{"id":"553d7597-80c9-4695-8d8f-1cb127283bdd","arxiv_id":"2510.03244","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A cross-modal model that renders multivariate time series as images, encodes them with a frozen SigLip2 vision model, and fuses the features with a temporal attention branch achieves state-of-the-art results on seven forecasting benchmarks.","lead":"This paper turns multivariate time series into images and feeds them to a frozen vision model, then fuses the visual features with a temporal model to forecast the series. It reports lower forecast errors than several large time series models on seven standard benchmarks while training only 7.45% of the parameters.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The unique role of the frozen SigLip2 encoder is not isolated: a random or lightweight visual encoder could plausibly yield the same small gains, so the 'pre-trained LVM' explanation is unsupported as stated.","rationale":"The reader's verdict is already CONDITIONAL and correctly notes that the central mechanistic claim is not isolated and that rendering and training details are missing. My concern agrees with that assessment but is narrower: the missing control is not only a comparison to prior visual methods, but an ablation that separates the frozen pretrained encoder from the visual pathway as such. The reported results are coherent, and the ablation table at least shows a positive long-horizon effect from the visual branch, so the problem is best addressed by additional experiments rather than outright rejection. The name inconsistency (VIFO versus VFEM) and the typo in Table 2's horizon list are secondary editorial issues. I therefore recommend keeping the conditional verdict unchanged.","tokens_in":8919,"tokens_out":5770,"duration_ms":52292,"concrete_test":"On ETTh1 and ETTh2, rerun the VIFO pipeline from Section 3 with the only change being that the frozen SigLip2 encoder is replaced by (a) a frozen randomly initialized encoder of identical architecture and (b) a small trainable CNN consuming the same rendered image, holding all other modules, optimization budgets, and seeds fixed. If either variant matches the Table 2 MSE within 0.005 on average over the three seeds, the central 'pre-trained LVM' claim is not supported; if both are clearly worse, the claim survives.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that rendering multivariate time series as an image and passing it through a frozen SigLip2 LVM is what captures cross-variable patterns (Section 1). The experiments do not isolate that mechanism. The time-series branch in Section 2.5 is already a spatiotemporal attention model, not channel-independent, so comparisons with channel-independent baselines do not prove that the visual pathway is the source of cross-variable gains. Table 3 shows that removing the visual processor changes ETTh1 average MSE from about 0.399 to 0.383 and ETTh2 average MSE from about 0.339 to 0.331; there is no control that removes pretraining, replaces the LVM with a smaller or random encoder, or tests whether the gain is simply extra capacity. Without such controls, the advertised 'pre-trained LVM' and 'large vision model' explanation is one of several possible explanations of the small deltas. Rendering details are also unspecified: how scalar values become RGB channels, how per-variable scaling is applied before display, and how images with M < patch size (e.g., ETTh1 has M=7 with L=512) are presented to a patch-based encoder. These details matter because a frozen natural-image model's output depends heavily on the color map and geometry.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes VIFO (also called VFEM in the abstract), a dual-branch multivariate time series forecasting model that renders the M×L input as an image, feeds it to a frozen SigLip2 vision encoder, and fuses the visual hidden states with a spatiotemporal-attention time-series branch via a cross-modal fusion layer. On seven benchmarks with look-back 512 and forecast horizons 96–720, the reported MSE and MAE are lower than all compared baselines, while only 7.45% of the model parameters are trainable. The central claim is that rendering multivariate series as images makes complex cross-channel patterns visible to a pretrained large vision model, yielding gains that channel-independent baselines cannot achieve.","tokens_in":9300,"tokens_out":3850,"duration_ms":34392,"significance":"If the central claim is validated, the result would be interesting: a frozen general-purpose vision encoder acting as a cross-variable feature extractor, combined with parameter-efficient fine-tuning, would offer a new and practical direction for multivariate forecasting. The choice of SigLip2 with variable input size is a sensible match to the large range of variable counts across datasets, and the reported 7.45% trainable-parameter fraction is a genuine efficiency feature. However, the paper's evidence does not yet isolate the proposed mechanism: the visual branch is ablated only as a whole, the rendering procedure is unspecified, and the reported differences are often small and lack uncertainty estimates. The significance of the central claim therefore depends on additional controls and reproducibility details.","major_comments":[{"comment":"The image-generation process is not specified. Section 2.1 defines X_vs as an M×L×c image generated by rendering, but the paper never states how scalar series values are mapped to RGB channels, whether per-variable normalization or RevIN is applied before rendering, what colormap or value-range rescaling is used, or how the encoder handles images whose M or L is smaller than the vision encoder's patch size. For example, ETTh1 has M=7 and L=512, while Traffic has M=862 and L=512, so the geometry of the rendered image varies substantially across datasets. Because the central claim is that a pretrained natural-image encoder extracts meaningful cross-channel patterns from these renderings, the missing rendering recipe blocks reproducibility and leaves open the possibility that the visual branch exploits color-map artifacts rather than the intended cross-variable structure.","section":"Sections 2.1 and 2.4"},{"comment":"The ablation study removes the entire visual pathway, so it cannot isolate the contribution of pretraining. The comparison between 'w/ all' and 'w/o VS modal' only shows that some visual pathway helps, not that a frozen SigLip2 LVM specifically is responsible; a randomly initialized vision encoder or a much smaller untrained encoder could plausibly yield similar small deltas. The paper should include controls that replace the frozen SigLip2 encoder with (a) a randomly initialized encoder of the same architecture, (b) a smaller vision encoder, and (c) a version with the image input permuted or corrupted, so that the claimed mechanism is distinguished from a simple increase in model capacity. Without such controls, the advertised 'pre-trained LVM' explanation is only one of several possible explanations of the observed gains.","section":"Table 3"},{"comment":"No uncertainty estimates are reported even though the setup states that each experiment was run three times with different random seeds. Many of the reported differences are small enough that seed noise could change the ranking: for example, in Table 3 the ETTh1 horizon-96 MSE difference between 'w/o VS modal' (0.355) and 'w/ all' (0.348) is 0.007, and in Table 2 several comparisons between VIFO and the second-best baseline are of similar magnitude. Reporting means, standard deviations, or confidence intervals over the three seeds is necessary to support the claims that VIFO is best on all seven datasets and that each ablated module contributes meaningfully.","section":"Section 3.1 and Tables 2–3"},{"comment":"The temporal modality branch is not channel-independent: it uses a spatiotemporal attention structure over segments, as described in Section 2.5 and reference [27]. Therefore the comparison with channel-independent baselines (PatchTST, GPT4TS, Chronos) does not by itself demonstrate that the visual pathway is the source of cross-variable gains. The observed improvements could in principle come from the spatiotemporal temporal branch already modeling cross-variable dependencies. The paper should explicitly compare against a variant of the same temporal branch that is channel-dependent without the visual branch, or otherwise show that the visual branch contributes cross-channel information beyond what the temporal branch already captures.","section":"Section 2.5 and Section 1"}],"minor_comments":[{"comment":"The model name is inconsistent: the abstract and the beginning of the full text use 'VFEM', while the title, the rest of the full text, and the tables use 'VIFO'. One name should be used throughout.","section":"Abstract/title"},{"comment":"The table header lists F∈{96,129,336,720}; this should be F∈{96,192,336,720}. The same typo appears in the caption.","section":"Table 2"},{"comment":"Several cells lack spacing between consecutive numbers, for example '0.2940.330' in the ETTh2 horizon-96 row and '0.2040.248' in the Weather horizon-192 row. The formatting should be corrected.","section":"Table 2"},{"comment":"The sentence about long-sequence performance refers to 'forecasting length of 360' while the experiments use horizons 336 and 720; this is presumably a typo for 336.","section":"Section 3.2"},{"comment":"The caption and text state that '7.45% of its parameters' are trainable; this should be phrased as '7.45% of the total model parameters' for clarity, since Table 1 separately lists the frozen and trainable fractions.","section":"Table 1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript appears to be a conference-length paper with a strong central claim about the mechanism of pretrained vision models for time series forecasting. The empirical results are consistently favorable, but the missing rendering details and absent mechanism-isolating ablations are load-bearing for the paper's main message and should be addressed before publication. I would not reject the paper outright, as the deficiencies are fixable with additional experiments and description within the manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: consistently good numbers, but the central claim about the frozen vision encoder is not isolated, and the rendering details are missing. Still worth a referee.\n\nWhat's actually new here is the specific stack: a variable-size SigLip2 encoder, frozen, reads an M×L image of the multivariate series; a trainable spatiotemporal attention branch (PSFormer) reads the raw series; and cross-modal attention fuses the two. Only 7.45% of parameters are trainable. That combination is not in the paper's references, and the results across seven benchmarks and four horizons are uniformly in VIFO's favor. Credit where due: the parameter efficiency is real, the variable-size encoder is a sensible way to handle arbitrary variate counts, and the ablation shows both branches contribute. The claim that the visual branch captures long-term patterns better is weakly supported but not absurd.\n\nThe soft spots are proportionate to the central claim. The main one: the paper says the LVM captures cross-channel patterns 'invisible to channel-independent models,' but the temporal branch is already channel-dependent via PSFormer's spatiotemporal attention. So beating PatchTST and TimesNet does not establish that the visual pathway is the source of the gains. The ablation without the visual branch moves ETTh1 average MSE from 0.383 to 0.399 and ETTh2 from 0.331 to 0.339. Small deltas, no error bars, and no control replacing SigLip2 with a random or much smaller encoder. A frozen random encoder of the same size could plausibly give the same capacity bump. Without that control, 'pre-trained LVM' remains one explanation, not the demonstrated one. The rendering pipeline is also under-specified: how raw values map to RGB, per-variable scaling, and how a 7×512 image is fed to a patch-based encoder. These details matter because the frozen model's output depends on the color map and geometry. Second, the novelty is incremental: rendering time series as images and running vision models on them is an established idea, and none of that prior work is cited or compared. Third, no error bars despite three seeds, so the small gaps in Table 3 could be noise. The VFEM/VIFO name inconsistency is trivial.\n\nBottom line: this is a serious, reproducible-in-principle architecture, and the empirical pattern is consistent enough that a careful referee should look at it. But the mechanism claim needs to be either isolated with proper controls or softened, and the details need to be supplied. I'd send it to peer review expecting major revision.","headline":"Plausible architecture with consistently good numbers, but the central claim about the frozen vision encoder is not isolated; still worth a serious referee.","tokens_in":9703,"tokens_out":2780,"would_cite":false,"duration_ms":24015,"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":"Frozen vision model reads series plots and beats time-series baselines","keywords":["multivariate time series forecasting","cross-modal fusion","large vision model","channel-independent architecture","time series visualization","spatiotemporal attention","parameter-efficient fine-tuning","SigLip2"],"falsifier":"Permute the row order of variables before rendering the image and re-run the same training and evaluation; if forecast error is unchanged, the visual branch is not using cross-variable spatial layout. Alternatively, take a dataset with hundreds of variables and a 720-step horizon and check whether the reported long-horizon advantage persists at that extreme aspect ratio.","tokens_in":8746,"feed_emoji":"📈","tokens_out":8259,"duration_ms":63998,"temperature":0.7,"pith_summary":"This paper introduces a model called VIFO in the body text (VFEM on the title page) that renders a multivariate time series as a single image, with each variable as a row and each time step as a column, then feeds that image through a frozen large vision model. The authors claim this lets the vision encoder extract cross-variable patterns such as lead-lag relationships, shared periodicities, and holiday effects that channel-independent time series models cannot see. Those visual features are fused with a trainable temporal branch through cross-modal attention, and only 7.45% of the model's parameters are trained. On seven benchmark datasets and horizons of 96, 192, 336, and 720 steps, VIFO reports the lowest MSE and MAE among the compared baselines, with errors degrading more slowly at long horizons. The intended contribution is a parameter-efficient way to bring spatial pattern recognition to multivariate forecasting.","feed_headline":"Frozen vision model reads series plots and beats time-series baselines","feed_subtitle":"Turning variables into image rows lets a frozen vision encoder see cross-channel patterns at 7.45% training cost.","key_machinery":"The central object is the variable-sized image rendering of the multivariate series: rows are variables, columns are time steps, and pixel color encodes the normalized value at that point. The model uses a frozen SigLip2-base-NaFlex encoder because it accepts arbitrary image heights and widths, letting one vision backbone handle datasets with very different numbers of variables and a fixed look-back of 512 steps. A projection layer maps the vision encoder's patch embeddings to the hidden dimension, while a parallel temporal branch applies reversible instance normalization and a spatial-temporal attention module to the raw series. A cross-modal attention layer fuses the two hidden representations before a mapping layer produces the forecast. The parameter economy is the distinctive part: roughly 375 million parameters are frozen, and only about 30 million, or 7.45%, are trainable, covering the projection, temporal network, and fusion network.","core_discovery":"The paper's central claim is that a pre-trained vision model, kept frozen, can serve as a feature extractor for multivariate time series if the series is rendered as an M-by-L pixel image. The authors assert that this visual representation exposes patterns that are visible to the human eye and to a vision transformer but invisible to channel-independent forecasting architectures: daily and weekly periodicity, morning and evening rush hours, weekend effects, small holidays, and stable lead-lag relations between variables. VIFO combines that frozen visual branch with a trainable time-series branch using spatial-temporal attention, fuses the two representations with cross-modal attention, and maps the result to a forecast. The paper reports that this configuration achieves the best MSE and MAE on ETTh1, ETTh2, ETTm1, ETTm2, Electricity, Weather, and Traffic across all tested horizons, and that the visual branch matters most for long-horizon forecasting.","pith_inferences":["A direct test of whether the vision branch reads cross-variable structure rather than global texture would be to permute the row order of variables before rendering; if forecasts do not change, the encoder is not using inter-variable spatial layout.","The same rendering could be paired with image-style augmentations such as row or column shifts, color jitter, or patch masking to regularize the temporal branch, which the paper does not explore.","For datasets with extreme aspect ratios, such as hundreds of variables against 512 time steps, the variable-size encoder's patch grid may stretch the visual signal; checking whether the reported long-horizon advantage persists at that extreme would test the generality of the mechanism.","A comparison against the same temporal branch with a randomly initialized vision encoder of equal parameter count would isolate whether the gains come from pre-training on natural images or simply from having a large frozen feature extractor."],"forward_implications":["If the claim holds, channel-independent architectures are not the only efficient way to handle varying dataset dimensions; a frozen vision backbone can supply cross-variable structure without per-dataset full fine-tuning.","Because the vision encoder is frozen, the approach suggests that time series forecasting can inherit the scale and transfer properties of large vision models at a fraction of the training budget required by channel-dependent models.","The visual branch's larger contribution at long horizons implies that rendering-based encoders may be especially useful when forecasts extend beyond one or two seasonal cycles.","The variable-size encoding scheme extends to datasets with hundreds of variables, such as Traffic, without changing the architecture, which removes a common obstacle for channel-dependent designs.","The small trainable fraction means the method could be adapted to new domains by training only the projection, temporal, and fusion modules, making it a practical route for low-resource forecasting settings."],"supporting_citations":[{"why":"Supplies the frozen SigLip2-base-NaFlex vision encoder that accepts variable-size images.","marker":"[28]"},{"why":"Provides the spatial-temporal attention module used in the time-series branch and the fusion layer.","marker":"[27]"},{"why":"Chronos is a channel-independent foundation-model baseline that VIFO must outperform.","marker":"[29]"},{"why":"Moirai is a channel-independent foundation-model baseline that VIFO must outperform.","marker":"[19]"},{"why":"PatchTST is a strong channel-independent small-model baseline included in the comparison.","marker":"[10]"},{"why":"Reversible instance normalization is applied in the temporal modality branch to handle distribution shift.","marker":"[26]"},{"why":"UniTST is a channel-dependent baseline that models inter-series and intra-series dependencies, used for comparison.","marker":"[30]"}],"fun_headline_variants":["Frozen vision model reads series plots for forecasts","See time series as images: frozen vision model forecasts","Freeze vision model, train 7.45%, beat baselines","Frozen vision + temporal fusion: better multivariate forecasts","Frozen vision model sees cross-variable patterns in plots"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that rendering the multivariate series as a rows-by-columns image preserves the cross-variable relationships that matter for forecasting, and that a vision model pre-trained on natural images can extract them from that rendering.","fun_headline_variants_meta":{"raw":{"variants":["Frozen vision model reads series plots for forecasts","See time series as images: frozen vision model forecasts","Freeze vision model, train 7.45%, beat baselines","Frozen vision + temporal fusion: better multivariate forecasts","Frozen vision model sees cross-variable patterns in plots"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.002096,"raw_usage":{"total_tokens":8121,"prompt_tokens":897,"completion_tokens":7224,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":513,"completion_tokens_details":{"reasoning_tokens":7158}},"tokens_in":513,"tokens_out":7224,"duration_ms":43161,"temperature":1.0,"reasoning_tokens":7158,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T15:46:21.133580+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Permute the row order of variables before rendering the image and re-run the same training and evaluation; if forecast error is unchanged, the visual branch is not using cross-variable spatial layout. Alternatively, take a dataset with hundreds of variables and a 720-step horizon and check whether the reported long-horizon advantage persists at that extreme aspect ratio.","supporting_citations":[{"cited_title":"MOMENT: A family of open time-series foundation models,","cited_arxiv_id":null,"evidence_quote":"Moirai is a channel-independent foundation-model baseline that VIFO must outperform."},{"cited_title":"Time-LLM: Time series forecasting by repro- gramming large language models,","cited_arxiv_id":null,"evidence_quote":"Reversible instance normalization is applied in the temporal modality branch to handle distribution shift."},{"cited_title":"Timecma: Towards llm- empowered multivariate time series forecasting via cross-modality alignment,","cited_arxiv_id":null,"evidence_quote":"Supplies the frozen SigLip2-base-NaFlex vision encoder that accepts variable-size images."},{"cited_title":"Moderntcn: A modern pure convolution structure for general time series analysis,","cited_arxiv_id":null,"evidence_quote":"Chronos is a channel-independent foundation-model baseline that VIFO must outperform."},{"cited_title":"Reversible instance normalization for accu- rate time-series forecasting against distribution shift,","cited_arxiv_id":null,"evidence_quote":"UniTST is a channel-dependent baseline that models inter-series and intra-series dependencies, used for comparison."}],"review_version":2}