{"id":"f4bf6675-d1c0-4225-9d7e-9f114072c523","arxiv_id":"2505.02417","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"T2S uses a length-adaptive VAE and flow-matching diffusion transformer to generate variable-length time series from text captions, trained on a new fragment-level caption dataset.","lead":"The paper introduces T2S, a diffusion model that generates time series from natural language captions, and TSFragment-600K, a dataset of over 600,000 time series-caption pairs. It claims state-of-the-art generation quality across 13 datasets, but the evaluation has gaps around baseline fairness and length generalization.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The SOTA claim is not yet supported: WAPE/MSE/MRR@10 are computed against the original series whose GPT-generated caption was used as the prompt, so the evaluation measures inversion of the captioning pipeline rather than semantic generation from new captions.","rationale":"The paper has real strengths: it introduces a large fragment-level dataset, proposes a coherent LA-VAE plus flow-matching DiT architecture, and includes ablations showing that text conditioning matters. Those contributions are independent of my concern. My concern is narrower but load-bearing: the empirical protocol does not test the paper's central generalization claim. Because every test caption is generated from the corresponding ground-truth series, the evaluation rewards inverting the captioning function or memorizing the training set rather than synthesizing a plausible series from a novel specification. The reader's weakest assumption about caption circularity is related, but the deeper issue is that WAPE, MSE, and MRR@10 are reconstruction metrics computed against the original series, so even a fair comparison across baselines does not validate semantic alignment. The absence of error bars and the restriction of arbitrary-length tests to {24, 48, 96} are additional but secondary weaknesses. If the proposed held-out caption experiment preserves T2S's ranking, the SOTA claim would become credible; until then, I would keep the reader's conditional verdict while requiring this re-evaluation.","tokens_in":14471,"tokens_out":6671,"duration_ms":90430,"concrete_test":"Construct a held-out evaluation set of at least 500 unseen time series whose captions are written by a different annotator (human or GPT-4o) and are not filtered through the same embedding-selection pipeline used for TSFragment-600K. Freeze the trained T2S model and recompute WAPE, MSE, and MRR@10 on this set against both the original series and human judgments of caption-series alignment. If T2S's margin over the same baselines shrinks materially, or if human raters cannot match generated series to their captions, the SOTA claim does not generalize beyond the auto-captioning loop.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (Sec. 6) is that T2S generates semantically aligned time series of arbitrary lengths and is state of the art across 13 datasets. The load-bearing assumption is that the reported metrics measure text-to-series semantic alignment. As implemented, they do not. TSFragment-600K captions are produced by GPT-4o-mini from the same time series fragments used for training and evaluation (Sec. 2.2), and the paper does not state a train/test split or a held-out set of human-authored captions. WAPE and MSE (Eq. 11) compare a generated series directly to the original series y_i, and MRR@10 (Eq. 12) ranks generated candidates by cosine similarity to that same ground-truth y_i. A model that memorizes training series and uses the caption as a retrieval key can therefore score well without demonstrating that it can satisfy a genuinely new caption, such as a user-written caption never paired with that series. The arbitrary-length part of the claim is also only tested at lengths 24, 48, and 96, with no evidence for unseen lengths. The comparison to DiffusionTS and TimeVAE is fair in the narrow sense that all methods see the same captions, but it inherits the same validity problem, so the relative ranking does not establish the absolute capability claimed.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper defines the text-to-time-series (T2S) generation task at three caption granularities (point, fragment, instance), introduces a new fragment-level dataset TSFragment-600K with over 600,000 GPT-generated caption-series pairs, and proposes a diffusion-based model T2S that combines a length-adaptive VAE (LA-VAE) with a Diffusion Transformer denoiser trained under flow matching and classifier-free guidance. Interleaved training across lengths is used to support generation at arbitrary lengths. The authors evaluate T2S on 13 datasets from 12 domains against DiffusionTS, TimeVAE, GPT-4o-mini, and Llama-3.1-8b, and claim state-of-the-art results on fragment-, point-, and instance-level benchmarks.","tokens_in":14758,"tokens_out":4880,"duration_ms":59563,"significance":"If the empirical claims are supported, the paper would make a useful contribution: it provides a large fragment-level text-time-series dataset, a domain-agnostic generation framework, and a length-adaptive latent VAE with a consistency loss. The theoretical framing is standard and internally consistent; there is no mathematical circularity in the flow-matching derivation. The release of code and dataset (GitHub link in Section 1) is a concrete strength. However, the current evaluation does not establish the central claim of semantically aligned generation from natural-language prompts, because the captions are generated from the same time series used for evaluation and the metrics measure fidelity to those original series. The 'arbitrary length' claim is also only tested at the three training lengths. These issues are fixable with additional experiments, but they are load-bearing for the stated SOTA claim.","major_comments":[{"comment":"The evaluation is circular with respect to the text-to-series semantic-alignment claim. TSFragment-600K captions are generated by GPT-4o-mini from the same time series fragments used for training and evaluation, and no train/test split or held-out human-authored caption set is described. WAPE and MSE in Eq. (11) directly compare generated series to the original y_i, and MRR@10 in Eq. (12) ranks candidates by cosine similarity to that same ground-truth y_i. A model that memorizes training series and uses the caption as a retrieval key can score well without demonstrating that it can satisfy a genuinely new caption. The authors should evaluate on held-out captions (e.g., human-written prompts or prompts not paired with any evaluated series) and report metrics that measure semantic match independently of the original series, such as retrieval against a caption-based latent space or human preference judgments.","section":"Section 2.2, Section 4.2, Eqs. (11)-(12)"},{"comment":"The 'arbitrary length' claim is only supported for the three training lengths {24, 48, 96}, plus a fixed length of 2048 for SUSHI. There is no evidence for interpolation to unseen lengths or extrapolation beyond the trained range. The paper should report generation quality at lengths not seen during training, e.g., 12, 72, 192, or explicitly restrict the claim to 'lengths within a specified range' with demonstrated coverage of that range.","section":"Section 3.2, Table 1, Table 2"},{"comment":"The comparison to DiffusionTS and TimeVAE is potentially confounded by data budget. The footnote states that T2S uses interleaved training across lengths while baselines require fixed-length training and evaluation. If T2S sees all lengths of the same underlying series during training while each baseline sees only one length, T2S may benefit from more effective training data even without better text-series alignment. The authors should provide an ablation in which baselines receive the same interleaved or multi-length data, or otherwise control for the total amount of training data seen by each method.","section":"Section 4.2, Table 1 footnote"},{"comment":"All reported numbers are point estimates with no error bars, multiple seeds, or significance tests. Given the small numeric gaps in some cells (e.g., MRR@10 values differing by 0.01-0.02 in Table 1), the SOTA claim is not statistically grounded. The authors should report mean and standard deviation over at least three independent runs and state whether the observed differences are significant.","section":"Section 4.1, Tables 1-2"}],"minor_comments":[{"comment":"The definition of MRR@10 is incomplete: the threshold in cos(·) > threshold is never specified, and the procedure for forming the set of 10 generated candidates is not described. Please define the threshold and the candidate generation protocol.","section":"Section 4.1, Eq. (12)"},{"comment":"The description 'we adapted the dataset by concatenating each time series point with its associated text' is unclear. Please specify how point-level annotations are converted into training prompts and what sequence lengths result from this adaptation.","section":"Section 4.1, Point-Level Dataset"},{"comment":"The notation is inconsistent: line 12 iterates 'lengthi = 1 to k' but line 13 filters by len(s) = i, while earlier k denotes the number of datasets and lengths are denoted l1,...,ld. Please clarify whether the loop is over datasets or over lengths, and define how multiple lengths within one dataset are handled.","section":"Section 3.3, Algorithm 1"},{"comment":"The token limit ⟨M⟩ is introduced but its value is never given. Please report the actual token limit used for caption generation.","section":"Section 2.2"},{"comment":"The consistency-loss weight λ in Eq. (10) is not specified anywhere. Please report its value and the influence of λ on reconstruction quality.","section":"Section 3.2"},{"comment":"There are small typos: 'noise estimater' should be 'noise estimator' in the text following Eq. (4).","section":"Section 3.1"}],"recommendation":"major_revision","confidential_remarks":"The central architectural contribution is plausible, but the evaluation circularity is severe enough that the abstract-level SOTA claim should not remain in its current form without a held-out caption evaluation. I would ask the editor to require the authors to either provide such an evaluation or substantially soften the claims before publication. The arbitrary-length claim also needs to be scoped to tested lengths unless additional experiments are added."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take: T2S is worth a look for the dataset and the length-adaptive VAE, but the evaluation as written doesn't support the \"state of the art, semantically aligned, arbitrary length\" headline. The central gap is that captions in TSFragment-600K are generated by GPT-4o-mini from the same time series that are later used as ground truth in the evaluation. There's no held-out set of human-authored captions, so WAPE, MSE, and MRR@10 measure how well the model can reconstruct a series from a caption that was generated from that same series. A model that memorizes training series and uses the caption as a retrieval key can look great without showing it can satisfy a genuinely new caption. That's a serious problem for the claim of semantic alignment.\n\nWhat is genuinely new: TSFragment-600K is a large fragment-level text-time series dataset spanning 12 domains; that alone is a community resource. The LA-VAE with consistency loss for variable-length inputs is a sensible architectural addition, and the interleaved training scheme is simple and plausibly effective. The diffusion components are standard borrowings from image/audio work, but that's not a flaw in itself. The ablation showing text guidance matters is nice, and the point/instance-level experiments on datasets without fragment captions partially mitigate the circularity, though they still evaluate against the original series.\n\nSoft spots beyond circularity: \"Arbitrary length\" is supported only at lengths 24, 48, and 96; no evidence for unseen lengths. Baselines are trained per-length while T2S sees all lengths, so the comparison conflates architecture with data exposure. There are no error bars anywhere; given the reported differences, we can't tell what's significant. Minor formatting issues in Table 1 aside, the math is straightforward and the citation pattern is fine.\n\nThis paper is for people working on time series generation, multimodal representation, or dataset construction. It deserves a serious referee: the dataset alone is worth peer review, but the revision needs a held-out caption evaluation (e.g., human-written captions on never-seen series), error bars, and a clearer statement of the length generalization. My recommendation: send to peer review with a request for major revision before acceptance.","headline":"T2S brings a useful new dataset and a clean variable-length latent design, but the current evaluation lets a model score well by memorizing its own captioning pipeline; the headline claims need a held-out caption test.","tokens_in":15293,"tokens_out":3196,"would_cite":true,"duration_ms":34956,"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":"A single diffusion model can generate semantically aligned time series of arbitrary length from natural-language captions.","keywords":["text-to-time-series generation","diffusion models","flow matching","diffusion transformer","length-adaptive variational autoencoder","time series captions","TSFragment-600K","arbitrary-length generation"],"falsifier":"Run T2S on a held-out set of captions written by humans for time series the model has never seen, and compare its WAPE, MSE, and MRR@10 with the published TSFragment-600K numbers; a large drop would show the model is tied to its own captioning pipeline rather than to general natural-language understanding.","tokens_in":14296,"feed_emoji":"📈","tokens_out":8804,"duration_ms":94102,"temperature":0.7,"pith_summary":"This paper argues that text-to-time-series generation can be moved out of single-domain settings. It introduces TSFragment-600K, a dataset of more than 600,000 fragment-level caption-series pairs, and T2S, a diffusion-based model that encodes variable-length series into a common latent space and conditions generation on natural-language captions using flow matching and a diffusion transformer. The authors report that T2S outperforms trained diffusion, variational-autoencoder, and zero-shot large-language-model baselines across 13 datasets from 12 domains, and that by training interleaved across lengths 24, 48, and 96 it can generate series of arbitrary length in that range from a single model. If these results hold, text prompts become a practical way to synthesize realistic temporal data in domains where data are scarce, imbalanced, or hard to collect.","feed_headline":"One model turns text prompts into time series of any length","feed_subtitle":"A diffusion model trained at multiple lengths claims to follow natural-language captions across 12 domains.","key_machinery":"The load-bearing machinery is the combination of four pieces. The first is TSFragment-600K, a dataset built by cutting series into fragments, asking a language model to caption them with expert-curated seed prompts, and selecting among five candidate captions by embedding cosine similarity. The second is LA-VAE, a variational autoencoder that upsamples variable-length latent codes to a fixed size and downsamples them back, with a consistency loss on the latent code to reduce blur. The third is flow matching, a diffusion variant in which the denoiser predicts the straight-line velocity from noise to data along optimal-transport paths, integrated with a diffusion transformer (T2S-DiT) that injects caption information through adaptive layer normalization. The fourth is interleaved training, which shuffles samples of different lengths together so one model learns all lengths without catastrophic forgetting.","core_discovery":"The paper's central claim is that text and time series can be aligned at the fragment level by one domain-agnostic generative model. T2S first maps a variable-length series into a fixed-size latent embedding with a length-adaptive variational autoencoder (LA-VAE), then runs a diffusion transformer (T2S-DiT) trained with flow matching to denoise that embedding under a natural-language caption. The authors build TSFragment-600K, more than 600,000 fragment-level caption-series pairs drawn from eight classical time series datasets, and train T2S interleaved across lengths 24, 48, and 96 so that inference can produce any length in that range. They report that T2S obtains the best WAPE, MSE, and MRR@10 scores among trained diffusion, VAE, and zero-shot large-language-model baselines on 13 datasets spanning 12 domains, including point-level and instance-level captions as well as fragment-level ones.","pith_inferences":["The evaluation captions were generated from the same series used for scoring, so the strongest version of the claim is that T2S reproduces series from captions produced by its own dataset pipeline; a test with human-written captions on held-out series would separate language understanding from pipeline inversion.","The paper only demonstrates arbitrary-length generation inside the 24-96 range used in interleaved training; extrapolation to drastically longer horizons or to lengths not aligned with the trained multiples remains untested.","The model is evaluated on univariate series, so extending the same latent-space design to multivariate, multi-channel series would be a natural next test of the domain-agnostic claim.","A retrieval baseline that simply matches existing series to caption embeddings would clarify whether T2S truly generates new temporal structure or reproduces memorized fragments."],"forward_implications":["If T2S works as reported, a single model can generate series of any length within its trained range, so practitioners no longer need to train separate generators for each fixed sequence length.","The TSFragment-600K captioning pipeline gives other domains a recipe for building fragment-level text-series data, which is the missing ingredient for text-conditioned generation outside specialized fields.","Because T2S handles point-, fragment-, and instance-level captions, one model can serve tasks from fine-grained event descriptions to coarse global summaries.","The reported gains over diffusion- and LLM-based baselines suggest that the flow-matching and DiT design, not the caption size alone, drives generation quality across granularities."],"supporting_citations":[{"why":"Provides the rectified-flow training objective that T2S adopts as its diffusion backbone.","marker":"[Liu et al., 2022]"},{"why":"Supplies the flow-matching framework for conditional generation used by T2S.","marker":"[Lipman et al., 2022]"},{"why":"Defines the diffusion transformer architecture that T2S-DiT adapts for time series.","marker":"[Peebles and Xie, 2023]"},{"why":"Demonstrates scaling rectified-flow transformers for high-resolution image synthesis, the design T2S transfers to time series.","marker":"[Esser et al., 2024]"},{"why":"DiffusionTS is the primary trained text-conditioned time series baseline that T2S must outperform.","marker":"[Yuan and Qiao, 2024]"},{"why":"TimeVAE is the variational-autoencoder baseline for text-conditioned time series generation.","marker":"[Desai et al., 2021]"},{"why":"Time-MMD supplies the point-level caption dataset used in evaluation.","marker":"[Liu et al., 2024a]"},{"why":"SUSHI supplies the instance-level caption dataset used in evaluation.","marker":"[Kawaguchi et al., 2025]"},{"why":"GPT-4o-mini is used to generate TSFragment-600K captions and is also evaluated as a zero-shot baseline.","marker":"[OpenAI, 2023]"},{"why":"Provides the source time series datasets from which TSFragment-600K fragments are drawn.","marker":"[Wu et al., 2021]"}],"fun_headline_variants":["Text-to-time series, any length you need, from one diffusion model","T2S turns captions into time series of arbitrary length","From text to time series: T2S adapts to any length","Diffusion model maps text to time series at any length","Prompt a time series, T2S delivers any length"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The benchmark assumes the evaluation captions test genuine language-to-series understanding, but the captions were produced from the same samples the model is later scored on, so the test may only measure how well the model inverts its own captioning pipeline.","fun_headline_variants_meta":{"raw":{"variants":["Text-to-time series, any length you need, from one diffusion model","T2S turns captions into time series of arbitrary length","From text to time series: T2S adapts to any length","Diffusion model maps text to time series at any length","Prompt a time series, T2S delivers any length"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000221,"raw_usage":{"total_tokens":1484,"prompt_tokens":1012,"completion_tokens":472,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":628,"completion_tokens_details":{"reasoning_tokens":384}},"tokens_in":628,"tokens_out":472,"duration_ms":5775,"temperature":1.0,"reasoning_tokens":384,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T00:51:14.535769+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run T2S on a held-out set of captions written by humans for time series the model has never seen, and compare its WAPE, MSE, and MRR@10 with the published TSFragment-600K numbers; a large drop would show the model is tied to its own captioning pipeline rather than to general natural-language understanding.","supporting_citations":[{"cited_title":"Scaling rectified flow transformers for high- resolution image synthesis","cited_arxiv_id":null,"evidence_quote":"Demonstrates scaling rectified-flow transformers for high-resolution image synthesis, the design T2S transfers to time series."},{"cited_title":"Sushi: A system for unified semantic human in- teraction","cited_arxiv_id":null,"evidence_quote":"SUSHI supplies the instance-level caption dataset used in evaluation."},{"cited_title":"Gpt-4o mini,","cited_arxiv_id":null,"evidence_quote":"GPT-4o-mini is used to generate TSFragment-600K captions and is also evaluated as a zero-shot baseline."}],"review_version":1}