{"id":"6f0fc8f0-bd6c-40fd-b2f6-068241c7d42b","arxiv_id":"2506.20167","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"SEED is an LLM-based forecasting architecture that couples variable-wise attention encoding with prototype-based semantic reprogramming and a frozen decoder.","lead":"The paper introduces SEED, a model that combines a structural encoder for multivariate time series with a frozen large language model to forecast future values. It reports lower prediction error than several strong baselines on five of eight benchmark datasets.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table I refutes the reported Solar-Energy win: SEED's MSE is 0.243 vs iTransformer's 0.233, while the text claims a match at 0.233.","rationale":"The reader's weakest assumption focuses on whether the prototype-based semantic grounding carries the reported gains. I agree that this mechanism is unvalidated, but the single most load-bearing issue is more direct: the paper's empirical headline is contradicted by its own table. The claimed best result on Solar-Energy is the clearest example, and the missing protocol and error bars mean the comparison cannot currently be verified. Because this can be fixed by correcting the table, specifying the protocol, and adding ablations, the existing CONDITIONAL verdict remains appropriate; nothing in my read moves it to accept or reject.","tokens_in":8488,"tokens_out":6870,"duration_ms":82714,"concrete_test":"Recompute Table I's Solar-Energy row from the official test predictions under the paper's stated protocol, and also under the standard Time-LLM/iTransformer protocol with a fixed look-back and horizons 96, 192, 336, and 720. If the corrected SEED MSE is 0.243 while iTransformer is 0.233, the five-dataset claim is false and the headline should be reduced to four datasets.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest claim is the empirical one: SEED achieves best overall MSE and MAE on five datasets including Solar-Energy. The paper's own Table I does not support this. For Solar-Energy, SEED has MSE 0.243 and iTransformer has 0.233; the text later states 'SEED matches the best MSE score of 0.233', which is inconsistent with the table. Since the central claim explicitly lists Solar-Energy as a dataset where SEED wins on both metrics, the reported evidence is internally contradictory. Moreover, the experimental section gives no prediction horizon, no input length, no error bars, and no ablation that isolates the prototype-based semantic reprogramming; the comparison protocol is underspecified. The load-bearing condition for the central claim is that the empirical comparison is accurate and complete; the Solar-Energy row shows that it is not, and the missing protocol prevents independent verification.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SEED, a four-stage architecture for multivariate time series forecasting: a token-aware structural encoder that treats each variable as a token and applies attention over the variable dimension, a patch projection and alignment module that maps the encoded representation into an LLM-compatible embedding space, a prototype- and prompt-based semantic reprogramming module, and a frozen autoregressive LLM decoder. The claimed contribution is a modular structural-semantic decoupling that yields improved or competitive MSE/MAE relative to five baselines on eight benchmark datasets. The empirical evidence is a single comparison table (Table I), with the text claiming best overall results on five datasets including Solar-Energy.","tokens_in":8687,"tokens_out":4534,"duration_ms":50917,"significance":"If the empirical claims were fully supported, SEED would be a plausible incremental contribution: it combines an iTransformer-style inverted encoder with Time-LLM-style reprogramming and a frozen LLM, and the four-stage decomposition is clearly described. The paper also evaluates on eight diverse datasets, which is a useful breadth. However, the current evidence is not sufficient to support the central claim. The Solar-Energy row of Table I directly contradicts the text, the experimental protocol omits forecast horizon, lookback length, and error bars, and no ablation isolates the prototype-based semantic mechanism. These are load-bearing issues because the paper's contribution is an architecture whose claimed advantage rests on those empirical results.","major_comments":[{"comment":"The central empirical claim is internally inconsistent for Solar-Energy. Table I reports SEED's MSE as 0.243 and MAE as 0.263, while iTransformer reports MSE 0.233 and MAE 0.262; the text states that 'SEED matches the best MSE score of 0.233' and earlier says both models achieve 'approximately 0.24'. These statements cannot all be true. Since Solar-Energy is explicitly listed among the five datasets where SEED is claimed to achieve the best overall MSE and MAE, the table as printed refutes the headline claim. In addition, the phrase 'consistent improvements' is contradicted by the same table: SEED loses to Time-LLM on ETTh1 (0.413 vs 0.408) and Traffic (0.388 vs 0.390) in MSE. Please correct the table or the text, and re-check every row for the same type of discrepancy.","section":"Section III, Table I"},{"comment":"The experimental protocol is underspecified. The paper never states the forecast horizon H or the input lookback length L for the reported results, nor does it report the number of random seeds or any variance measure. The table gives single point estimates, so differences such as ETTh2 MSE 0.330 versus Time-LLM 0.334 or Traffic MAE 0.264 versus Time-LLM 0.263 are not interpretable without error bars or significance tests. To make the comparison reproducible and the claimed gains meaningful, the authors must report the exact values of H and L, the number of runs, and mean plus/minus standard deviation, together with an appropriate statistical test for pairwise differences.","section":"Section III"},{"comment":"The semantic reprogramming module is load-bearing for the paper's stated contribution, but no ablation isolates its effect. The claim that learnable prototypes and the attention-based convex combination in Eq. (6) provide 'semantic grounding' is not tested. An ablation that replaces Eq. (6) with a plain linear projection of the same dimension, or uses fixed random prototypes, would show whether any gains come from the prototype mechanism itself or simply from the additional trainable parameters. Without such an ablation, the role of the semantic reprogramming stage remains unsupported.","section":"Section II-C and Section III"},{"comment":"The statement after Eq. (6) that the convex combination \\(\\tilde z_i\\) 'preserves the original temporal pattern of \\(z_i\\)' is not justified. The output is a convex combination of prototype vectors whose weights are derived from \\(z_i\\), so the result lies in the prototype span; no reconstruction error, proximity constraint, or formal argument is given that would preserve temporal structure. This is a conceptual claim that needs either a derivation or an empirical verification, and it underlies the paper's narrative that numerical patterns are made 'semantically grounded'.","section":"Section II-C"}],"minor_comments":[{"comment":"In the ETTm2 row, '0.2850.333' appears to be a typo for '0.285 0.333'; please fix the missing separator.","section":"Table I"},{"comment":"The symbol P is overloaded: in Eq. (3) it denotes the patch length, while in Section II-C it is reused for the set of prototypes \\(\\mathcal P\\). Please use distinct notations.","section":"Section II-C"},{"comment":"The token matrix T is defined as \\(\\mathbb R^{N\\times d}\\) in Section II-A but as \\(\\mathbb R^{N\\times D\\times L}\\) in Section II-B; the relationship between these two notations is not explained and should be clarified.","section":"Sections II-A and II-B"},{"comment":"The symbol N is used inconsistently: it denotes the number of variables in Section II-A, the number of patches in Section II-C and II-D, while Section II-B defines the number of patches as M. Please unify the notation.","section":"Section II-D"}],"recommendation":"major_revision","confidential_remarks":"The architecture is coherent and the central problem is correctable, so I recommend major revision rather than rejection. However, I would not accept the paper until the Solar-Energy numbers are reconciled, the protocol (horizon, lookback, seeds, error bars) is fully specified, and the semantic reprogramming module is ablated. If the corrected table shows that SEED does not actually win on five datasets, the paper's central claim will need to be substantially rewritten."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, SEED is a clean modular recipe that stacks iTransformer-style variable attention, patch projection, prototype-based reprogramming, and a frozen LLM. That is a legitimate combination, and the write-up is easy to follow. Second, the headline empirical claim does not hold up as printed: Table I shows SEED's Solar-Energy MSE at 0.243 while iTransformer gets 0.233, and the text says SEED \"matches the best MSE score of 0.233.\" That is a direct contradiction in the paper's own evidence, and Solar-Energy is one of the five datasets in the claimed win list. This has to be fixed before the results are usable.\n\nWhat the paper does well: the modular decomposition is genuinely useful for thinking about structure/semantic interfaces. The authors are explicit that the encoder is initialized from iTransformer and the decoder is a frozen LLaMA/GPT, so there is no mystery about what is borrowed. The prototype-as-semantic-anchor idea is a modest but reasonable extension of Time-LLM's reprogramming, and the formalism in Section II is readable. If the code and full protocol are released, a reimplementation would be straightforward.\n\nWhere it falls short: the evaluation is the soft spot, and it is load-bearing. No error bars, no significance tests, no ablations. The paper attributes gains to the semantic reprogramming module but never isolates it. The setup omits the forecast horizon, input length, per-dataset hyperparameters, and even the LLM variant used for the main table (only \"frozen LLaMA-2 or GPT-2\" is mentioned). The \"consistent improvements\" phrasing is also contradicted by the table: SEED loses to Time-LLM on ETTh1 and Traffic. None of these are fatal to the architecture, but they mean the empirical contribution is currently unsupported as written.\n\nBottom line: the architecture is sensible and the writing is clear, but the evidence needs heavy revision. This paper belongs in a workshop or a revision cycle, not in final form. I would send it to a serious referee with a request for code, ablations, and a corrected table; if the authors fix the Solar-Energy numbers and add an ablation that isolates the prototype mechanism, the result could be a useful incremental contribution to the LLM-time-series subfield.","headline":"A sensible modular recipe, but the headline empirical claim is internally contradicted by its own table on Solar-Energy; needs a corrected table and better protocol before the results can be trusted.","tokens_in":9171,"tokens_out":2421,"would_cite":false,"duration_ms":23625,"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":"SEED couples a variable-wise structural encoder with prototype-based semantic tokens and a frozen LLM, and the paper reports best MSE and MAE on five of eight benchmark datasets.","keywords":["multivariate time series forecasting","large language models","structural encoding","semantic reprogramming","prototype-based attention","frozen LLM decoder","patch projection","inverted attention"],"falsifier":"Run SEED on the claimed datasets with the prototype module removed (or with prototypes frozen to random vectors) and compare MSE and MAE against the full model. If the full model does not beat the prototype-free version on ETTh2, ETTm1, Weather, ECL, and Solar-Energy, then the semantic reprogramming mechanism is not doing the work claimed. A second check is to sweep the number of prototypes K; flat performance across K would also indicate the prototypes are not functioning as semantic anchors.","tokens_in":8322,"feed_emoji":"📈","tokens_out":6472,"duration_ms":64457,"temperature":0.7,"pith_summary":"The paper tries to establish that multivariate time series forecasting improves when structural encoding is separated from semantic inference: a variable-wise attention encoder first models cross-variable dependencies, and a frozen large language model then reads prototype-based semantic tokens to produce forecasts. The reported evidence is that SEED obtains the best MSE and MAE on five of eight benchmark datasets (ETTh2, ETTm1, Weather, ECL, and Solar-Energy) when compared with Time-LLM, iTransformer, Autoformer, Informer, and Reformer. If true, this means a single frozen LLM can serve as a task-agnostic decoder across heterogeneous domains while all learning happens upstream, which would lower the cost of adapting forecasting systems to new tasks. The paper's central bet is that learnable prototypes genuinely ground numerical tokens in a semantically useful space; this grounding is not isolated by an ablation.","feed_headline":"Frozen LLM plus structural encoder tops five forecasting benchmarks","feed_subtitle":"SEED reports best MSE and MAE on ETTh2, ETTm1, Weather, ECL, and Solar-Energy, beating Time-LLM and iTransformer.","key_machinery":"The load-bearing mechanism is a four-stage pipeline. A token-aware structural encoder, based on the inverted attention idea from iTransformer, transposes the multivariate input so that each variable is a token and attention runs over the variable dimension; a temporal projection collapses each variable's trajectory to a latent vector. A patch projection module then reshapes the encoded tensor into non-overlapping patches along time, flattens each patch, and linearly projects it to the LLM embedding dimension with a learned positional embedding. A semantic reprogramming module compares each projected patch token with K learnable prototypes using scaled dot-product attention, replaces the token with a convex combination of prototypes, and prepends task prompt embeddings. Finally, a frozen autoregressive LLM (GPT-2 or LLaMA-2) consumes the sequence and a learned linear head maps its output token back to the value space. The prototypes are the component intended to supply semantics, which is why the paper calls them 'semantic anchors'.","core_discovery":"The central claim is that the structural-semantic gap in time series prediction can be closed by a modular pipeline in which the LLM never sees raw numbers and never gets fine-tuned. SEED's variable tokens, produced by inverted attention over variables, are patched, projected into the LLM embedding space, and then replaced by convex combinations of K learnable prototypes before being fed, with task prompts, to a frozen autoregressive LLM. On the paper's reported numbers, this pipeline achieves the best overall MSE and MAE on ETTh2 (0.330/0.379), ETTm1 (0.329/0.372), Weather (0.225/0.257), ECL (0.158/0.252), and Solar-Energy (0.243/0.263, with iTransformer matching its MSE at about 0.24), and it trails Time-LLM on ETTh1 and Traffic in MSE while keeping the lowest MAE on Traffic. The conclusion the authors draw is that decoupling structural representation learning from semantic inference is a viable route to a unified, transferable forecasting system.","pith_inferences":["An ablation the paper does not report would settle the central bet: remove the prototype module (or replace prototypes with random fixed vectors) and keep everything else fixed. If accuracy on the five claimed datasets does not drop, the semantic-grounding story is not the cause of the gains.","The reported margins over Time-LLM are small on ETTh2 (0.330 vs 0.334 MSE), and the paper trains with a single fixed seed, so multi-seed runs with confidence intervals would show whether the ranking is stable enough to guide architecture choices.","Because the prototype layer is a convex dictionary of K vectors, one can ask whether the learned prototypes align with interpretable temporal patterns (trend, seasonality, spikes) by clustering them or by inspecting which prototypes receive high attention per dataset; the paper does not analyze this.","A direct extension would be to feed the same upstream encoder into multiple different frozen LLMs (e.g., GPT-2 vs LLaMA-2) and check whether the reported gains persist, which would test the claim that the framework is agnostic to the decoder."],"forward_implications":["A frozen LLM can serve as the decoder across several heterogeneous domains without fine-tuning, so adapting SEED to a new task reduces to retraining the upstream encoder and prototypes.","The same modular interface should transfer to other temporal tasks such as imputation and anomaly detection by swapping the task prompt and the output head.","Variable-wise attention plus LLM decoding yields the largest gains on the high-dimensional datasets (Weather, ECL) and on the ETT family, suggesting the design addresses inter-variable structure more than long-range temporal memory.","On Traffic and ETTh1, where SEED trails Time-LLM in MSE but still leads in MAE, the system appears more robust to outlier-heavy signals, a property the authors attribute to the frozen decoder.","Since the LLM is frozen, the framework's predictive capacity is bounded by the encoder, projection, and prototype layers; improving those should directly improve forecasts without any language-model retraining."],"supporting_citations":[{"why":"Supplies the inverted variable-wise attention mechanism that SEED's token-aware encoder is built on and provides the backbone initialization.","marker":"[16]"},{"why":"Supplies the prompt-and-reprogramming strategy that SEED's semantic reprogramming module adapts, and is the primary LLM-based baseline in the comparison.","marker":"[18]"},{"why":"Establishes patch-based tokenization for time series, which SEED's patch projection extends to multivariate structural embeddings.","marker":"[15]"},{"why":"One of the frozen autoregressive decoders SEED uses for prediction.","marker":"[20]"},{"why":"The other frozen autoregressive decoder used in the experiments.","marker":"[21]"},{"why":"A sparse-attention transformer baseline that SEED is compared against in the benchmark table.","marker":"[13]"},{"why":"A decomposition-based transformer baseline that SEED is compared against in the benchmark table.","marker":"[14]"}],"fun_headline_variants":["SEED: Frozen LLM plus structural encoder tops five benchmarks","Structural encoder + frozen LLM beats baselines on five datasets","SEED closes structural-semantic gap for time series forecasting","Frozen LLM with structural encoder wins on five forecasting tasks","SEED: Decoupling structure and semantics for time series prediction"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that taking convex combinations of learnable prototypes turns numerical patch embeddings into tokens that a frozen language model can reason about; if that semantic grounding contributes nothing beyond extra trainable parameters, the reported gains could come from the structural encoder and training setup alone.","fun_headline_variants_meta":{"raw":{"variants":["SEED: Frozen LLM plus structural encoder tops five benchmarks","Structural encoder + frozen LLM beats baselines on five datasets","SEED closes structural-semantic gap for time series forecasting","Frozen LLM with structural encoder wins on five forecasting tasks","SEED: Decoupling structure and semantics for time series prediction"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.0002,"raw_usage":{"total_tokens":1380,"prompt_tokens":958,"completion_tokens":422,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":574,"completion_tokens_details":{"reasoning_tokens":337}},"tokens_in":574,"tokens_out":422,"duration_ms":3952,"temperature":1.0,"reasoning_tokens":337,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T22:54:40.225604+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run SEED on the claimed datasets with the prototype module removed (or with prototypes frozen to random vectors) and compare MSE and MAE against the full model. If the full model does not beat the prototype-free version on ETTh2, ETTm1, Weather, ECL, and Solar-Energy, then the semantic reprogramming mechanism is not doing the work claimed. A second check is to sweep the number of prototypes K; flat performance across K would also indicate the prototypes are not functioning as semantic anchors.","supporting_citations":[{"cited_title":"itrans- former: Inverted transformers are effective for time series forecasting,","cited_arxiv_id":null,"evidence_quote":"Supplies the inverted variable-wise attention mechanism that SEED's token-aware encoder is built on and provides the backbone initialization."},{"cited_title":"Time-LLM: Time series forecasting by reprogramming large language models,","cited_arxiv_id":null,"evidence_quote":"Supplies the prompt-and-reprogramming strategy that SEED's semantic reprogramming module adapts, and is the primary LLM-based baseline in the comparison."}],"review_version":1}