{"id":"3e8ad19a-5185-422f-bbb4-80b9cb6ebd87","arxiv_id":"2501.05403","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"TimeDP conditions a diffusion model on prototype assignments extracted from a few example series, achieving strong in-domain and few-shot unseen-domain time series generation.","lead":"The paper introduces TimeDP, a diffusion model that generates time series across many domains by using a few example sequences as 'domain prompts', without needing domain labels or text descriptions. It reports state-of-the-art generation quality on 12 benchmark datasets and shows few-shot generation on two unseen datasets.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Unseen-domain capability rests on an unvalidated assumption that a frozen random prototype basis plus a source-trained assignment network can prompt any new domain; only two near-source datasets are tested.","rationale":"The in-domain results are credible: TimeDP wins most of the twelve in-domain MMD/K-L/MDD comparisons across sequence lengths, and the ablations show the conditioning mechanism matters. So I would not reject the paper. The load-bearing weakness is the scope of the unseen-domain claim. The current experiment uses only Stock and Web; Stock is a daily financial series similar to Exchange, and Web is a daily count series similar to NN5. The paper provides no mechanism showing the frozen random prototypes form a universal basis, and PAM is never trained on data outside the source set. This matches the Reader's weakest_assumption, so I agree with the CONDITIONAL verdict. My proposed leave-one-domain-out and basis-orthogonal synthetic check would directly test whether the few-shot prompt mechanism transfers to domains whose structure is not near the training set. If it fails, the abstract's broad claim should be weakened to 'strong capability on domains similar to training domains.' I also flag Algorithm 1's undefined p_u and the random reset of P as a reproducibility gap consistent with the Reader's concern.","tokens_in":23875,"tokens_out":8423,"duration_ms":88071,"concrete_test":"Leave out the entire economic domain (NN5, Fred-MD, Exchange) from training and retrain TimeDP on the remaining nine datasets; then evaluate few-shot generation on the held-out economic datasets and on one synthetic domain whose spectrum is deliberately orthogonal to the frozen prototypes (e.g., square waves or high-frequency chirps) at K=3, 10, and 100, against the unconditional and fine-tuned baselines. If TimeDP's MMD and K-L on held-out domains are no better than its unconditional variant, or fail to improve with K, then the 'strong unseen-domain generation' claim is an artifact of similarity between Stock/Web and the training domains rather than a general prompt mechanism.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim includes 'strong unseen domain generation capability,' and the Method section states that prototypes 'are not restricted to the domains in training sets' and can represent any unseen dataset D_j from K few-shot samples. This requires two things that are never established: (1) a fixed, randomly initialized and frozen set of 16 prototype vectors can express the temporal structure of an arbitrary new domain, and (2) the PAM weight extractor, trained only on the 12 source datasets, maps few-shot samples from a genuinely novel domain to a conditioning vector the denoising U-Net has learned to follow. No training signal enforces either property; the only evidence is Table 2 on Stock and Web, two domains that are plausibly close to Exchange and NN5 in the training set. If an unseen domain's patterns are not spanned by the frozen basis, or its assignments fall outside the source-domain condition manifold, generation should degrade toward the unconditional output. The claim 'for any unseen dataset' is therefore broader than the data support. A secondary reproducibility gap is that Algorithm 1's 'unconditional identifier p_u' is never defined, so the exact training objective is not fully specified in the text.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"TimeDP is a multi-domain time series diffusion model that conditions generation on a \"domain prompt\" derived from a set of randomly initialized and then frozen prototype vectors. The Prototype Assignment Module (PAM) maps each input sequence to a sparse weight vector over the prototypes; the weights are added to the cross-attention logits of a U-Net denoiser. The model is trained jointly on 12 datasets with balanced sampling; at inference, prompts are extracted from a few-shot sample of the target domain, including domains unseen in training. The paper reports in-domain generation results on 12 datasets using MMD, K-L, and MDD metrics, plus few-shot unseen-domain results on Stock and Web, and ablations on prototype count and component removal.","tokens_in":24019,"tokens_out":8590,"duration_ms":82182,"significance":"TimeDP addresses a real gap: multi-domain time series generation without explicit class labels or text descriptions. The core idea of using a small set of basis vectors and sparse assignments as a conditioning interface is simple and potentially useful. The in-domain benchmark is reasonably strong: the proposed model achieves the best MMD and K-L scores on most of the 12 datasets against established baselines, and the ablation shows that the conditioning mechanism gives a large improvement in MMD over the same model without prompts. The release of code and the use of five seeds are commendable. However, the paper's broader claims—that the prototypes \"learn\" time-series semantics and that the method generalizes to \"any unseen dataset\"—are not supported as stated, because the prototype vectors are frozen and the unseen-domain evidence is limited to two near-source datasets.","major_comments":[{"comment":"","section":"Semantic Prototype Module; Abstract; Conclusion"},{"comment":"","section":"Algorithm 1 (Training algorithm), line 5"},{"comment":"","section":"Unseen Domain Generation; Table 2"},{"comment":"","section":"Table 3 (Ablation Study)"}],"minor_comments":[{"comment":"","section":"Algorithm 2 (Sampling with domain prompts)"},{"comment":"","section":"Equation (6)"},{"comment":"","section":"Contributions and Related Work"},{"comment":"","section":"Experiment Settings"},{"comment":"","section":"Table 2 and Appendix Figures 6/7"},{"comment":"","section":"Abstract and Background"}],"recommendation":"major_revision","confidential_remarks":"The paper is likely of interest to the AAAI audience because multi-domain time series generation without labels is a useful direction and the in-domain empirical results are strong. The main risks are (1) the missing definition of the unconditional identifier, which makes the training procedure not fully reproducible; and (2) the overclaimed unseen-domain generalization, supported only by two conveniently chosen near-source datasets. If the authors add a clearly specified training algorithm and either broaden the unseen-domain evaluation or temper the claim, the paper could become acceptable. I would also flag that the \"learned prototypes\" narrative should be reconciled with the frozen prototype implementation, as this affects how the contribution is perceived."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"TimeDP is a solid, useful method paper. The new thing is real: conditioning a diffusion U-Net on prototype-assignment weights extracted from a few real samples, with no domain labels and no text. That combination — frozen random prototype dictionary, learned assignment network, cross-attention conditioning — is not something I've seen in the cited prior work. The empirical work is also more thorough than the average generation paper: 12 datasets, four sequence lengths, 5 seeds, several baselines, and ablations. On the in-domain benchmarks, TimeDP beats the class-conditional TimeVQVAE-C on most datasets, and the ablation study shows both the PAM and the prompt conditioning matter. The code is available. That part holds up.\n\nThe soft spots are mostly about framing and one missing definition. The \"unseen domain\" claim is broader than the evidence. Only two held-out datasets are tested, Stock and Web, and both are plausibly close to Exchange and NN5 in the training set. The paper asserts that the frozen random prototypes can represent any new domain and that the source-trained assignment network will map few-shot samples to useful prompts, but there is no mechanism guaranteeing that. If a genuinely novel domain produced assignments outside the training manifold, generation would degrade toward unconditional output. A claim like \"strong unseen-domain capability on two held-out datasets\" would be accurate; \"for any unseen dataset\" is not supported.\n\nSecond, Algorithm 1 has an undefined term: \"Randomly set P as unconditional identifier p_u.\" I could guess it means randomly dropping the condition, but the exact training objective is not fully specified in the text. That is a reproducibility gap.\n\nThird, the paper repeatedly calls the prototypes \"learned\" and refers to \"time series semantic prototypes,\" but the prototype vectors are randomly initialized and frozen. What is actually learned is the assignment network and the attention projections. The visualization in Figure 2 is nice and shows those projections can extract meaningful structure, but the paper should be precise about what is learned versus fixed. This is a wording problem, not a fatal flaw.\n\nMinor: no commit hash or environment specification in the code link, so independent verification takes some effort.\n\nWho is this for? People working on time series generation, data augmentation, or privacy-preserving release of time series will get value from it. It is a practical method paper, not a field-rearranging one. It deserves a serious referee. I would send it to peer review with a request for revision: define p_u, soften the unseen-domain claim to match the evidence, and fix the \"learned prototypes\" language.","headline":"A useful multi-domain time series diffusion method with credible in-domain results, but the unseen-domain generalization claim runs ahead of the evidence.","tokens_in":24628,"tokens_out":1998,"would_cite":true,"duration_ms":21927,"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, conditioned on few-shot 'domain prompts' built from a frozen random dictionary of time-series prototypes, generates high-quality data for trained domains and adapts to unseen domains without fine-tuning.","keywords":["multi-domain time series generation","diffusion models","domain prompts","prototype assignment","few-shot generation","unseen domain generation","time series prototypes","conditional generation"],"falsifier":"Hold out a domain whose time series contain frequency components absent from the twelve training datasets—for example, a synthetic chirp with a time-varying frequency, or a series of aperiodic spikes. Give TimeDP 10 or 100 samples from that domain and compute the MMD and K-L divergence between its generated samples and the true test set, alongside the same metrics for the model's own unconditional output. If the prompted generation is not closer to the target distribution than the unconditional output, or if increasing the sample count from 3 to 100 does not improve fidelity, the claim that frozen random prototypes can express unseen domains is falsified.","tokens_in":23565,"feed_emoji":"📈","tokens_out":8765,"duration_ms":74357,"temperature":0.7,"pith_summary":"The paper proposes TimeDP, a single diffusion model that generates time series for many domains at once without domain labels or text descriptions. Instead of conditioning on a class or a natural-language prompt, TimeDP learns to construct a 'domain prompt' from a few example sequences of the target domain: a small network assigns weights to a fixed set of 16 random orthogonal prototype vectors, and those weights steer the denoising process through cross-attention. The paper claims this suffices for best-in-class in-domain generation on 12 real-world datasets and, more strikingly, for few-shot generation on domains never seen in training (Stock and Web), outperforming both unconditional baselines and baselines fine-tuned on the same few shots. If true, this offers a label-free and text-free route to multi-domain and zero-shot time-series synthesis, which matters for data augmentation, privacy-preserving synthetic data, and financial or medical scenario simulation.","feed_headline":"Few-shot prompts let one model generate time series in unseen domains","feed_subtitle":"A learned prompt built from a few samples reproduces a domain's distribution better than fine-tuning.","key_machinery":"The load-bearing objects are the time-series semantic prototypes and the prototype assignment module (PAM). Prototypes are $N_p=16$ vectors $P \\in \\mathbb{R}^{N_p \\times d}$, initialized as random orthogonal vectors and frozen; each vector is meant to encode an elementary time-series feature such as trend or seasonality, acting like a 'word' in a shared dictionary. PAM is a small network $\\varphi$ that maps an input sequence $x$ to a weight vector $m = \\varphi(x)$, with negative entries set to zero for sparsity. The weights $m$ are added to the scaled dot-product attention scores in each cross-attention layer of the denoising U-Net, so the predicted noise $\\hat{\\epsilon} = \\epsilon_{\\theta,P}(x_n, n, m)$ is conditioned on the assigned prototypes. At sampling time, the same PAM is applied to $K$ few-shot samples of the target domain to form the domain prompt $m_{D_i}=\\{m^i_1,\\dots,m^i_K\\}$, and the model generates by repeated denoising conditioned on these prompts.","core_discovery":"TimeDP's central claim is that a small set of frozen random orthogonal vectors can act as a universal time-series 'dictionary,' and that a learned prototype assignment module can express any domain—trained or unseen—as a weighted combination of these vectors. These weights, called domain prompts, are fed into a U-Net diffusion model by biasing the cross-attention logits between the noisy latent and the prototype keys and values. During sampling, K few-shot samples from the target domain are mapped through the assignment network to obtain a prompt population, and the diffusion model generates new sequences conditioned on it. On 12 datasets across energy, transport, nature, and economic domains, TimeDP reports the best Maximum Mean Discrepancy and K-L divergence on most datasets, beating a class-conditioned vector-quantized baseline without using labels; on the unseen Stock and Web datasets it reports the best MMD, K-L, and marginal-distribution scores with 3, 10, or 100 prompt samples, improving as the few-shot count grows, while fine-tuned baselines do not consistently improve.","pith_inferences":["Not explored in the paper: learning the prototype vectors during training instead of freezing them; if random orthogonal vectors already work, a learned, updated basis might cover a wider range of temporal patterns and improve unseen-domain transfer further.","The prompting mechanism is modality-agnostic in principle: any signal that can be decomposed into a small set of elementary patterns (images, audio, spatiotemporal data) could reuse the same few-shot prompt-extraction idea.","The ablation shows that removing PAM or the prompt barely changes K-L divergence but sharply worsens MMD, suggesting the prompt mainly preserves sequential dependence rather than marginals; a direct check of autocorrelation or spectral fidelity of generated sequences would test this interpretation.","The paper's unseen-domain experiments use only two target datasets; a stronger probe would be to hold out entire categories (for instance, all economic domains) and test whether prompts still separate them, which would clarify whether the frozen basis is truly universal or merely sufficient for these data."],"forward_implications":["One trained TimeDP model can replace per-dataset generators: the same weights serve every domain, and the only per-domain input is a few dozen sample sequences.","Unseen-domain generation needs no fine-tuning: a handful of samples defines the prompt, and quality improves as the number of samples grows from 3 to 100.","Domain labels and natural-language descriptions become unnecessary for conditional time-series generation, removing a bottleneck for new or evolving domains whose patterns are hard to describe.","The prototype dictionary provides an interpretable control axis: prompting with one-hot weights yields distinct seasonal or trend patterns, so a user can steer generation in a human-readable way.","Because the prompt is a distribution over a shared basis, domains are explicitly related rather than treated as independent, which the paper argues helps the model capture inter-domain similarity."],"supporting_citations":[{"why":"Supplies the denoising diffusion probabilistic model backbone and the simplified noise-prediction objective the method trains with.","marker":"Ho, Jain, and Abbeel 2020"},{"why":"Foundational diffusion formulation the paper builds on for the forward and reverse Markov process.","marker":"Sohl-Dickstein et al. 2015"},{"why":"Provides the conditional DDPM form used to inject the domain prompt as conditioning term c.","marker":"Ho and Salimans 2021"},{"why":"Motivates the time-series 'basis' idea that the semantic prototype module implements as a shared dictionary.","marker":"Ni et al. 2023"},{"why":"Defines the multi-domain unified-training setting and channel-independent preprocessing the paper adopts.","marker":"Woo et al. 2024"},{"why":"Supports the channel-independent uni-variate formulation used to handle heterogeneous time series.","marker":"Nie et al. 2023"},{"why":"The TimeVQVAE baseline, including its class-conditioned variant TimeVQVAE-C, which TimeDP compares against as a label-based conditional generator.","marker":"Lee, Malacarne, and Aune 2023"},{"why":"The TimeGAN baseline that TimeDP outperforms on in-domain and unseen-domain metrics.","marker":"Yoon, Jarrett, and van der Schaar 2019"},{"why":"The GT-GAN baseline combining GANs, flows and ODEs used in the comparison.","marker":"Jeon et al. 2022"},{"why":"The TimeVAE baseline with a trend/seasonality decoder structure used in the comparison.","marker":"Desai et al. 2021"}],"fun_headline_variants":["One diffusion model, many domains: few-shot prompts do the trick","TimeDP: few-shot prompts generate time series in unseen domains","Universal time-series dictionary enables few-shot domain generation","Prompt from few samples adapts diffusion model to new domains"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that 16 fixed, randomly chosen and frozen vectors can represent the temporal structure of any domain, including domains never seen in training, and that the assignment network trained only on source domains will map a few samples from any new domain to a meaningful prompt; if a new domain's patterns cannot be expressed by those vectors, the prompt carries no useful information and generation falls back to unconditional output.","fun_headline_variants_meta":{"raw":{"variants":["One diffusion model, many domains: few-shot prompts do the trick","TimeDP: few-shot prompts generate time series in unseen domains","Universal time-series dictionary enables few-shot domain generation","Prompt from few samples adapts diffusion model to new domains"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000226,"raw_usage":{"total_tokens":1469,"prompt_tokens":944,"completion_tokens":525,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":560,"completion_tokens_details":{"reasoning_tokens":457}},"tokens_in":560,"tokens_out":525,"duration_ms":5060,"temperature":1.0,"reasoning_tokens":457,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:14:31.176393+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Hold out a domain whose time series contain frequency components absent from the twelve training datasets—for example, a synthetic chirp with a time-varying frequency, or a series of aperiodic spikes. Give TimeDP 10 or 100 samples from that domain and compute the MMD and K-L divergence between its generated samples and the true test set, alongside the same metrics for the model's own unconditional output. If the prompted generation is not closer to the target distribution than the unconditional output, or if increasing the sample count from 3 to 100 does not improve fidelity, the claim that frozen random prototypes can express unseen domains is falsified.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The GT-GAN baseline combining GANs, flows and ODEs used in the comparison."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the multi-domain unified-training setting and channel-independent preprocessing the paper adopts."},{"cited_title":"H.; Sinthong, P.; and Kalagnanam, J","cited_arxiv_id":null,"evidence_quote":"Supports the channel-independent uni-variate formulation used to handle heterogeneous time series."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The TimeVQVAE baseline, including its class-conditioned variant TimeVQVAE-C, which TimeDP compares against as a label-based conditional generator."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The TimeGAN baseline that TimeDP outperforms on in-domain and unseen-domain metrics."}],"review_version":1}