{"id":"3ee9da11-7a52-4999-9c13-148f2258ba51","arxiv_id":"2501.19364","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"CoSTI applies consistency training to multivariate time series imputation, claiming diffusion-level accuracy at a fraction of the inference cost.","lead":"This paper adapts consistency models, a fast single-step generative framework, to fill missing values in multivariate time series and traffic and clinical data. It reports up to 98% faster imputation than diffusion models with accuracy close to diffusion baselines, though the speed measurement and the single-step framing require scrutiny.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 3 timing appears to count one forward pass, while Algorithm 2's deterministic imputation uses 100 passes and a median; if so, the 98% speedup is overstated by roughly 100x.","rationale":"The reader's weakest assumption is exactly the load-bearing issue: Algorithm 2 requires 100 stochastic forward passes and a median, while Table 3 and the complexity analysis claim a single sampling step / single forward pass. I agree this is the central risk. The paper is otherwise a reasonable application of consistency training to MTSI, with ablations and a code release that would settle the question. Since the reader's verdict is already CONDITIONAL and the requested clarification matches the condition, no verdict adjustment is needed. However, the concern is serious enough that without a clarifying timing measurement, the central 98% claim should not be accepted as stated. I do not see a stronger alternative concern: the accuracy deficits on PEMS08 and ETTh1 are notable, but the paper explicitly frames overall accuracy as 'comparable', and the timing asymmetry is the claim that determines acceptance.","tokens_in":24528,"tokens_out":5891,"duration_ms":64441,"concrete_test":"Inspect the released repository (github.com/javiersgjavi/CoSTI) and time the exact pipeline used for Tables 4-6: run Algorithm 2 with 100 samples and median imputation over one full test set (e.g., METR-LA or AQI-36) on the same RTX A5000, and compare the wall-clock to Table 3 and Table A.13. If the full 100-sample pipeline matches Table 3, the claim stands. If it is approximately 100x larger (or any factor substantially exceeding the batching advantage), Table 3 reports a single-sample timing, and the headline speedup should be recomputed with matched sample counts, e.g., CoSTI-100 versus DDPM-1 and CoSTI-1 versus DDPM-50/100.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim hangs on the inference-time comparison in Table 3. Algorithm 2 (Section 4.3) runs j=1..100 stochastic forward passes per window and then takes the median; Section 4.1.4 sets N=100 for this deterministic imputation. Yet Section 5.4.1 says Table 3 was obtained \"using a single sampling step\", and Section 4.3's complexity argument says \"CoSTI requires only a single forward pass\". These two accounts cannot both describe the timed pipeline. If the measured times include only one of the 100 passes, the true per-window work for CoSTI is ~100 model evaluations, so the step-count advantage over T=50/100 diffusion steps disappears (and reverses for T=50). For AQI-36, CSDI uses T=100 and CoSTI uses 100 samples, so step count alone predicts equal forward passes; the reported 44x speedup would then require a 44x per-forward-pass architecture advantage, which is neither claimed nor supported. Section 5.4.1 also does not state whether the DDPM baselines' timings include one imputation sample or multiple, so the comparison may be asymmetric. This is a resolvable but load-bearing ambiguity.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CoSTI, an adaptation of consistency models to multivariate time series imputation. The method trains a consistency model with a spatio-temporal U-Net architecture, conditional information, curriculum learning over the number of noise levels, and a deterministic imputation stage based on the median of 100 stochastic forward passes. The authors report imputation accuracy comparable to DDPM-based baselines (CSDI, PriSTI, TIMBA) across several datasets and missing-data scenarios, and claim a reduction of inference time of up to 98% relative to the diffusion baselines. The paper also includes benchmark comparisons, sensitivity analyses, ablations, and a downstream task evaluation.","tokens_in":24714,"tokens_out":3473,"duration_ms":36665,"significance":"If the main claim holds, CoSTI would be the first application of consistency models to multivariate time series imputation and would offer a practically useful accuracy/speed trade-off in a domain where diffusion models are accurate but slow. The paper has clear strengths: the code and Docker setup are released, the experimental coverage is broad (six datasets, multiple missingness scenarios, ablations, downstream evaluation), and the accuracy numbers are mostly competitive with the diffusion baselines. However, the central speed claim rests on a timing comparison whose protocol is internally inconsistent with the published sampling algorithm, and this issue must be resolved before the contribution can be assessed.","major_comments":[{"comment":"There is a direct internal contradiction between the timed inference protocol and the sampling algorithm. Algorithm 2 runs j=1..100 stochastic forward passes per window and then takes an element-wise median, and Section 4.1.4 states N=100 for this deterministic imputation. Section 4.3, however, claims that 'CoSTI requires only a single forward pass' and derives a ~T-fold speedup over T-step diffusion models. Section 5.4.1 states that Table 3 was obtained 'using a single sampling step'. If the reported CoSTI times include only one of the 100 passes, then the true per-window work is about 100 model evaluations; the step-count advantage over T=50 or T=100 diffusion steps disappears (and for AQI-36, where CSDI uses T=100, the forward-pass counts are equal). In that case the reported 44x speedup on AQI-36 would have to come from a per-forward-pass architectural advantage that is neither claimed nor demonstrated. The authors must state unambiguously whether Table 3 reports end-to-end time for the full Algorithm 2 procedure or for a single forward pass, and must report the corresponding baselines under the same protocol.","section":"§4.1.4, §4.3 (Algorithm 2), §5.4.1 (Table 3)"},{"comment":"The complexity analysis in Section 4.3 is not a valid substitute for an empirical timing comparison once Algorithm 2 is used. Even if a single consistency-model forward pass is cheaper than one diffusion step, the paper must report the total number of model evaluations per imputed window for each method, including the 100 samples used for the median and the number of stochastic samples used for each diffusion baseline. At present, the text does not state whether the CSDI/PriSTI/TIMBA timings include one stochastic imputation sample or several, so the comparison may be asymmetric in either direction. The step-count argument in Section 4.3 and the '1/T times faster' statement in Section 5.4.1 are only valid if every method is measured with the same number of stochastic samples per window.","section":"§5.4.1 (Table 3), §4.3"},{"comment":"The dataset-specific second noise level sigma_i2 is described as 'experimentally determined' and its values are reported in Table 2, but the selection procedure is not specified. If sigma_i2 was tuned using the same test scenarios that are later reported in Tables 4 and 5, the accuracy advantage of CoSTI-2 (and the claimed 'controllable trade-off') could be optimistically biased. The authors should state whether sigma_i2 was chosen on validation data only, and if not, how the reported CoSTI-2 numbers should be interpreted.","section":"§5.3 (Table 2), §5.4.1 (Tables 4-5)"}],"minor_comments":[{"comment":"In Eq. (10), the second displayed formula is written as c_in(σ) = ln(σ)/4, but from the context it should be c_noise(σ) = ln(σ)/4; this appears to be a typo.","section":"Eq. (10)"},{"comment":"The expression for c_skip(σ) appears to be missing a square in the denominator: the standard form is σ_data^2 / ((σ - σ_min)^2 + σ_data^2), while the paper writes (σ - σ_min) without the square. Please check and correct.","section":"Eq. (4)"},{"comment":"The comment 'keep only missing values' next to the line ̂X_t ← [ ̃X_t ⊙ M_t + ̂X_t ⊙ (1 - M_t)] is misleading: the operation keeps the observed values from the input and the imputed values at missing positions, which is the opposite of keeping only missing values.","section":"Algorithm 2"},{"comment":"The statement that CoSTI is 'approximately 1/T times faster' than diffusion models conflates step count with wall-clock time; it should be phrased as a reduction in the number of network evaluations, with wall-clock speedup depending on per-forward-pass cost and any overhead from the 100-sample median loop.","section":"§5.4.1"}],"recommendation":"major_revision","confidential_remarks":"The central speed claim is currently unverifiable because the published sampling algorithm and the timed protocol are inconsistent. The issue is potentially fixable by clarifying and re-measuring, which is why I do not recommend rejection at this stage, but the revision must provide a transparent timing protocol and, ideally, report wall-clock times for the exact Algorithm 2 procedure used to produce the reported imputations."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is the first application of consistency models to multivariate time series imputation, and the engineering is mostly solid. But the central speed claim is built on a timing ambiguity that the paper does not resolve. If the reported times count only one of Algorithm 2's 100 forward passes, the 98% reduction is roughly 100x overstated.\n\nWhat's genuinely new: CoSTI brings consistency training to MTSI, with a conditional U-Net architecture borrowing from CSDI/PriSTI/TIMBA. The evaluation is broad: six datasets, multiple missingness scenarios, sensitivity analysis, downstream task, ablations of the conditional head, STFEM, NEM, and curriculum schedule. The ablations are informative and the best-config choices (linear scheduler, AdamW-ScheduleFree) are backed by experiments. Code and Docker are promised. That's real work.\n\nThe problem is the timing. Algorithm 2 runs 100 stochastic imputations per window and takes a median. Section 5.4.1 says Table 3 was obtained \"using a single sampling step,\" and Section 4.3 claims \"a single forward pass.\" Those cannot both describe the measured pipeline. If Table 3 times exclude the other 99 passes, the per-window cost is 100x higher, and the step-count advantage over T=50/100 diffusion steps essentially disappears. For AQI-36, CSDI uses T=100 and CoSTI with 1 step still needs 100 samples, so equal forward passes; the reported 44x speedup would then require a 44x per-pass advantage, which is not argued. The authors also don't say whether baseline diffusion timings include one sample or many, so the comparison might be asymmetric. This is not a minor cosmetic inconsistency; it's the load-bearing pillar of the paper.\n\nAccuracy-wise, the paper claims \"comparable,\" but CoSTI is noticeably worse on PEMS08 (MAE 11.09 vs 9.87 for CSDI) and ETTh1 (0.38 vs 0.33), and closer elsewhere. The 2-step variant helps but doesn't close the gap. Not disqualifying, but the \"on par\" language is too strong.\n\nOther notes: sigma_i2 is tuned per dataset on the same benchmarks, a mild overfitting to the test sets, but common in this literature. Self-citations to TIMBA are appropriate; no citation problem.\n\nWho is this for? Anyone working on fast generative imputation or consistency models for time series. It deserves peer review; the idea is worth pursuing. But as it stands, I would not accept the speed claim without a clear statement of the number of forward passes in the timed pipeline and a matching code check. If the timing is confirmed to include all 100 passes, the contribution stands; if not, the headline fails. My recommendation: send to referees, require the timing clarification and a rerun if necessary.","headline":"First consistency-model adaptation for MTSI, but the headline 98% speedup is compromised by an unresolved contradiction between Algorithm 2's 100-pass median and the text's 'single forward pass' claim.","tokens_in":25319,"tokens_out":4155,"would_cite":false,"duration_ms":39873,"reading_group":"maybe","serious_thinker":"unclear","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"CoSTI adapts consistency models to multivariate time series imputation, matching diffusion-based accuracy while cutting inference time by up to 98%.","keywords":["multivariate time series imputation","consistency models","consistency training","generative imputation","spatio-temporal data","diffusion models","missing data","real-time inference"],"falsifier":"Run Algorithm 2 end-to-end on METR-LA's test set, timing all 100 forward passes plus the median aggregation on the same GPU, and compare with one complete 50-step diffusion reverse chain; if the per-window time ratio is not close to 1:50, the up-to-98% reduction claim is refuted.","tokens_in":24284,"feed_emoji":"⏱️","tokens_out":10517,"duration_ms":101492,"temperature":0.7,"pith_summary":"CoSTI adapts consistency models — generative models trained to map a noised input directly back to the clean signal in one step — to multivariate time series imputation. The paper argues that with conditional inputs (a linear-interpolation placeholder, the missing-value mask, and a fixed spatial graph) and a curriculum over noise levels, a consistency model can match the imputation accuracy of diffusion-based state-of-the-art while reducing inference time by up to 98%. If this is right, generative imputation becomes practical for real-time ICU monitoring, traffic control, and energy systems, where the iterative denoising of DDPMs has been too slow. The evidence spans six datasets with point, block, and natural missingness, and includes a downstream node-prediction check.","feed_headline":"Consistency models cut imputation time up to 98%","feed_subtitle":"Single-step generative imputation matches diffusion accuracy, enabling real-time ICU and traffic monitoring.","key_machinery":"The central object is the consistency function $\\boldsymbol{f}_\\theta(\\mathbf{x}_\\sigma, \\sigma)$, parameterized as $c_{\\text{skip}}(\\sigma)\\mathbf{x} + c_{\\text{out}}(\\sigma) F_\\theta(\\ldots)$ and implemented by a U-Net with spatio-temporal feature extraction modules (STFEM), noise-extraction modules (NEM), and a separate conditioning branch that ingests the linear-interpolation estimate, the mask, and the graph. The training objective enforces self-consistency between adjacent noise levels of the PF-ODE using a Pseudo-Huber metric, while a linear curriculum scheduler raises the number of noise levels from 10 to 200 during training. This machinery replaces the diffusion reverse chain: one consistency evaluation maps noise to data where a DDPM would need $T = 50$ or $T = 100$ iterative denoising steps.","core_discovery":"The core discovery is that consistency training transfers from image generation to structured spatio-temporal imputation without needing a pretrained diffusion teacher. CoSTI learns a consistency function $f_\\theta$ that takes a $\\sigma$-noised observation together with conditional information and outputs the clean imputation, trained so that outputs for adjacent noise levels along the probability-flow ODE agree. At inference, each sample is produced in one (or two) network evaluations, and the final deterministic imputation aggregates 100 such samples by element-wise median. The paper reports per-dataset inference times of 0.005 to 0.48 hours versus 0.22 to 18.19 hours for DDPM-based models, with MAE and MSE close to those baselines across all six datasets.","pith_inferences":["The same conditional consistency formulation should transfer to forecasting by treating the future window as the missing block and using observed history as conditioning; this reuses Algorithm 1 without architectural changes.","The 100-sample median used for deterministic imputation could double as an uncertainty estimator: the spread of the samples can be summarized as confidence intervals for clinical or traffic decisions, which the paper does not report.","The fixed-graph assumption is the least-transferable design choice; on networks with changing topology, a learnable or time-varying adjacency matrix would be the natural next test, and the ablation results indicate the conditioning head would need to absorb that change."],"forward_implications":["Generative imputation can operate at speeds suitable for real-time ICU monitoring and traffic control rather than offline batch processing.","Because CoSTI is trained from scratch with consistency training, it avoids the extra cost of first training a diffusion model and then distilling it.","The two-step variant CoSTI-2 offers a direct accuracy/speed dial, improving MAE and MSE on most datasets with only one additional forward pass.","On the reported accuracy-time trade-off, CoSTI occupies the Pareto frontier, so no evaluated method is simultaneously faster and more accurate in the tested scenarios."],"supporting_citations":[{"why":"Defines consistency models and consistency training; the framework CoSTI adapts.","marker":"[11]"},{"why":"Introduces conditional score-based diffusion for time series imputation; supplies the conditional architecture lineage and a main accuracy baseline.","marker":"[17]"},{"why":"Proposes PriSTI's conditional diffusion framework and the spatio-temporal feature extraction ideas CoSTI builds on; a main accuracy baseline.","marker":"[18]"},{"why":"Provides the Karras noise schedule, c_skip/c_out parameterization, and PF-ODE formulation used in training.","marker":"[31]"},{"why":"Supplies improved consistency-training techniques (lognormal noise sampling, curriculum schedulers) that CoSTI adapts and compares against.","marker":"[26]"},{"why":"TIMBA is the most recent DDPM baseline; this paper's results provide the reported CSDI, PriSTI, and TIMBA comparisons.","marker":"[25]"},{"why":"Defines the benchmark datasets, graph construction, and evaluation protocol that CoSTI inherits.","marker":"[6]"},{"why":"Supplies consistency-model training recipes for optimizers and curriculum design that inform CoSTI's setup.","marker":"[27]"}],"fun_headline_variants":["CoSTI: one-step imputation, 98% faster than diffusion","Single-step generative imputation: diffusion accuracy at 98% less time","Imputation in one step: CoSTI cuts time by 98%, keeps accuracy","Consistency models: real-time imputation without accuracy loss","98% faster imputation with consistency models, diffusion-level precision"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The speed claim rests on the reported inference time covering the full imputation procedure in Algorithm 2, which runs 100 stochastic forward passes per window and takes the median; if Table 3 counts only one of those passes, the 98% reduction would shrink by roughly a factor of 100 and the central comparison would not hold.","fun_headline_variants_meta":{"raw":{"variants":["CoSTI: one-step imputation, 98% faster than diffusion","Single-step generative imputation: diffusion accuracy at 98% less time","Imputation in one step: CoSTI cuts time by 98%, keeps accuracy","Consistency models: real-time imputation without accuracy loss","98% faster imputation with consistency models, diffusion-level precision"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000944,"raw_usage":{"total_tokens":3993,"prompt_tokens":867,"completion_tokens":3126,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":483,"completion_tokens_details":{"reasoning_tokens":3031}},"tokens_in":483,"tokens_out":3126,"duration_ms":23289,"temperature":1.0,"reasoning_tokens":3031,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T20:21:19.502355+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run Algorithm 2 end-to-end on METR-LA's test set, timing all 100 forward passes plus the median aggregation on the same GPU, and compare with one complete 50-step diffusion reverse chain; if the per-window time ratio is not close to 1:50, the up-to-98% reduction claim is refuted.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines consistency models and consistency training; the framework CoSTI adapts."},{"cited_title":"Tashiro, J","cited_arxiv_id":null,"evidence_quote":"Introduces conditional score-based diffusion for time series imputation; supplies the conditional architecture lineage and a main accuracy baseline."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Proposes PriSTI's conditional diffusion framework and the spatio-temporal feature extraction ideas CoSTI builds on; a main accuracy baseline."},{"cited_title":"Karras, M","cited_arxiv_id":null,"evidence_quote":"Provides the Karras noise schedule, c_skip/c_out parameterization, and PF-ODE formulation used in training."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies improved consistency-training techniques (lognormal noise sampling, curriculum schedulers) that CoSTI adapts and compares against."},{"cited_title":"TIMBA: Time series Imputation with Bi-directional Mamba Blocks and Diffusion models","cited_arxiv_id":"2410.05916","evidence_quote":"TIMBA is the most recent DDPM baseline; this paper's results provide the reported CSDI, PriSTI, and TIMBA comparisons."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the benchmark datasets, graph construction, and evaluation protocol that CoSTI inherits."}],"review_version":1}