{"id":"8618b6b7-8291-4303-bf4f-ddfc4917ef5e","arxiv_id":"2602.03564","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"CoGenCast couples a Qwen-based encoder-decoder with flow matching and reports strong MSE/MAE on ten time-series benchmarks.","lead":"This paper combines a pre-trained language model with a flow-matching generator to forecast time series, reporting better accuracy than LLM-only or diffusion-only baselines. It also proposes a one-step sampling scheme to keep inference fast.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Pretrained-weight transfer after attention-topology reconfiguration is untested; a randomly initialized same-architecture control is needed to support the core LLM-pretraining claim.","rationale":"The reader's weakest assumption identifies exactly the experimental control that is missing: a same-architecture, randomly initialized baseline. I agree this is the most load-bearing concern because the paper's novelty and explanation rest on reusing pretrained LLM weights after an attention-topology change. The existing Table 6 comparison against a 'vanilla Transformer' cannot rule out that Qwen's advantage comes merely from having more parameters or from the bidirectional encoder/causal decoder+flow architecture, independent of pretraining. If the pretrained initialization provides no benefit, the central 'pre-trained LLM' claim is either false or untested, and the architecture could be repackaged as a from-scratch transformer with no loss. I do not think this requires rejection: the reported numbers could still be correct, and a random-init control might support the claim. But the paper is conditional on that control being run. A secondary concern is the one-step derivation gap in Eq. (8): an integral over tau normally requires multiple function evaluations, yet the paper claims one-step/1-NFE generation. This is worth checking against the released code, but it affects the efficiency claim rather than the attribution of accuracy gains, so I prioritize the pretraining transfer control. The paper deserves credit for releasing code and running broad ablations; the missing control is a concrete, easily fixable gap.","tokens_in":33577,"tokens_out":5230,"duration_ms":57052,"concrete_test":"Train the full CoGenCast model from scratch with the same hyperparameters, datasets, and all architectural components (bidirectional LLM encoder, causal-plus-cross-attention decoder, denoising decoder), but initialize all backbone parameters randomly (e.g., standard GPT-style init) instead of from Qwen3-0.6B. Run this control on at least Energy, ETTh1, Exchange, and Wind and compare MSE/MAE against Table 6. If the random-init control matches Qwen-init within the reported margins, the pretrained-weight transferability assumption fails and the 'pre-trained LLM' attribution is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that reconfiguring a pre-trained decoder-only LLM (Qwen3-0.6B) into an encoder-decoder via attention-topology changes preserves and exploits pretrained knowledge (§3.3). The only backbone comparison offered (Table 6, §4.6) is Qwen3-{0.6B,1.7B,4B} against a 'vanilla Transformer', not against the same CoGenCast architecture with randomly initialized weights. Table 6 therefore cannot separate the contribution of pretrained weights from the contribution of larger parameter count or of the encoder-decoder-plus-flow architecture. If changing causal self-attention to bidirectional self-attention and adding cross-attention destroys the value of the pretrained weights, the observed gains would be attributable to scale and architecture, not to 'pre-trained LLM' semantic understanding. The paper argues in Appendix D.1 that the full model beats even a 1.7B encoder-only variant, but that comparison holds the pretrained initialization fixed for both sides and does not test transferability. Since the abstract, title, and contribution list all rest on the phrase 'pre-trained LLM', this missing random-init control is load-bearing: it is the one experiment that would validate or invalidate the pretraining-based attribution of the reported improvements.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CoGenCast, a generative time-series forecasting framework that (i) reconfigures a pretrained decoder-only LLM (Qwen3-0.6B) into an encoder-decoder by changing causal self-attention to bidirectional self-attention in the encoder and adding cross-attention in the decoder, and (ii) couples this backbone with a flow-matching denoising decoder that predicts an interval-conditioned average velocity field. Training uses an autoregressive patch loss and a JVP-corrected flow-matching objective (Eq. 7); inference generates patches autoregressively and, the authors claim, samples each patch in one function evaluation (Eq. 8). Experiments on ten benchmarks compare against LLM-based, generative, and transformer baselines, with ablations on architecture, AR-flow components, context features, backbone scale, NFE, patch size, and noise schedule.","tokens_in":33960,"tokens_out":9844,"duration_ms":103645,"significance":"If the empirical results are robust, the proposed framework is a meaningful step: it is one of the few systems that combine LLM semantic conditioning with continuous stochastic generation while avoiding iterative sampling. The evaluation spans ten datasets, per-horizon results are provided in the appendix, and the code is released, all of which are strengths. The empirical claim is not circular, since the flow objective is a training loss rather than a quantity fitted to benchmark results. However, the manuscript overstates the consistency of the wins, and the central attribution of gains to pretrained LLM weights is not yet supported by the experimental design. The methodological novelty of the JVP-corrected average-velocity objective also needs a derivation or explicit reference. These issues are fixable but require additional experiments and analysis.","major_comments":[{"comment":"The paper claims CoGenCast \"consistently outperforms previous compared baselines,\" but Table 2 shows the opposite on ETTh2: CoGenCast MSE 0.159 vs TimeDART 0.145 and PatchTST 0.142 (MAE 0.246 vs 0.253/0.251). No standard deviations, seeds, or significance tests are provided anywhere, and many between-method gaps in Table 2 are small (e.g., Solar 0.221 vs 0.232; ETTm2 0.109 vs 0.122). Please add multi-seed statistics or revise the claim to \"competitive on most benchmarks.\"","section":"§4.2, Table 2; abstract and contribution 3"},{"comment":"The load-bearing claim that reusing pretrained Qwen3-0.6B weights is beneficial is not tested. Reconfiguring causal self-attention to bidirectional self-attention and adding cross-attention may destroy the value of the pretrained weights, but the only backbone comparisons are against a vanilla Transformer (varying pretraining, architecture, and parameter count together) or against larger Qwen variants that also use pretrained weights. A same-architecture randomly initialized control (same encoder-decoder plus flow decoder, same patching/text embedding scheme) is required to attribute the observed gains to pretraining rather than scale or architecture. Appendix D.1's \"full 0.6B > encoder-only 1.7B\" comparison does not provide this control.","section":"§3.3 and Table 6"},{"comment":"The one-step/low-latency claim is not consistent with the written equation. Eq. (8) defines yout_j as y_j(0) + ∫_0^1 u(zτ,τ,zdec_out)dτ, an integral over a continuum; evaluating this is not a single function evaluation. If the denoising decoder outputs the average velocity û_j over [0,1], the one-step output should be y_j(0) + û_j. Please either rewrite Eq. (8) accordingly or state explicitly that 1-NFE refers to a one-step Euler approximation of the integral, and report measured latency/NFE.","section":"§3.4, Eq. (8)"},{"comment":"The JVP-corrected objective is insufficiently justified. The regression target v_j − (r−t) ∂u_j/∂t depends on the model's own time-derivative, so the loss is not a fixed supervised target; the text does not prove that minimizers learn the average velocity over [t,r]. This is the main methodological novelty, and the cited mean-flow papers (Geng et al. 2025a,b) are not connected to the derivation. Please provide a derivation or reference, and include a small synthetic check of the estimated velocity/straightness.","section":"§3.3, Eq. (7)"}],"minor_comments":[{"comment":"The header states L=336, while Section 4.1 and Table 2 use L=96. Please clarify which look-back setting produced the main results and ensure the appendix tables match the main-text averages.","section":"Appendix I, Table 8"},{"comment":"Notation is inconsistent: Eq. (6) uses hat z_j but Eq. (8) uses z_τ for the noised latent. Please unify.","section":"§3.3–§3.4, Eqs. (6) and (8)"},{"comment":"The figure contains a typo (\"knowdge\" for \"knowledge\") and reports only MSE without error bars or raw values in the main text.","section":"Figure 3"},{"comment":"The text calls the predictive intervals \"well-calibrated,\" but no quantitative coverage or calibration metrics are reported. Figure 6 alone is qualitative and does not support that wording.","section":"§4.6, Figure 6"},{"comment":"Mean flows and improved mean flows are cited but not discussed. Please articulate the difference between the proposed interval-conditioned objective and those methods.","section":"§3.3, Related Work"}],"recommendation":"major_revision","confidential_remarks":"The skeptical concern about the missing random-init control is decisive and, in my view, correct: without it, the title/abstract claim about pretrained LLM value is not established. The ETTh2 counterexample to \"consistent\" wins, the Eq. (8)/NFE inconsistency, and the underived Eq. (7) also need attention. None of these appears unfixable, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: CoGenCast pairs a Qwen3-0.6B decoder-only LLM, reconfigured into an encoder-decoder by flipping attention masks, with a mean-flow one-step denoising decoder. The combination is genuinely new, the experiments are broad, and the code is out. But the paper sells itself as consistently SOTA, which Table 2 doesn't support, and the one-step inference story is sloppy. The bigger issue: nothing tests whether the pretrained weights survive the attention-topology surgery.\n\nWhat it does well: The architecture is coherent. The average-velocity objective is imported from the mean-flow line (Geng et al.), but applying it to time series with cross-attention conditioning is a legitimate new configuration. Ablations cover architecture, AR/Flow components, context features, NFE, patch size, and noise schedule. The flow loss is a training objective, so the central empirical claim is not circular. Ten datasets with a released repo is real work.\n\nWhere it's soft: (1) Overclaiming. Table 2 shows CoGenCast loses on ETTh2 to TimeDART and PatchTST, and the abstract and conclusion still say \"consistently outperforms previous compared baselines.\" No error bars or seeds appear anywhere, so the wins may be noise. (2) Eq. 8 presents the output as an integral of the velocity field while the text claims a single function evaluation. One can approximate that integral with a single average-velocity evaluation, but the paper never says how; right now it reads as a genuine derivation gap. (3) The main text fixes L=96, but the full results in Appendix Table 8 are for L=336. The headline averages and the appendix numbers don't line up; a referee will catch that immediately. (4) The missing random-init control. Table 6 compares Qwen backbones to a vanilla Transformer, not to the same CoGenCast architecture with randomly initialized weights. The D.1 ablation holds pretrained init fixed on both sides. So the gains attributed to \"pre-trained LLM\" could just as easily come from parameter scale or the encoder-decoder-plus-flow structure. This is the one experiment that would give the paper's central claim its legs.\n\nBottom line: the architecture is plausible, the paper is readable, and the flaws are fixable. For anyone working on LLM-based time series or flow-matching forecasters, this is a useful data point, but the current version shouldn't be read as strong SOTA evidence. I'd send it to peer review, asking for error bars, a random-init control, and a clean rewrite of the inference description. I'd also read the revision.","headline":"CoGenCast is a plausible hybrid LLM+flow forecaster with broad experiments, but the 'consistently outperforms' claim overreaches and the pretrained-weight attribution is untested because the random-init control is missing.","tokens_in":34362,"tokens_out":3582,"would_cite":true,"duration_ms":36520,"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":"Coupling a pre-trained LLM, reconfigured only by attention-topology changes, with an interval-averaged flow-matching decoder yields one-step forecasts that the paper reports consistently outperform eight baselines across ten time-series ben","keywords":["time series forecasting","generative modeling","flow matching","large language models","attention topology","one-step generation","probabilistic forecasting","encoder-decoder"],"falsifier":"Train the same encoder–decoder–flow architecture from scratch (random initialization, same 0.6B scale, same data) and compare MSE/MAE on the ten benchmarks; statistical parity with CoGenCast would falsify the claim that pretraining, rather than architecture or training procedure, drives the gains.","tokens_in":33530,"feed_emoji":"⚡","tokens_out":4896,"duration_ms":48857,"temperature":0.7,"pith_summary":"CoGenCast argues that time series forecasting needs both semantic understanding of context and stochastic modeling of continuous dynamics, and that neither pure LLM nor pure diffusion/flow approaches deliver both. The paper's proposal is a single hybrid: a pre-trained decoder-only LLM whose attention topology is changed to form a bidirectional encoder plus causal decoder with cross-attention, followed by a flow-matching denoising decoder that predicts the average velocity over an integration interval. Because the learned trajectories are near-straight, the model generates each future patch in one function evaluation. On ten public benchmarks the paper reports consistent error reductions over eight baselines — roughly 11% lower MSE than LLM-based methods and over 7% lower than transformer-based methods — plus gains from cross-domain training. The load-bearing premise is that the LLM's pretrained weights survive the attention reconfiguration, a premise the paper does not test against a randomly initialized control.","feed_headline":"Recycled LLM + one-step flow beats time-series baselines by 11%","feed_subtitle":"Attention-only reconfiguration turns a decoder-only LLM into an encoder–decoder forecaster that needs just one sampling step.","key_machinery":"The load-bearing mechanism is the attention-topology reconfiguration of a pre-trained decoder-only LLM: replacing causal self-attention with bidirectional self-attention in the encoder and adding cross-attention in the causal decoder, initialized entirely from the LLM's pretrained weights. The second mechanism is interval-conditioned average-velocity flow matching: instead of predicting instantaneous velocity, the denoising decoder predicts the average velocity over an interval, with a JVP-based loss term that penalizes curvature; this keeps the transport path nearly straight, enabling one-step (1-NFE) sampling.","core_discovery":"The central claim is that the conditional forecasting distribution is best modeled by jointly learning an autoregressive semantic representation and a continuous stochastic transport, and that pre-trained LLM weights can be repurposed for this by modifying only the attention topology: the encoder uses bidirectional self-attention to fuse look-back patches and textual context, the decoder keeps causal self-attention and adds cross-attention, and a denoising decoder predicts a velocity field averaged over a time interval rather than an instantaneous velocity. A Jacobian-vector-product-corrected MSE loss drives the trajectories straight, so a single Euler step suffices at inference. The paper r","pith_inferences":["A decisive control the paper does not run: same reconfigured architecture with randomly initialized weights. If that control matches CoGenCast, the advantage would come from the architecture and training recipe rather than from pretrained weights.","The interval-averaging trick could transfer to other generative tasks — e.g., time-series imputation or multimodal generation — wherever a nearly straight probability path can be enforced by the JVP objective.","The attention-topology reconfiguration recipe suggests a general way to adapt decoder-only pretrained models to encoder–decoder tasks, which could be tested on text summarization or translation with minimal fine-tuning.","Because the paper's main table and appendix headers report different look-back lengths (L=96 vs L=336), a reader should confirm which setting produced the headline numbers before generalizing the comparison."],"forward_implications":["If the reported gains hold, one-step forecasting with LLM-scale semantic conditioning becomes practical for latency-sensitive applications, since a single function evaluation replaces iterative denoising.","Cross-domain training consistently lowers MSE versus in-domain training, implying that a single CoGenCast model can serve multiple domains without per-domain fine-tuning.","Removing either the autoregressive LLM backbone or the flow decoder degrades accuracy, supporting the paper's dual-capability thesis.","LLM backbones outperforming a vanilla transformer at equal scale suggests pretrained language model weights, not just scale, contribute to forecasting accuracy — provided the transferability assumption holds.","Extending the look-back window from 96 to 336 steps yields progressive improvements, so the architecture can exploit longer history without structural changes."],"fun_headline_variants":["Attention surgery turns LLM into a one-step time-series forecaster","Hybrid LLM-flow model forecasts time series with one Euler step","CoGenCast: bidirectional LLM + flow matching for unified forecasting","Pre-trained LLM reworked for stochastic time-series generation"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The pretrained LLM weights remain useful after swapping causal self-attention for bidirectional attention and inserting cross-attention; if they do not, the reported advantage over a same-scale randomly initialized transformer disappears, and this is not tested against a randomly initialized control.","fun_headline_variants_meta":{"raw":{"variants":["Attention surgery turns LLM into a one-step time-series forecaster","Hybrid LLM-flow model forecasts time series with one Euler step","CoGenCast: bidirectional LLM + flow matching for unified forecasting","Pre-trained LLM reworked for stochastic time-series generation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000659,"raw_usage":{"total_tokens":2835,"prompt_tokens":711,"completion_tokens":2124,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":455,"completion_tokens_details":{"reasoning_tokens":2051}},"tokens_in":455,"tokens_out":2124,"duration_ms":17829,"temperature":1.0,"reasoning_tokens":2051,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T04:53:54.731471+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same encoder–decoder–flow architecture from scratch (random initialization, same 0.6B scale, same data) and compare MSE/MAE on the ten benchmarks; statistical parity with CoGenCast would falsify the claim that pretraining, rather than architecture or training procedure, drives the gains.","supporting_citations":[],"review_version":1}