10.","section":"Table 10, Table 11"},{"comment":"The phrase 'For simplicity, we represent all variables as nominal tokens' is clear, but the later label availability study (Appendix F.3) uses up to 16 labels; a brief note on how continuous variables (e.g., age) are discretized would help reproducibility.","section":"Section 5.4"},{"comment":"The expected-value conditionality comparison is informative, but the axes or units of the subplots are not always labeled; consider adding a legend and axis titles to make the panels self-contained.","section":"Section 7.2, Figure 9"},{"comment":"The 200k/100k/50k/25k/12.5k sample sizes appear in Table F.26 but the sampling procedure is not fully described. State whether these are random samples from the combined 2019–2023 NTS data and how stratification was handled.","section":"Appendix F.2"},{"comment":"There are several typographical and formatting issues, e.g., 'V ariation Auto-Encoder theory' in the section header, occasional double hyphens in Table E.22, and inconsistent use of 'V AEs' vs 'VAEs'. A final proofread is recommended.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern about latent-label entanglement lands: the paper's own MI and bias numbers put the central controllability claim in question. However, the issue is not fatal — the authors already identify the mechanism and could address it with additional evaluation. Given the otherwise thorough empirical design, I would encourage a revision that either quantifies the operational impact of the entanglement or explicitly redefines the claimed scope of conditionality."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Fred, here's the short version: ActVAE does what it says — it conditions a schedule-generating VAE on sociodemographic labels and improves on both a most-likely-schedule RNN and a non-conditional generative RNN. But the gain over the non-conditional baseline is modest, and the conditional capability comes with a systematic shrinkage: label effects are underestimated by roughly 10%, which the authors trace to incomplete disentangling of labels from the latent. That's not hidden; they report it plainly.\n\nWhat's genuinely new is the application: a conditional VAE for full-day activity schedules, built on their earlier ContRNN. The architecture is a standard CVAE — no new learning machinery — but the label-conditioned density evaluation framework they build is a real contribution, and they use it thoroughly: held-out test data, five runs, EMD on participations/transitions/timings, plus feasibility, creativity, and mutual-information estimates. They also ship open-source code (Caveat) and use the UK National Travel Survey, so the results are reproducible.\n\nSoft spots, in proportion. The 'joint density estimation' label overstates things: it covers first-order marginals only, as Section 6.14 admits. The MI analysis shows I(z;y)=0.17 vs I(x;y)=0.458 — around 37% of the available label information sits in the latent, violating the z⊥y assumption that makes sampling z~p(z) a clean conditional generator. The paper acknowledges this and reports the ~10% bias, but it doesn't directly test the operating scenario that motivated the work: generating schedules for a genuinely new label distribution and checking whether the label effects are recovered. The forecasting case study is qualitative. So the central conditional claim holds, but with a caveat: the model is a conditionally-biased generator, not a clean controllable one.\n\nNo significance tests were run; given five runs per model and the size of the EMD differences, the main comparisons would survive eyeballing, but intervals would help. The improvement over GenerativeRNN is often small — the paper says so itself, and its MI finding that labels explain only ~16% of schedule variation suggests conditional capability is inherently limited by the data.\n\nWho's this for: anyone working on activity-based demand models or synthetic schedule generation. The paper is honest, well-documented, and reproducible. It deserves a serious referee. I'd want the revision to include a counterfactual validation (hold out a label segment, train on the rest, generate for the held-out segment) and a classifier-based conditional metric like the one they suggest. Not a desk reject.","headline":"ActVAE conditions schedule generation on labels and beats its baselines, but the conditional gain is modest and the known ~10% label-effect bias means the controllable-generation claim needs a stronger counterfactual test.","tokens_in":35258,"tokens_out":4512,"would_cite":true,"duration_ms":40394,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"ActVAE, a conditional generative model, produces diverse and realistic activity schedules that match the joint distribution of schedules and labels better than either a purely conditional or a purely generative baseline.","keywords":["activity-based demand modelling","conditional variational autoencoder","activity schedules","generative modelling","joint density estimation","mutual information","disentanglement","deep learning"],"falsifier":"A direct falsifier would be a two-sample test between real and ActVAE-generated schedules for a specific label group (e.g., employed males aged 30-39); if the generated conditional distribution differs significantly from the real one on key statistics (work participation, activity start times, sequence length), the paper's central claim fails. More mechanistically, one could compute the mutual information I(z; y) on a held-out set and check whether it is near zero (the paper finds it near 0.17 nats vs. the 0.46 available); a value that does not shrink with increased β would indicate that the a","tokens_in":34349,"feed_emoji":"🗓️","tokens_out":4202,"duration_ms":35841,"temperature":0.7,"pith_summary":"ActVAE is a conditional variational autoencoder that models a whole day's activity schedule as a sequence of activity types with durations, conditioned on personal and household attributes such as age, income, and car access. The paper's central claim is that explicitly combining a structured latent generative process with label conditioning yields schedules that are both individually realistic and collectively diverse, and that this joint approach matches the real distribution of schedules-and-labels better than a conditional-only model (which always outputs the most likely schedule) or a generative-only model (which ignores labels). If correct, the model offers a fast, drop-in scheduling component for activity-based transport demand models and, more broadly, shows that most day-to-day variability in schedules is random rather than explained by demographics. The paper also demonstrates that the model can forecast changes in scheduling under demographic projections, using as few as 25,000 training samples.","feed_headline":"Combining labels and randomness yields realistic activity schedules","feed_subtitle":"ActVAE generates millions of diverse daily schedules that match trip-survey distributions, for transport demand modelling.","key_machinery":"The central object is ActVAE, a conditional variational autoencoder (CVAE) in which a schedule is encoded into a six-dimensional latent vector z with a standard normal prior, and the decoder maps (z, y) to a variable-length sequence of (activity type, duration) tokens; controllable generation relies on the assumption that z and y are independent, so sampling z from the prior and conditioning on y produces samples from p(x|y). The architecture uses stacked LSTM units for the schedule encoder/decoder, a label-encoder block that sums embeddings of categorical attributes, and a combined reconstruction loss (cross-entropy for activity types, MSE for durations) with a β-weighted KL regularisation","core_discovery":"On its own terms, the paper establishes that a conditional VAE—termed ActVAE—can learn the joint distribution of 24-hour activity schedules and demographic labels from roughly 59,000 UK travel-diary schedules, and can then generate new schedules conditional on arbitrary label sets. Trained with a continuous encoding of activity types and durations, and an LSTM-based encoder/decoder with a six-dimensional Gaussian latent bottleneck, ActVAE outperforms both a discriminative conditional baseline and a non-conditional generative baseline on a suite of joint-density metrics: participation rates, transition probabilities, and activity timing. The authors find that label-conditioned variation is re","pith_inferences":["If the disentanglement defect is cured—e.g., by increasing β or using auxiliary objectives—the conditional accuracy could improve substantially while retaining generative diversity, since the paper's own mutual-information numbers suggest a direct trade-off between them.","The claim that 'labels explain only ~16% of schedule variation' is conditional on the chosen label set; including richer context like weather, day-of-week, or network-level accessibility might raise that share, though the paper's label-availability experiment suggests diminishing returns.","The joint-density evaluation framework, conditioning on each label marginal, could be extended to higher-order interactions (e.g., age × gender) to catch cases where ActVAE's additive label encoder might miss interactive effects.","A natural stress test for the model would be to benchmark it on a second country's travel diary with a different label system, to see whether the architecture transfers without tuning."],"forward_implications":["ActVAE can be plugged into existing activity-based demand models as a single joint generator, replacing hand-built compositions of choice sub-models.","The model can simulate counterfactual scenarios—e.g., population ageing or policy changes—by feeding new label distributions, as demonstrated in the forecasting case study.","Explicitly modelling random variation is necessary: a purely conditional model that outputs the most likely schedule fails to match the real diversity of schedules, even with the same conditional architecture.","The approach is data-efficient, remaining viable with as few as 25,000 schedule samples, and can be trained on multi-year data with year-conditioning to improve performance.","The residual ~10% under-estimation of label effects means that scenario forecasts are biased toward the population mean, which should be considered when interpreting predictions."],"fun_headline_variants":["Conditional VAE generates realistic activity schedules from demographics","ActVAE: deep generative model for diverse daily schedules","Generating human-like activity schedules with a conditional VAE","Deep learning crafts realistic daily schedules from label data","Randomness and labels combine to model daily activity patterns"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The central assumption is that the latent random variable z that drives schedule diversity is independent of the conditioning labels y; the paper's own mutual-information analysis shows that ActVAE's latents still encode about 37% of the label information, and the model consequently under-estimates label effects by roughly 10%—so if one requires exact conditionality, the load-bearing premise is not fully satisfied.","fun_headline_variants_meta":{"raw":{"variants":["Conditional VAE generates realistic activity schedules from demographics","ActVAE: deep generative model for diverse daily schedules","Generating human-like activity schedules with a conditional VAE","Deep learning crafts realistic daily schedules from label data","Randomness and labels combine to model daily activity patterns"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000297,"raw_usage":{"total_tokens":1499,"prompt_tokens":623,"completion_tokens":876,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":367,"completion_tokens_details":{"reasoning_tokens":799}},"tokens_in":367,"tokens_out":876,"duration_ms":7358,"temperature":1.0,"reasoning_tokens":799,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T18:36:49.161991+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct falsifier would be a two-sample test between real and ActVAE-generated schedules for a specific label group (e.g., employed males aged 30-39); if the generated conditional distribution differs significantly from the real one on key statistics (work participation, activity start times, sequence length), the paper's central claim fails. More mechanistically, one could compute the mutual information I(z; y) on a held-out set and check whether it is near zero (the paper finds it near 0.17 nats vs. the 0.46 available); a value that does not shrink with increased β would indicate that the a","supporting_citations":[],"review_version":1}