{"id":"27db53f8-1282-4d63-8490-d14ccb544295","arxiv_id":"2505.04445","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"M2Rec blends Mamba sequence modeling, a learnable FFT noise filter, and LLM text embeddings through two scalar weights, reporting 3% HR@10 gains over Mamba baselines.","lead":"This paper combines Mamba, Fourier transforms, and language-model embeddings into a sequential recommender that predicts a user's next interaction. The authors report small accuracy gains over earlier Mamba models, but the claimed time-based periodicity detection and dynamic gating are not fully supported by the model as written.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. 11 runs FFT over interaction positions, not timestamps, so the claimed daily/weekly periodic-pattern mechanism is not grounded in the model's input.","rationale":"The reader and I identify the same load-bearing concern, and it is an internal inconsistency rather than a disagreement with consensus. Eq. 11's t is the position index in Xu, and no timestamp or inter-arrival information appears anywhere in the model input, so the FFT branch filters sequence-position frequencies, not daily or weekly temporal cycles. This directly undermines the paper's central contribution as stated in the abstract, Figure 1, and Section III-D: the model cannot explicitly model the multi-scale periodic user behaviors used to motivate the architecture. The concern is specific and testable by reindexing the same interaction data by position versus timestamp. I do not recommend moving from the reader's CONDITIONAL verdict to REJECT because the empirical SOTA and efficiency claims could in principle survive as generic frequency-domain feature extraction; the remedy is either to add timestamp-aware sampling/resampling to Eq. 11 or to weaken the periodic-pattern claim. The table inconsistencies and lack of a code commit hash noted by the reader are real verification blockers and reinforce the conditional status, but the FFT-position issue is the most load-bearing conceptual flaw. The anonymous code link is a positive step toward reproducibility, though it cannot be executed without a fixed revision.","tokens_in":17596,"tokens_out":7969,"duration_ms":83542,"concrete_test":"Recompute the Fig. 1-style spectrum on ML-1M in two ways: (a) apply Eq. 11 to user event sequences indexed by interaction position t = 1..T; (b) apply the same FFT to the same events indexed by actual wall-clock timestamps resampled onto a uniform daily grid. If the position-indexed spectrum lacks the daily and weekly peaks that the timestamp-indexed spectrum exhibits, the AFFT branch of M2Rec cannot be capturing the motivating periodic patterns. A complementary behavioral check: train M2Rec on two copies of ML-1M with identical item order but all inter-arrival intervals compressed by a constant factor. If the model is truly time-aware, HR@10 should change; if results are identical, the AFFT branch is blind to wall-clock periodicity.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Eq. 11 applies the DFT to Xu[t] with t = 1..T as the position index of items in the user interaction sequence. The model never receives timestamps: the input to Algorithm 1 is Xu: (B, T, |V|) plus context C used only for LLM embeddings, and no inter-arrival times are used in Section III-D. The DFT assumes uniform sampling in the indexed variable; in interaction-position space the unit is interactions, not days. Because user interactions are irregularly spaced (ML-1M averages 165.6 interactions/user while Amazon Beauty averages 8.9), a spectral peak at k/T in Eq. 11 corresponds to cycles per interaction and cannot be equated with the 1 cycle/day and ~0.14 cycle/day peaks displayed in Fig. 1. Consequently the adaptive low-pass mask M := |F| <= theta in Eq. 12 removes position-frequency content, not temporal noise. The central claimed mechanism, explicit multi-scale modeling of daily/weekly/monthly periodic user behavior, is therefore not operational in the model as written. The 3.2% HR@10 gain may still exist as a generic sequence-smoothing or feature-transformation effect, but the paper's headline periodic-pattern contribution and its Fig. 1 motivation are unsupported by the architecture.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes M2Rec, a sequential recommendation model that combines a Mamba-based state-space encoder with an adaptive Fast Fourier Transform (AFFT), LLM-generated item embeddings, and a learned gate for fusing feature streams. The authors claim state-of-the-art performance on four datasets, with HR@10 improved by about 3.2% over Mamba-based baselines and faster inference than Transformer baselines. The manuscript includes an architecture description, complexity analysis, main results, ablations, robustness tests, and hyperparameter studies.","tokens_in":17795,"tokens_out":3954,"duration_ms":39272,"significance":"If the reported improvements are reproducible, the paper makes a useful empirical contribution to efficient sequential recommendation, particularly the combination of a linear-time Mamba backbone with frequency-domain filtering and semantic embeddings. The experiments cover a reasonable range of baselines and datasets, and the paper includes ablation, robustness, and efficiency analyses, which are strengths. However, the central mechanistic claim—that the FFT module explicitly models daily, weekly, and monthly periodic user behavior—is not grounded in the model's actual input, and several internal inconsistencies in the data and hyperparameter tables currently undermine confidence in the reported results.","major_comments":[{"comment":"The DFT in Eq. (11) is applied to Xu[t] with t as the position index in the interaction sequence. Algorithm 1's input is Xu: (B, T, |V|) and context C used only for LLM embeddings; no timestamps or inter-arrival times enter the model. Consequently, the frequency axis is cycles per interaction, not cycles per day, and the peaks at 1 cycle/day and ~0.14 cycles/day shown in Fig. 1 cannot be obtained from Eq. (11). The paper's central claim that M2Rec explicitly captures daily, weekly, and monthly periodic user behavior is therefore not operational in the architecture as written. This is a load-bearing mismatch between the motivating analysis and the actual mechanism; please either modify the model to consume timestamps or reframe the contribution as position-frequency filtering/sequence smoothing.","section":"Section III-D, Eq. (11) and Algorithm 1"},{"comment":"The dataset statistics are internally inconsistent. The text states that Texas has 27.2M interactions and is the largest dataset, but Table III reports only 1,344,379 interactions for Texas. In addition, the reported average interactions per user do not match Table III: New York has 478,903/6,195 = 77.3 vs the stated 106.4, and Texas has 1,344,379/24,559 = 54.7 vs the stated 66.0. These discrepancies affect the validity of the experimental setup and must be corrected and reconciled.","section":"Section IV-A and Table III"},{"comment":"The Mamba Layer rows are verbatim identical to the θ rows (e.g., HR@10 values 4.72, 4.69, 4.56, 4.58, 4.59 appear in both blocks). As presented, the paper reports no distinct hyperparameter results for the number of Mamba layers, so the RQ6 layer analysis is unsupported. Moreover, the accompanying text says a lower cut ratio captures more high-frequency components, but Eq. (12) defines M := |F| ≤ θ, so a lower θ retains fewer high-frequency components; the interpretation is reversed.","section":"Table V"},{"comment":"The loss in Eq. (17) is written as a full-vocabulary binary cross-entropy over all items and all time steps, which is not the standard softmax cross-entropy used in sequential recommendation and has different optimization and computational properties. The text calls it cross-entropy without specifying the output activation (e.g., sigmoid vs softmax). Additionally, Algorithm 1 declares the output as (B, T, 1), which is inconsistent with producing a distribution over the item vocabulary. Please clarify the actual loss, the prediction head, and how the large vocabulary is handled.","section":"Section III-E, Eq. (17) and Algorithm 1"}],"minor_comments":[{"comment":"The abstract claims 20% faster inference than Transformer baselines, but Section IV-E reports only qualitative statements and plots without a quantitative speedup value or standard-deviation information; please provide the actual measured numbers.","section":"Abstract and Section IV-E"},{"comment":"The phrase 'through Long Short-Term Memories (LLMs)' appears to be a typo; it should be 'Large Language Models (LLMs)'.","section":"Section III-E"},{"comment":"The text says results are 'presented in Figure IV', but the reference should be to Table IV.","section":"Section IV-D"},{"comment":"The DFT uses N in the exponential but sums to T; please use one consistent length symbol throughout the frequency-domain equations.","section":"Equation (11) and surrounding text"},{"comment":"The parameter list repeats 'SSM state expansion factor' twice; also the descriptions of the expansion factor, state dimension, and convolution width should be clarified.","section":"Section IV-A Parameter Settings"},{"comment":"The phrase 'these approaches remain constrained on periodity caption' is grammatically unclear and should be rewritten.","section":"Section V-A"}],"recommendation":"major_revision","confidential_remarks":"The paper has a promising empirical direction, but the current version needs substantial corrections. The copied rows in Table V, the dataset-statistics mismatch, and the loss/output specification are concrete issues that a careful revision can address. The central periodic-pattern claim, however, requires either a change to the architecture (e.g., using timestamps) or a significant reframing of the contribution. I recommend a major revision rather than rejection because the empirical question—whether the proposed AFFT-LLM-gate combination helps—remains potentially valuable and testable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take. The paper proposes a concrete architecture combo—Mamba with an adaptive FFT, LLM item embeddings, and a gated fusion—for sequential recommendation. That exact combination is new, and the reported gains over Mamba4Rec/RecMamba are consistent across four datasets. The efficiency claim also holds up: adding FFT to Mamba keeps the same asymptotic complexity, and the LLM embeddings are a one-time precomputation. The ablation suggests each component contributes.\n\nThe biggest problem is the periodic-pattern story. Eq. 11 applies the DFT over sequence positions t, and the model never receives timestamps. Fig. 1 plots daily/weekly cycles in days, but the FFT in the model is over interaction indices, so a spectral peak at k/T is cycles per interaction, not cycles per day. The headline 'multi-scale periodic user behavior' mechanism is therefore not operational in the architecture as written. The FFT might still help as a sequence smoother or feature transformer, but the paper's own motivation is unsupported.\n\nSecond, the 'gate mechanism' is just a scalar weighted sum of two embeddings (alpha and beta). That's not adaptive multi-modal fusion; it's a linear combination. Minor relative to the first point, but the language overstates it.\n\nThird, the experimental tables have inconsistencies: Texas is quoted as 27.2M interactions in Sec. IV-A but Table III says 1.34M; New York's average density doesn't match its user/item counts; Table V's Mamba-layer rows are identical to the theta rows. These are likely copy-paste errors, but they make the reported numbers hard to trust without code. The anonymous code link has no commit hash, so verification is limited.\n\nNet: the empirical claim is plausible but unproven as presented. The position-vs-time FFT issue is load-bearing for the paper's narrative, though the architecture could still work for other reasons. A serious referee should be sent in—not to rubber-stamp, but to require the authors to fix the periodic-pattern rationale, correct the tables, and clarify what the gate actually does. This is not a desk-reject; it's a major-revision candidate.","headline":"Plausible architecture combo with a real positional-FFT mechanic mismatch; the reported gains are worth checking but the periodic-story and data tables need fixing.","tokens_in":18366,"tokens_out":3378,"would_cite":false,"duration_ms":30684,"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":"This paper claims that M2Rec—a Mamba-based sequential recommender augmented with an adaptive Fourier filter, fixed LLM text embeddings, and a learned gate—beats existing Mamba-based and Transformer-based models on next-item prediction…","keywords":["sequential recommendation","Mamba","state space models","Fast Fourier Transform","LLM embeddings","adaptive gating","multi-scale periodic patterns","noise filtering"],"falsifier":"Train M2Rec and an otherwise identical variant that also receives timestamps or inter-arrival gaps on ML-1M. If the position-only FFT truly captures daily and weekly behavior, adding time information should not change Hit Rate@10; if it improves, the paper's stated periodic mechanism is not grounded in the model's actual input. A second check: keep item order fixed but replace all timestamps with random ones; the position-only model must be unaffected by construction, while a time-aware model should degrade.","tokens_in":17351,"feed_emoji":"📈","tokens_out":6415,"duration_ms":61919,"temperature":0.7,"pith_summary":"The paper proposes M2Rec, a sequential recommendation model that tries to fix three weaknesses it attributes to existing Mamba-based recommenders: they miss periodic user behavior, ignore item-text semantics, and fuse multimodal features poorly. To fix this, it runs the user's item sequence through an adaptive Fast Fourier Transform before the Mamba state-space layer (filtering high frequencies as noise), supplements the item-ID embeddings with fixed embeddings from a general-purpose text-embedding LLM, and combines the two branches through a learned gate. The paper claims that on ML-1M and three location datasets this combination reaches state-of-the-art results, improving Hit Rate@10 by 3.2% over prior Mamba-based models and running about 20% faster at inference than Transformer baselines. A sympathetic reader would care because, if right, it shows Mamba can be augmented cheaply to capture periodic and semantic structure that pure ID-sequence models miss, without giving up linear-time efficiency.","feed_headline":"Fourier and LLM text push Mamba recommenders past baselines","feed_subtitle":"Adaptive FFT filters noise, LLM embeddings add semantics, and a gate fuses both for state-of-the-art next-item prediction.","key_machinery":"The central object is the AFFM block (Adaptive Fast Fourier Transform powered Mamba), which implements Equations 11 through 15. The user's item sequence $X_u[t]$ is Fourier-transformed over positions $t$, multiplied by a binary mask that keeps frequencies below threshold $\\theta$, multiplied by learned complex weights $W$, inverse-transformed (IFFT) back to the time domain, and then processed by the Mamba SSM with SiLU gating. This module grounds the paper's claims about periodic-pattern capture and noise reduction. The second mechanism is the gate fusion in Equation 16, which forms the prediction as a linear map of $\\alpha$ times the LLM embeddings plus $\\beta$ times the Mamba output; the paper treats $\\alpha$ and $\\beta$ as adaptable parameters.","core_discovery":"The central claim is that M2Rec outperforms both Transformer-based and Mamba-based sequential recommenders across four datasets (ML-1M, New York, California, Texas), with the best Hit Rate@10 of 0.3224 on ML-1M, a 3.30% relative improvement over the strongest Mamba baseline, and analogous gains in NDCG@10 and MRR@10. The paper attributes the gains to three components: the adaptive FFT (AFFT) module, which transforms the item-position signal into the frequency domain, masks high frequencies, multiplies by a learned kernel, and transforms back before entering the Mamba SSM; fixed LLM embeddings of item names, categories, and descriptions; and a gate that balances the two embedding streams. The paper also claims the FFT branch makes the model more robust to injected Gaussian noise and that the gate makes performance stable across different LLM embedding backbones.","pith_inferences":["Because the FFT in Equation 11 is applied to position index $t$, not to wall-clock time, the paper's daily- and weekly-cycle story is really about position-index frequencies; a direct test would be to feed the model actual timestamps or inter-arrival gaps and see whether the periodic mechanism still helps.","The gate in Equation 16 appears to use scalar parameters $\\alpha$ and $\\beta$ shared across users and positions; a per-user or per-item gate would be a natural extension if context-dependent balancing is the goal.","The LLM embeddings are fixed (bge-large-en-v1.5), so the paper's results speak to content enrichment rather than LLM reasoning; using a trainable or instruction-tuned LLM encoder could change the accuracy-efficiency trade-off.","If the FFT's noise filtering is the source of robustness, the same adaptive filter could be applied to other sequence encoders besides Mamba—a testable transfer to attention-based recommenders."],"forward_implications":["If the central claim is right, Mamba-based sequential recommenders can capture periodic user behavior by preprocessing sequences in the frequency domain, so future work need not replace Mamba with attention to model periodicity.","The gate mechanism makes the model robust to which LLM produces the text embeddings, so teams can swap embedding backbones without retuning the rest of the model.","The adaptive FFT acts as a noise filter, so M2Rec should degrade more gracefully than Mamba4Rec or BERT4Rec when embeddings are corrupted.","Because the FFT adds only $O(BTd\\log T)$ time, the approach remains practical for long user histories where Transformers' quadratic attention is prohibitive."],"supporting_citations":[{"why":"Supplies the selective state-space backbone that M2Rec extends with FFT and gate; the efficiency claim depends on Mamba's linear complexity.","marker":"[16]"},{"why":"Mamba4Rec is the primary Mamba-based recommendation baseline M2Rec must beat, and the motivating example of a Mamba recommender that ignores periodic patterns and semantics.","marker":"[17]"},{"why":"RecMamba is the second Mamba-based baseline, representing lifelong sequential recommendation, and the claimed 3.2% Hit Rate@10 improvement is measured against this class.","marker":"[18]"},{"why":"BERT4Rec is the Transformer baseline used for efficiency comparisons and long-sequence prediction experiments.","marker":"[14]"},{"why":"FEARec is the frequency-domain attention baseline whose design M2Rec contrasts with its AFFT-in-Mamba approach.","marker":"[38]"},{"why":"Provides the fixed bge-large-en-v1.5 text embeddings that form the semantic branch of M2Rec.","marker":"[27]"}],"fun_headline_variants":["FFT-gated Mamba beats Transformers in recommendation","Multi-scale Mamba fuses frequency, semantic cues for next-item","M2Rec: Faster Mamba with Fourier and LLM embeddings","Periodic patterns via FFT boost Mamba recommenders","Adaptive gate balances Mamba, FFT, and LLM features"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that Fourier-transforming the item order in a user's history captures genuine daily and weekly behavior cycles, even though the model never sees timestamps; if user interactions are irregularly spaced in time, that premise fails and the periodic-pattern explanation for the gains collapses.","fun_headline_variants_meta":{"raw":{"variants":["FFT-gated Mamba beats Transformers in recommendation","Multi-scale Mamba fuses frequency, semantic cues for next-item","M2Rec: Faster Mamba with Fourier and LLM embeddings","Periodic patterns via FFT boost Mamba recommenders","Adaptive gate balances Mamba, FFT, and LLM features"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000203,"raw_usage":{"total_tokens":1404,"prompt_tokens":985,"completion_tokens":419,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":601,"completion_tokens_details":{"reasoning_tokens":331}},"tokens_in":601,"tokens_out":419,"duration_ms":3977,"temperature":1.0,"reasoning_tokens":331,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:28:24.778907+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train M2Rec and an otherwise identical variant that also receives timestamps or inter-arrival gaps on ML-1M. If the position-only FFT truly captures daily and weekly behavior, adding time information should not change Hit Rate@10; if it improves, the paper's stated periodic mechanism is not grounded in the model's actual input. A second check: keep item order fixed but replace all timestamps with random ones; the position-only model must be unaffected by construction, while a time-aware model should degrade.","supporting_citations":[{"cited_title":"Uncovering selective state space model’s capabilities in lifelong sequential recommendation,","cited_arxiv_id":null,"evidence_quote":"RecMamba is the second Mamba-based baseline, representing lifelong sequential recommendation, and the claimed 3.2% Hit Rate@10 improvement is measured against this class."},{"cited_title":"Frequency enhanced hybrid attention network for sequential recommendation,","cited_arxiv_id":null,"evidence_quote":"FEARec is the frequency-domain attention baseline whose design M2Rec contrasts with its AFFT-in-Mamba approach."}],"review_version":1}