{"id":"3054243a-4c6d-48ea-bb45-3d88b9a5798c","arxiv_id":"2611.16322","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":11,"one_line_summary":"Deep DualMamba uses bidirectional selective state space models to forecast temperature with claimed 14.2% lower RMSE and 5x speedup over Transformers.","lead":"This paper proposes Deep DualMamba, a deep learning framework for climate temperature prediction that uses bidirectional state space models instead of Transformers. It claims to achieve lower error and faster inference on UK temperature data while providing theoretical convergence guarantees.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek/deepseek-v4-flash","headline":"Empirical results contradict the central claim: Table IV shows DualMamba-4 test RMSE 0.0292 vs Mamba-4 0.0278, so the claimed 14.2% improvement over baselines is not supported by the paper's own data.","rationale":"The reader's verdict is REJECT with high confidence, and the weakest assumption identified includes the empirical contradiction in Table IV. My stress-test confirms this is the single most load-bearing concern: the paper's headline empirical claim is directly contradicted by its own reported numbers. The reader also noted lack of code, error bars, and comparison with recent SOTA, but the internal inconsistency alone is sufficient for rejection. The theoretical analysis (Theorems 1-3) is not rigorous and relies on unverified assumptions (e.g., PL condition, continuous kernel operator interpretation of discrete Mamba), but even if the theory were correct, the empirical failure would still invalidate the paper's main contribution. The hyperparameter sensitivity showing near-identical performance across all settings further suggests the model is not actually learning meaningful temporal structure. I agree with the reader's assessment and recommend no change to the verdict.","tokens_in":22760,"tokens_out":807,"duration_ms":10141,"concrete_test":"Recompute the test-set RMSE for DualMamba-4 and Mamba-4 from Table IV using the original data and code. If the values are correct, then DualMamba does not outperform standard Mamba, and the central claim of superior accuracy is false. Additionally, compute 95% confidence intervals via bootstrapping (1000 resamples) for each model's test RMSE to assess whether any differences are statistically significant. Report the exact baseline used for the claimed 14.2% improvement and verify the calculation.","verdict_should_be":"REJECT","load_bearing_attack":"The paper's strongest claim is that Deep DualMamba achieves superior predictive accuracy with a reduced parameter footprint, specifically 14.2% lower RMSE than baselines. However, Table IV directly contradicts this: on the test set, Mamba-4 achieves RMSE 0.0278 while DualMamba-4 achieves 0.0292—DualMamba is worse. For the 3-depth variants, Mamba-3 test RMSE is 0.0308 and DualMamba-3 is 0.0314, again worse. The only metric where DualMamba-4 leads is training RMSE (0.0257 vs 0.0281), but that is irrelevant for generalization. The paper claims '14.2% lower RMSE' but provides no calculation or reference to which baseline or which metric this applies. The abstract and introduction state '14.2% lower RMSE and 5x faster inference than Transformers,' but Table IV shows Transformer-3 test RMSE 0.0534 and Transformer-4 0.0659, so a comparison to Transformers would indeed show improvement—but the claim is phrased as 'compared to MLP and Transformer baselines' (contributions bullet) and the 14.2% figure is not traceable to any specific entry. The reader correctly identified this inconsistency. This is not a minor typo; it undermines the central empirical claim of the paper. If the model does not outperform standard Mamba, the entire motivation for the bidirectional architecture collapses. The paper also lacks error bars, confidence intervals, or statistical significance tests, so we cannot assess whether any observed differences are meaningful. The hyperparameter sensitivity analysis (Table V) shows nearly identical RMSE across all historical/prediction horizons (e.g., Radcliffe: 0.1194–0.1201), which is suspiciously flat and inconsistent with a model that is supposed to capture long-range dependencies. This further weakens the empirical case.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek/deepseek-v4-flash","summary":"This paper proposes Deep DualMamba, a forecasting architecture that stacks forward and backward Mamba blocks per layer, fuses their hidden states, adds an adaptive residual connection with coefficient alpha, and applies a feed-forward network before projecting to the prediction horizon. The framework is evaluated on eight UK temperature station series from the CET dataset with input horizons of 600-1000 steps and prediction horizons of 300-500 steps, against MLP, RNN, GRU, LSTM, Transformer, and vanilla Mamba baselines. The paper claims a 14.2% RMSE improvement over MLP/Transformer baselines (and a 32% error reduction in the abstract), roughly 5x faster inference than Transformers, universal approximation and linear convergence guarantees (Theorems 1-3), and horizon-insensitive performance in a hyperparameter ablation (Tables V-VI).","tokens_in":23257,"tokens_out":15501,"duration_ms":93604,"significance":"The architecture description (Algorithm 2 and the DualMamba block) is clear, and the paper usefully documents its hypotheses, including the kernel-operator assumption in Eq. (14) and the PL condition in Assumption 1. If the empirical claims were substantiated, the contribution would be a moderate incremental engineering result: a bidirectional residual variant of Mamba with linear complexity that improves over vanilla Mamba and Transformers on a real climate-temperature benchmark. However, the paper provides no code, checkpoints, or error bars; the headline numbers (14.2%, 32%, 5x) are not traced to any reported measurement; the theoretical theorems are generic statements about state-space models rather than architecture-specific results; and, as detailed below, the central accuracy claim is contradicted by the paper's own Table IV. On the evidence reported, the significance of the contribution cannot currently be assessed as positive.","major_comments":[{"comment":"The central empirical claim is contradicted by the paper's own results. In Table IV, the test RMSE of DualMamba-4 is 0.0292 versus 0.0278 for Mamba-4, and DualMamba-3 gives 0.0314 versus 0.0308 for Mamba-3; the only metric on which DualMamba-4 beats Mamba-4 is training RMSE (0.0257 vs 0.0281), which is irrelevant to generalization. The statement in Section V.C that \"compared to villa-Mamba, our proposed DualMamba also achieves better forecasting performance\" is therefore not supported by the table it cites. The contribution bullet claiming \"14.2% lower RMSE\" cannot be traced to any entry of Table IV, and the abstract's \"32%\" error-reduction figure is inconsistent with the 14.2% figure. These numbers must be re-derived from the reported tables or the claims must be corrected before any accuracy statement can be evaluated.","section":"Section V.C, Table IV, contributions bullet (p. 3)","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"Table IV shows data-integrity problems. First, the GRU-3 row is a verbatim duplicate of the MLP-3 row (MSE 0.0487/0.0507/0.0579, RMSE 0.2208/0.2251/0.2406, MAE 0.1672/0.0738/0.0076, MAPE 27.96/28.52/28.91), which is implausible for two different architectures. Second, the MSE and RMSE columns are mutually inconsistent across rows: for MLP-4 test, MSE is 0.0742 while RMSE is 0.0563, yet sqrt(0.0742) = 0.2724; for Mamba-4 test, RMSE 0.0278 matches sqrt(0.000777), which implies the MSE column is scaled by 10^-2 in some rows but not in others. Third, Table V reports an RMSE of 0.3456 for Radcliffe with L=1000, T=300, whereas Table IV, which covers the same stations and the same horizon configuration, reports DualMamba RMSE values of 0.0257-0.0292; the factor-of-about-13 gap is never explained. No standard deviations, confidence intervals, or multiple-seed results are reported anywhere.","section":"Table IV and Table V","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The universal approximation proof begins by asserting that the Mamba block implements a continuous kernel operator (Ku)(t) = integral K(t,s;theta)u(s)ds. This assertion is not satisfied by the implemented architecture: in Algorithm 1, the discretized state parameters are input-dependent (Delta, B, and C are computed from x'), so the map u -> y is nonlinear and context-dependent and cannot be written as a fixed Volterra kernel acting linearly on u. The cited universal approximation result for Volterra operators [41] therefore does not apply to the discrete, selective model that is actually trained. Steps 2 and 3 of the proof (composition density and residual flow-map approximation) are likewise asserted without verifying the hypotheses of the cited deep-operator and neural-ODE results for this architecture.","section":"Section IV.A, Theorem 1, Eq. (14)","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The linear convergence rate in Eq. (18) depends on the Polyak-Lojasiewicz condition in Assumption 1. The paper itself limits this assumption in Remark 1 to a claim that the gating mechanisms \"ensure the PL condition holds locally,\" but no argument is given that the training trajectory remains in such a basin, so the rate is not established for the actual optimization run. The contractivity bound ||A_bar||_2 <= e^{-gamma Delta} < 1 in Eq. (19) concerns only the discretized state-transition matrix and does not imply the full-layer Jacobian bound in Eq. (21), since the full layer includes input-dependent projections and the SiLU nonlinearity. In addition, the model is trained with AdamW per Section V.B while the proof's Step 3 invokes an Adam convergence framework [44]; no argument connects that framework to AdamW under Assumption 1.","section":"Section IV.B, Theorem 2, Assumption 1, Eqs. (18)-(21)","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The bound N_DualMamba = O(kappa^{L+T} log(1/epsilon)) in Eq. (23) decreases as the input and output dimensions L and T grow when kappa < 1, which is not a plausible scaling for approximating a family of maps from R^L to R^T; the \"compression factor\" kappa is never defined. The supporting claims-the information-preservation inequality in Eq. (26) and the tensor-train representation in Eq. (28) with bond dimension alpha_t <= D-are asserted without proof, and the error bound e^{-cD} in Eq. (29) is not connected to the stated rate. The comparison also places the output dimension T in the MLP rate's exponent N^{-r/(L+T)} in Eq. (24), which inflates the MLP's difficulty, since the output dimension multiplies the number of output units rather than entering the input-dimension curse of dimensionality. Consequently, the limit claim lim_{epsilon->0} N_MLP/N_DualMamba = infinity in Section VI is not established.","section":"Section IV.C, Theorem 3, Eqs. (23)-(29)","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The evaluation is too thin to support the paper's comparative claims. The baselines omit the current standard long-term forecasting methods (e.g., PatchTST, iTransformer, DLinear, TimesNet) that the paper itself acknowledges in its related-work discussion; all results come from single runs with no error bars or significance tests; and the claimed 5x inference speedup and the 1.1x-of-MLP latency statement in Section V.B are not backed by any timing measurements. The dataset description is also internally inconsistent: Section V.A describes a 1 km-resolution gridded monthly CET product over 1961-2024, while Table I lists station series beginning in 1815 with tens of thousands of samples, and the use of 1000-step histories with 300-step prediction horizons is never reconciled with monthly sampling.","section":"Section V.A-B and Section V.C","authors_acknowledge":null,"authors_acknowledge_p":null}],"minor_comments":[{"comment":"The text contains a dangling reference: \"the description in Figure??\" was never filled in with a figure number.","section":"Section III.B","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"Figure 5's caption names Radcliffe, Rothamsted, Ross on Wye, and Stonyhurst, but the figure shows three panels, while Figure 6 shows four panels under a caption ending in \"Respectively.\"","section":"Figures 5 and 6","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"Terminology is inconsistent: \"villa-Mamba\" and \"Mamba-villa\" in Section V.C presumably mean vanilla Mamba; \"preceding horizon\" is used where \"prediction horizon\" is meant; and the abstract's \"theoretical analyzed\" should read \"theoretically analyzed.\"","section":"Throughout","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"Table III lists hidden widths of 60 for the Mamba and baseline variants, while Section V.B states that DualMamba uses hidden layers of 64 neurons; the exact configuration used for the reported runs should be stated.","section":"Table III and Section V.B","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"MAPE is defined with the observed value in the denominator; near-zero temperature values make this metric unstable, and the paper should state how such observations were handled when reporting MAPE in Table IV.","section":"Eq. (35)","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The claim that this work is \"the first demonstration of Mamba's viability in real-world climate prediction\" is difficult to reconcile with refs. [27]-[32], which the paper itself cites for Mamba applications to weather and traffic forecasting.","section":"Section I","authors_acknowledge":null,"authors_acknowledge_p":null}],"recommendation":"reject"},"author_rebuttal":null,"desk_editor":{"model":"deepseek/deepseek-v4-flash","letter":"Yongxiang Lei's paper applies a bidirectional Mamba to UK temperature data. The reader's take is accurate and I largely agree. Let me be direct.\n\nWhat is actually new: very little. Bidirectional Mamba for time series was already published as Bi-Mamba4TS (ref [28]), which the paper cites. The adaptive residual coefficient is a minor tweak. The application to climate data is a new domain, but that alone does not constitute a research contribution.\n\nWhat the paper does well: it assembles a forecasting pipeline on real UK Met Office data and runs a reasonable set of baselines. The writing is clear in places. That is where the credit ends.\n\nThe soft spots are load-bearing. First, the central claim of 14.2% lower RMSE is unsupported. Table IV shows DualMamba-4 test RMSE 0.0292 versus Mamba-4 0.0278—DualMamba is worse. The paper never explains where the 14.2% figure comes from. This is not a minor inconsistency; it undermines the entire empirical case. Second, the theoretical analysis (Theorems 1-3) assumes the Polyak-Łojasiewicz condition and a continuous kernel operator representation without verifying either for the discrete architecture. These are generic arguments that do not depend on the proposed modifications. Third, the hyperparameter sensitivity analysis (Table V) shows nearly identical RMSE across all historical/prediction horizons (e.g., Radcliffe: 0.1194–0.1201). That flatness is suspicious and inconsistent with a model claimed to capture long-range dependencies. Fourth, the paper lacks error bars, confidence intervals, code, and comparison with recent state-of-the-art methods like PatchTST, TimesNet, or DLinear. Fifth, the validation MSE for DualMamba-3 (0.1591) versus training MSE (0.0765) suggests overfitting or data leakage.\n\nWho this paper is for: perhaps a practitioner looking for a quick baseline on UK temperature data, but they would be better served by the existing Mamba or Bi-Mamba4TS implementations. The paper does not provide a reliable contribution.\n\nRecommendation: desk reject. The empirical contradiction alone is fatal. The theoretical analysis is not rigorous, the novelty is marginal, and the evaluation is incomplete. This does not deserve referee time.","headline":"The paper's central empirical claim is contradicted by its own Table IV, and the theoretical analysis is generic and unverified. The bidirectional Mamba idea is not new. This is not ready for peer review.","tokens_in":23830,"tokens_out":590,"duration_ms":9728},"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek/deepseek-v4-flash","headline":"DualMamba claims 14.2% lower RMSE and 5x faster inference than Transformers for UK temperature forecasting.","keywords":["time series forecasting","temperature prediction","state space models","Mamba","deep learning","climate informatics","bidirectional encoder"],"falsifier":"Train a Transformer, a standard Mamba, and DualMamba on a public long-sequence benchmark (e.g., ETT, Weather, Electricity) with identical hyperparameter tuning; if DualMamba does not achieve significantly lower RMSE than both baselines, the claimed advantage is not general.","tokens_in":24157,"feed_emoji":"🌡️","tokens_out":1310,"duration_ms":10297,"temperature":0.7,"pith_summary":"This paper presents Deep DualMamba, a neural network for climate temperature prediction that replaces the attention mechanism of Transformers with a bidirectional state space model. The authors claim that this architecture achieves higher accuracy with fewer parameters and faster inference on monthly temperature data from eight UK stations. They also provide theoretical proofs of universal approximation and linear convergence under specific assumptions. The work is framed as the first operational application of a Mamba-based model to real climate data, moving beyond synthetic benchmarks. If the claims hold, DualMamba offers a computationally cheaper alternative to Transformers for long-sequence time series forecasting.","feed_headline":"DualMamba claims 14% lower error than Transformers on UK temperature data","feed_subtitle":"A bidirectional state space model achieves linear complexity and faster inference for long-sequence climate forecasting.","key_machinery":"The central mechanism is the bidirectional Mamba block, which processes input sequences in both forward and reverse directions using selective structured state space models (SSMs). The forward and reverse representations are fused via addition, then passed through a feed-forward network with an adaptive residual connection (coefficient α). This stack is repeated L times to form the encoder. The SSM uses input-dependent discretization of continuous parameters (A, B, C) via a step size Δ, allowing the model to selectively focus on relevant temporal features. The hardware-aware parallel algorithm (Algorithm 1) ensures linear-time computation by exploiting the recurrent structure of the discretized SSM.","core_discovery":"The paper argues that a bidirectional Mamba encoder, combined with residual connections and a hardware-aware parallel algorithm, can outperform both traditional recurrent networks and Transformers in predicting monthly UK temperatures. The authors report a 14.2% reduction in RMSE and a fivefold speedup in inference compared to Transformer baselines on the Central England Temperature dataset. They also assert that the architecture achieves linear complexity with sequence length and exhibits stable training dynamics across eight stations. The core claim is that the selective state space mechanism captures long-range dependencies more efficiently than attention, and that the bidirectional processing overcomes the unidirectional limitation of standard Mamba.","pith_inferences":["The empirical comparison does not include recent Transformer-free forecasting methods (PatchTST, TimesNet, DLinear), so the reported advantage over 'Transformers' may not hold against the current state of the art; a broader benchmark would clarify the practical significance.","The hyperparameter sensitivity analysis shows nearly identical performance across different historical and prediction horizons (Table V), which undermines the claim that the architecture is especially effective for long-term forecasting; if performance is flat, simpler baselines might suffice.","The theoretical analysis (Theorems 1-3) assumes continuous kernel operators and the Polyak-Łojasiewicz condition, which are not verified for the discrete architecture and real data; the practical convergence guarantees are weaker than the proofs suggest.","The paper does not release code or trained models, making it difficult to reproduce the results or test the claims on new data; reproducibility is a prerequisite for operational adoption."],"forward_implications":["If DualMamba's accuracy advantage generalizes beyond UK temperature data, it could replace Transformers in other long-sequence climate forecasting tasks (e.g., precipitation, wind) with lower computational cost.","The linear complexity with sequence length makes the architecture attractive for real-time or edge deployment where memory and latency are constrained.","The theoretical convergence guarantee (Theorem 2) suggests that training is stable even with deeper stacks, potentially enabling larger models without vanishing/exploding gradients.","The bidirectional design could be adapted to other domains where temporal context from both past and future is beneficial (e.g., speech, biosignals)."],"supporting_citations":[{"why":"Introduces the Mamba architecture (selective state space model) that DualMamba builds upon.","marker":"[19]"},{"why":"Provides the discretization method for continuous SSM parameters (Δ, A, B → Ā, B̄) used in the Mamba block.","marker":"[36]"},{"why":"Describes the structured state space model (S4) that underlies the selective SSM mechanism.","marker":"[37]"},{"why":"Supplies the universal approximation theorem for Volterra operators used in Theorem 1 proof.","marker":"[41]"},{"why":"Deep operator network framework used to argue density of bidirectional composition in operator space.","marker":"[42]"},{"why":"Neural ODE framework used to justify residual connections as identity approximators in Theorem 1.","marker":"[43]"},{"why":"Adam convergence framework used in Theorem 2 proof of linear convergence.","marker":"[44]"},{"why":"Provides the approximation error scaling for ReLU MLPs used in Theorem 3 comparison.","marker":"[45]"}],"fun_headline_variants":[],"cache_read_input_tokens":0,"weakest_assumption_plain":"The theoretical analysis assumes that the discrete Mamba block behaves as a continuous kernel operator and that the loss landscape satisfies the Polyak-Łojasiewicz condition, neither of which is verified for the actual architecture or data.","fun_headline_variants_meta":{"raw":{"variants":[]},"model":"deepseek/deepseek-v4-flash","effort":"low","cost_usd":9.2e-05,"raw_usage":{"cost":9.2365e-05,"is_byok":false,"cost_details":{"upstream_inference_cost":9.2365e-05,"upstream_inference_prompt_cost":8.7269e-05,"upstream_inference_completions_cost":5.096e-06},"total_tokens":987,"prompt_tokens":959,"completion_tokens":28,"prompt_tokens_details":{"audio_tokens":0,"video_tokens":0,"cached_tokens":0,"cache_write_tokens":0},"completion_tokens_details":{"audio_tokens":0,"image_tokens":0,"reasoning_tokens":0}},"tokens_in":959,"tokens_out":28,"cost_source":"openrouter_usage","duration_ms":1920,"image_pages":[],"temperature":1.0,"billed_attempts":1,"reasoning_tokens":0,"transport_provider":"openrouter","cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-11-13T06:36:57.429735+00:00","model_set":{"reader":"deepseek/deepseek-v4-flash"},"pipeline_version":"pith-pipeline@v1.4.1-daily-deepseek-or-medium","input_kind":null,"input_chars":null,"input_total_chars":null,"passes_missing":null,"published_standing":null,"published_confidence":null,"objection_audit":null,"run_record":null,"basis":{"kind":"not_recorded","source":null,"chars_given":null,"chars_total":null},"receipt":null,"falsifier":"Train a Transformer, a standard Mamba, and DualMamba on a public long-sequence benchmark (e.g., ETT, Weather, Electricity) with identical hyperparameter tuning; if DualMamba does not achieve significantly lower RMSE than both baselines, the claimed advantage is not general.","supporting_citations":[{"cited_title":"Mamba: Linear-Time Sequence Modeling with Selective State Spaces","cited_arxiv_id":"2312.00752","evidence_quote":"Introduces the Mamba architecture (selective state space model) that DualMamba builds upon."},{"cited_title":"Combining recurrent, convolutional, and continuous-time models with linear state-space layers,","cited_arxiv_id":null,"evidence_quote":"Provides the discretization method for continuous SSM parameters (Δ, A, B → Ā, B̄) used in the Mamba block."},{"cited_title":"Efficiently Modeling Long Sequences with Structured State Spaces","cited_arxiv_id":"2111.00396","evidence_quote":"Describes the structured state space model (S4) that underlies the selective SSM mechanism."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the universal approximation theorem for Volterra operators used in Theorem 1 proof."},{"cited_title":"Learning nonlinear operators via deeponet based on the universal approximation theorem of operators,","cited_arxiv_id":null,"evidence_quote":"Deep operator network framework used to argue density of bidirectional composition in operator space."},{"cited_title":"Neural ordinary differential equations,","cited_arxiv_id":null,"evidence_quote":"Neural ODE framework used to justify residual connections as identity approximators in Theorem 1."},{"cited_title":"On the Convergence of Adam and Beyond","cited_arxiv_id":"1904.09237","evidence_quote":"Adam convergence framework used in Theorem 2 proof of linear convergence."},{"cited_title":"Error bounds for approximations with deep relu networks,","cited_arxiv_id":null,"evidence_quote":"Provides the approximation error scaling for ReLU MLPs used in Theorem 3 comparison."}],"lane_accuracy":{"status":"not yet measured","lane":{"model":"deepseek/deepseek-v4-flash","input_kind":null},"url":"https://pith.science/accuracy"},"review_version":1}