{"id":"6b0d4d52-4656-463c-a34f-b88e8eebc08a","arxiv_id":"2505.13175","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"SGCMA transfers an HMM state-transition prior learned from text into time series patches, then aligns patch embeddings to language tokens in each state, enabling a frozen GPT-2 to forecast as well as or better than tuned baselines.","lead":"This paper introduces SGCMA, a method that lets a frozen large language model forecast time series by first borrowing the hidden-state structure of language, learned with a Hidden Markov Model, and aligning temporal patches to those states. It reports state-of-the-art accuracy on long-term, short-term, few-shot, and zero-shot benchmarks without fine-tuning the LLM.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The transferability claim is untested because A_time is a learnable parameter initialized from A_text, not constrained to it; a random-initialization ablation is needed.","rationale":"The reader's weakest_assumption identifies the same core issue: transferability of the language-derived transition graph to time series is unverified. My read sharpens this with a concrete mechanism from the method itself. A_time is not a fixed structural prior; it is a learnable parameter initialized from A_text and trained end-to-end on the forecasting objective. With no proximity penalty, the optimization can freely adapt A_time to whatever structure helps forecasting, so even if the final A_time resembles A_text, this resemblance may reflect initialization inertia or convergence to a nearby optimum rather than a meaningful language-to-time-series transfer. The existing ablation in Table 5 removes structure alignment entirely, but it does not replace the language prior with a neutral control such as a random or identity transition matrix; therefore it cannot distinguish the contribution of the prior from the contribution of the extra trainable machinery. Figure 4 is also uninformative for transferability because HMM state labels are arbitrary and the L1 distance is computed after training; without a permutation-invariant comparison or a before-training measurement, the reported 0.046 distance does not demonstrate that the learned time-series states correspond semantically to language states. I do not claim the method is wrong. The reported forecasting gains may be real and reproducible. The problem is that the paper's stated explanation for those gains—that language-like structural dynamics are what activate the frozen LLM—is not supported by the experiments as reported. The proposed control experiment would settle this: if a random transition matrix achieves the same performance, then the language prior is not the active ingredient, and the central claim would need to be reframed. If the random-initialization baseline is clearly worse, the transferability claim would gain substantial support. Given no code, no error bars, and the missing control, the current CONDITIONAL verdict remains appropriate; my analysis does not move it to ACCEPT or REJECT, but it does identify the single experiment that should be run before the claim is accepted.","tokens_in":11798,"tokens_out":3435,"duration_ms":35780,"concrete_test":"Run SGCMA with the same hyperparameters on ETTh1, ETTm1, and Weather (H=96/192/336/720) under four conditions: (a) A_time initialized from A_text (original); (b) A_time initialized from a random/identity/uniform transition matrix; (c) A_time fixed to A_text (non-trainable); (d) A_time initialized from an HMM trained on the target time series itself or on a random walk. Report mean and standard deviation over three seeds. If (b) matches (a) within noise, the language prior is not load-bearing; if (c) also matches, the prior is unnecessary; if (d) outperforms (a), the language-specific origin is not special.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that sequence-level structural consistency with language, not token-level mapping, activates frozen LLMs. The decisive unverified link is Section 3.2: the language HMM's A_text and π_text are used only as initial values for learnable A_time and π_time in Eq. (8). There is no regularization keeping A_time near A_text, no permutation-invariant state correspondence, and no control experiment replacing the language prior with a random/identity/uniform transition matrix or with an HMM trained on unrelated time series. The Table 5 ablation removes Structure Alignment entirely or uses uniform state weights; it does not test whether the specific language-derived transition structure matters. Figure 4's L1 distance of 0.046 is computed after training and uses state indices that are arbitrary labels, so it cannot establish semantic transfer. Consequently, the reported gains may come entirely from the trainable prototypes, cross-attention, MEMM parameterization, and extra capacity, rather than from the language structure prior. This is a correctness risk for the headline claim, not just a missing baseline.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes SGCMA, a framework for time series forecasting with frozen LLMs. It trains an HMM on text to obtain a state-transition graph, uses it to initialize learnable transition parameters in a MEMM-like module that assigns state probabilities to temporal patches, aligns patch embeddings with top-k language tokens per state via cross-attention, and feeds the resulting embeddings into a frozen GPT-2 with a linear head. Experiments cover long-term, short-term (M4), few-shot, and zero-shot settings, with ablations and a sensitivity analysis over the number of latent states.","tokens_in":12044,"tokens_out":5917,"duration_ms":53665,"significance":"If the mechanism claimed were established, SGCMA would be a novel and potentially influential contribution: it is one of the few methods that align time series and language at the sequence-structure level rather than at the token-feature level, and it avoids LLM fine-tuning. The empirical scope is broad, and the idea of transferring an HMM-derived transition graph is interesting. However, the paper's central assertion—that language-derived structural priors are what drive the gains—is not actually tested, because the language parameters only initialize free learnable parameters. The paper also contains internal inconsistencies in the reported SOTA results and lacks error bars and reproducibility details. The contribution is promising but requires substantial additional evidence.","major_comments":[{"comment":"The transfer of the language structure is implemented only by initializing π_time and A_time from π_text and A_text; these are then trained without any regularization toward the text values, and no permutation-invariant correspondence between text and time states is defined. Therefore, the experiments do not establish that a language-derived transition structure is used at inference. To support the central claim, the authors should compare against (i) random initialization of A_time/π_time, (ii) a uniform or identity transition matrix, and (iii) an HMM trained on an unrelated time series or on shuffled text. If all variants perform equally, the paper's thesis is falsified by its own design.","section":"§3.2, Eq. (8)"},{"comment":"The L1 distance of 0.046 between A_time and A_text is computed after training on raw state indices, which are arbitrary labels; the same matrices can have large L1 distance under a permutation of state indices. Without Hungarian/permutation matching and without reporting the distance at initialization or a random-matrix baseline, this number cannot be interpreted as evidence of structural transfer.","section":"§4.4, Figure 4"},{"comment":"The abstract and §4.1/§4.3 claim state-of-the-art performance, but the tables show clear counterexamples: on Electricity and Traffic in Table 1, iTransformer achieves lower MSE than SGCMA (0.175 vs 0.189 and 0.422 vs 0.462); in Table 4 for h2→h1, SGCMA (0.574) is worse than TimeLLM (0.550) and GPT4TS (0.560). The aggregate improvement percentages should be recomputed honestly, or the claims should be qualified to 'most datasets' with explicit mention of the exceptions.","section":"Tables 1 and 4"},{"comment":"All results are reported as averages over three runs with no standard deviations, confidence intervals, or significance tests. Many conclusions rest on small margins (e.g., Table 5: 0.434 vs 0.440 on ETTh1; Table 4: 0.574 vs 0.550), so without variance information the reported improvements cannot be distinguished from noise. Please provide error bars or per-run results for the main tables and the ablation.","section":"§4 (experimental setup)"},{"comment":"The text corpus used for HMM training is never specified (no name, size, tokenizer, or preprocessing), and no code is released. This is a reproducibility gap for the core structural prior. At minimum, the corpus and HMM training details (number of EM iterations, initialization, vocabulary size) must be stated, and ideally the code should be made available.","section":"§3.1"}],"minor_comments":[{"comment":"In Eq. (7), the feature vector is written as x_p although the patch embedding was defined as z_p in Eq. (5), and the numerator uses μ_n while the denominator uses μ_c; please unify the notation and the summation indices.","section":"Eq. (7)"},{"comment":"The summation in Eq. (11) is written as 'K' but the number of hidden states is N; K is already used for the top-k tokens, so the index should be N.","section":"Eq. (11)"},{"comment":"The sentence before Eq. (10) says the temperature τ is set to d_h = d_llm/H, while Eq. (10) defines τ = sqrt(d_h); one of these is a typo and should be corrected.","section":"Eq. (10) and surrounding text"},{"comment":"The caption of Figure 4 should state which dataset and prediction setting are used, since the main text does not specify.","section":"Figure 4 caption"},{"comment":"The paper does not specify whether the channel-independent patching shares parameters across channels; please clarify in the setup.","section":"§3.2, Channel-Independent Patching"}],"recommendation":"major_revision","confidential_remarks":"This is a borderline case. The core idea is interesting and the empirical scope is good, but the claimed mechanism is not verified by the experiments as designed, and several SOTA statements are contradicted by the paper's own tables. If the authors add the missing controls and honestly revise the claims, the paper could be suitable for publication; in the current form, I cannot recommend acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, quick take: SGCMA is a genuinely new combination—HMM-derived transition prior, MEMM-style state decoding, top-k token cross-attention—and the reported margins over strong baselines are mostly positive. The writing is clear and the experiments cover long-term, short-term, few-shot, zero-shot, with ablations. The structural-transfer idea is worth taking seriously.\n\nThat said, the load-bearing claim is not yet supported. Section 3.2 uses A_text and pi_text only to initialize learnable A_time and pi_time. There is no regularization keeping A_time near A_text, and the state indices are arbitrary labels. So the L1 distance of 0.046 in Figure 4 is computed after the model has already tuned those parameters; it tells you the model chose a transition matrix close to text, but it doesn't tell you the text prior caused anything. The stress-test note is right: without a control replacing A_text with random/identity/uniform, or with an HMM trained on unrelated time series, the gains could come entirely from the trainable prototypes, cross-attention, and extra capacity. That's a missing experiment, not a cosmetic issue, because the abstract's claim is specifically about structural consistency.\n\nOther soft spots are more routine: no error bars despite claiming three runs, no code release, the text corpus for the HMM is unnamed, and a couple of table entries contradict the SOTA claims—iTransformer actually beats SGCMA on Electricity and Traffic in Table 1, and in Table 4 the h2->h1 zero-shot SGCMA MSE is worse than TimeLLM's. Those need fixing.\n\nOn balance: the method is coherent and the empirical pattern is encouraging enough that a serious referee should look at it. If the authors add the random-init control, show the HMM training setup, and tighten the claims, the central idea could be defensible. As written, I'd call it a promising conditional, not an established result.\n\nFor your use: it's worth a reading group discussion on how to design transferability controls, less as a settled forecasting result. I wouldn't cite it yet given the missing control and code.","headline":"A novel alignment pipeline with credible initial results, but the paper's core claim—that language structure transfers—is unverified because the transition matrix is free to drift away from its text initialization.","tokens_in":12606,"tokens_out":1680,"would_cite":false,"duration_ms":16228,"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":"SGCMA transfers a language-derived state-transition graph into time series patches, then aligns patch embeddings to language tokens, allowing a frozen GPT-2 to achieve state-of-the-art forecasting accuracy.","keywords":["time series forecasting","large language models","cross-modal alignment","hidden Markov model","maximum entropy Markov model","structure transfer","frozen LLM","state-transition graph"],"falsifier":"Train SGCMA with the language HMM's transition matrix replaced by a random or identity transition matrix of the same size, keeping all other components and the training budget identical. If forecasting performance on ETTh1/ETTm1 does not degrade substantially, the claim that the language-derived structural prior is the source of the gains is falsified; the gains would then come from the added trainable transformer, prototypes, and cross-attention instead.","tokens_in":11590,"feed_emoji":"📈","tokens_out":6412,"duration_ms":53266,"temperature":0.7,"pith_summary":"The paper argues that prior attempts to align time series with large language models fail because they match token-level features, whereas the real power of LLMs lies in modeling sequence-level structure. It proposes Structure-Guided Cross-Modal Alignment (SGCMA), which learns a hidden state-transition graph from natural language with a Hidden Markov Model, transfers that graph to time series patches through a Maximum Entropy Markov Model, and aligns patch embeddings to language tokens by cross-attention. The resulting structure-aware, semantically aligned embeddings are fed into a frozen GPT-2, and experiments report state-of-the-art forecasting accuracy on long-term, short-term, few-shot, and zero-shot benchmarks. The load-bearing idea is that giving time series a language-like state-transition structure is what unlocks the frozen LLM's generalization.","feed_headline":"Language state graphs sharpen frozen-LLM time series forecasts","feed_subtitle":"Frozen GPT-2 with a language state graph outperforms specialized forecasters on long- and short-term benchmarks.","key_machinery":"The central mechanism is the state-transition graph shared across modalities. The HMM provides the structure: an initial-state distribution $\\pi_{\\text{text}}$, a transition matrix $A_{\\text{text}}$, and an emission matrix $B_{\\text{text}}$; the transition matrix is transferred to the temporal domain as a learnable hot-start prior. A shallow transformer-based Maximum Entropy Markov Model (MEMM) then decodes a state-probability sequence over patches, combining the language-derived transition dynamics with each patch's own embedding. Semantic alignment uses top-k tokens per state from the emission matrix as prototypes in a cross-attention layer, and the final patch embedding is the state-probability-weighted expectation of the attended embeddings.","core_discovery":"SGCMA's central claim is that sequence-level structural consistency, not token-level mapping, is the right way to couple time series with pretrained language models. A Hidden Markov Model trained on text yields an initial-state distribution, a transition matrix, and an emission matrix; the transition matrix is then hot-started as a learnable prior in a shallow transformer-based MEMM that annotates each temporal patch with a state probability distribution. For semantic alignment, each patch attends to the top-k tokens of each hidden state as prototypes, and the final embedding is the state-probability-weighted average of those attended embeddings. Fed into a frozen GPT-2 followed by a linear head, this pipeline reports average MSE/MAE reductions of 3.7% over iTransformer in long-term forecasting and 3.6% over N-HiTS in short-term forecasting, with further gains in few-shot and zero-shot settings.","pith_inferences":["The paper does not ablate the origin of the structural prior; replacing the language HMM transition matrix with a random or identity matrix, or with an HMM trained on other time series, would test whether the language-derived structure itself, rather than the added trainable components, produces the reported gains.","The same structure-transfer recipe could transfer to other sequential modalities with Markovian latent dynamics, such as audio, DNA, or sensor streams, wherever a state-transition metaphor is appropriate.","If the structural prior is truly doing the work, the latent states should map to interpretable temporal regimes (rising, falling, stationary); checking whether the assigned states align with obvious trend patterns in the data would provide a direct interpretability check."],"forward_implications":["Frozen LLMs can serve as effective time series forecasters if the input embeddings carry language-like sequential structure, without any fine-tuning of the LLM weights.","The reusable structural prior extracted from text can be applied to time series from arbitrary domains, supporting few-shot and zero-shot transfer.","The ablation results imply that both structure alignment and semantic alignment are necessary: removing either degrades forecasting accuracy.","Because the LLM is frozen, the method's gains come from the input representation quality, so better structural priors should translate directly into better forecasts."],"supporting_citations":[{"why":"Supplies the HMM formulation (initial distribution, transition and emission matrices) used to extract the language state-transition structure.","marker":"[27]"},{"why":"Provides the Maximum Entropy Markov Model formulation that the structure alignment module adapts to decode temporal patch states with a hot-started transition matrix.","marker":"[22]"},{"why":"Defines the frozen GPT-2 backbone used as the LLM that consumes the aligned patch embeddings for forecasting.","marker":"[30]"},{"why":"Time-LLM, the token-level alignment baseline that SGCMA compares against in long-term, few-shot, and zero-shot experiments.","marker":"[13]"},{"why":"GPT4TS, an earlier LLM-adaptation baseline that repurposes a pretrained language model for time series and serves as a comparison point.","marker":"[44]"},{"why":"iTransformer, the best non-LLM long-term forecasting baseline, against which SGCMA reports an average 3.7% MSE/MAE reduction.","marker":"[19]"},{"why":"N-HiTS, the previous best short-term forecasting model, against which SGCMA reports a 3.6% improvement on the M4 benchmark.","marker":"[5]"},{"why":"PatchTST, the patch-based transformer baseline underlying the patching design and a key few-shot comparison.","marker":"[23]"}],"fun_headline_variants":["Structure, not tokens, aligns LLMs for time series","Language state graphs sharpen frozen-LLM forecasts","Sequence-level alignment beats token mapping for LLM forecasting","State graphs from text improve time series prediction with LLMs","Structure-guided cross-modal alignment lifts LLM forecast accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that the latent state-transition structure estimated from natural language text is a meaningful structural prior for arbitrary numerical time series, so that transferring that graph helps rather than distorts the forecasting task.","fun_headline_variants_meta":{"raw":{"variants":["Structure, not tokens, aligns LLMs for time series","Language state graphs sharpen frozen-LLM forecasts","Sequence-level alignment beats token mapping for LLM forecasting","State graphs from text improve time series prediction with LLMs","Structure-guided cross-modal alignment lifts LLM forecast accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000261,"raw_usage":{"total_tokens":1617,"prompt_tokens":991,"completion_tokens":626,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":607,"completion_tokens_details":{"reasoning_tokens":549}},"tokens_in":607,"tokens_out":626,"duration_ms":6227,"temperature":1.0,"reasoning_tokens":549,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:18:03.282600+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train SGCMA with the language HMM's transition matrix replaced by a random or identity transition matrix of the same size, keeping all other components and the training budget identical. If forecasting performance on ETTh1/ETTm1 does not degrade substantially, the claim that the language-derived structural prior is the source of the gains is falsified; the gains would then come from the added trainable transformer, prototypes, and cross-attention instead.","supporting_citations":[{"cited_title":"Maximum entropy markov models for information extraction and segmentation","cited_arxiv_id":null,"evidence_quote":"Provides the Maximum Entropy Markov Model formulation that the structure alignment module adapts to decode temporal patch states with a hot-started transition matrix."},{"cited_title":"One fits all: Power general time series analysis by pretrained lm","cited_arxiv_id":null,"evidence_quote":"GPT4TS, an earlier LLM-adaptation baseline that repurposes a pretrained language model for time series and serves as a comparison point."},{"cited_title":"Nhits: Neural hierarchical interpolation for time series fore- casting","cited_arxiv_id":null,"evidence_quote":"N-HiTS, the previous best short-term forecasting model, against which SGCMA reports a 3.6% improvement on the M4 benchmark."}],"review_version":1}