{"id":"f6473104-8560-4620-948e-ce63e88f30ef","arxiv_id":"2506.02916","paper_version":5,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"MMM4Rec uses a time-aware state space model with shared cross-modal projections and Fourier-filtered fusion to make multi-modal sequential recommenders transfer faster and more accurately.","lead":"MMM4Rec applies state space duality (Mamba-style temporal decay) to multi-modal sequential recommendation, aligning text and image features with shared projections and fusing them through a new cross-SSD block with Fourier filtering. The paper reports better retrieval accuracy and roughly 10x faster fine-tuning convergence than two transferable baselines on Amazon domains.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"TiCoSSD fusion is undefined as written: complex bD_f from eq. (16) is fed into real-valued D, Δ, mask, and Softplus operations without any real-extraction rule; the core fusion module is not reproducible.","rationale":"The reader's weakest_assumption identifies the same point I would: the Cross-SSD fusion equation is the linchpin of the architecture, and its undefined complex-to-real step is a genuine correctness gap, not a stylistic issue. I do not see a more load-bearing objection. The convergence-efficiency claim, while dependent on early-stopping protocol and missing error bars, is supported by appendix curves and Table 4; the fine-tuning loss typo in eq. (22) is almost certainly a typo because the numerator uses the correct target item; the missing-image handling is unspecified but follows MISSRec's convention and has a full-modality appendix check. The strongest alternative concern—that the 'algebraic constraints' are just weight sharing plus a time-aware mask and the novelty is incremental—affects significance, not correctness. Because the undefined fusion step can be resolved by inspecting the released code or adding one sentence, a CONDITIONAL verdict is right; if the code turns out to contain no real-extraction and a corrected forward pass changes the results, the verdict should move to REJECT. My recommendation is therefore UNCHANGED: keep CONDITIONAL pending the eq. (16)-(18) fix.","tokens_in":22366,"tokens_out":6520,"duration_ms":78586,"concrete_test":"Inspect and execute the released code at github.com/AlwaysFHao/MMM4Rec: (1) compute bD_f from eq. (16) as a complex tensor, then run the TiCoSSD forward pass for one batch; if the forward pass errors or requires an undocumented .real()/.abs() conversion, the equations are incomplete. (2) If code runs, locate where the complex-to-real conversion occurs and re-derive eqs. (16)-(18) accordingly; then re-run Table 2 row Scientific with that corrected definition to confirm the 6.23% R@50 improvement is preserved. (3) Check eq. (18) dimensions by printing Δ.shape and X.shape in the code; if Δ^T X is not L×N, the paper's formula is inconsistent with the implementation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central empirical claim depends on TiCoSSD (eqs. 16-18), and that module is not mathematically well-defined as written. Eq. (16) produces a complex vector bD_f ∈ C^L from the inverse FFT of filtered spectra. The next step substitutes this bD_f for the real time-difference vector D in the TiSSD computations of §3.3.1. That substitution breaks immediately: eq. (5) would make bD complex; eq. (6) applies Softplus to Δ·bD, yielding a complex Δ; eq. (7) then makes A and B complex; eq. (8) builds a complex mask L; and eq. (18) outputs a complex M. The paper never says to take real parts, magnitudes, or any other projection, and eq. (19) then feeds M through LayerNorm and FFN, which are not defined for complex tensors in the stated framework. Eq. (18) also has a dimension problem: with Δ an L-vector, Δ^T X is 1×N, not L×N, so even the real-valued version is not consistent as written. Because this is the fusion stage that produces the user representation used in both losses (eqs. 21-22), the reported SOTA and 10× convergence results are not tied to a reproducible model as specified. This is fixable—e.g., by defining bD_f := real(F^{-1}(·)) or a learned real projection—but the fix must be stated and verified.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MMM4Rec, a multi-modal sequential recommendation framework built on State Space Duality (SSD) / Mamba-style temporal modeling. The method consists of (i) frozen pretrained image/text encoders with lightweight adapters, (ii) sequence-level cross-modal alignment through weight-shared Time-aware SSD (TiSSD) modules, and (iii) a cross-modal fusion stage (TiCoSSD) that combines dual-channel Fourier filtering with a structurally adapted SSD. Both pre-training and fine-tuning use a simple cross-entropy objective with in-batch and full-corpus negative sampling, respectively. The central empirical claim is that MMM4Rec achieves state-of-the-art retrieval performance on five downstream Amazon domains and converges about 10x faster than the transferable baselines MMSRec and MISSRec, with ablations supporting the contributions of the time-aware, weight-sharing, and Fourier-filtering components. The paper also reports robustness in ID-removed and modality-missing settings.","tokens_in":22746,"tokens_out":5348,"duration_ms":59768,"significance":"If the mathematical definitions are corrected, the paper would offer a substantial empirical study: it compares against 14 baselines across 10 Amazon domains, includes both with-ID and ID-removed settings, provides ablation studies, full convergence curves, and an implementation link. The design goal of avoiding contrastive auxiliary losses and using a single cross-entropy objective is a noteworthy practical strength, and the convergence-speed results, if reproducible, would be valuable for transfer-efficient recommendation. However, the core TiCoSSD fusion module is not mathematically well-defined as written: a complex-valued signal from the inverse FFT is substituted into real-valued time-gap operations, and eq. (18) is dimensionally inconsistent. Because this module produces the user representation used in both losses, the headline results cannot currently be traced to a reproducible model. The issues are local and fixable, but they are load-bearing for the paper's central claim.","major_comments":[{"comment":"The output bD_f of the inverse FFT is defined in C^L, but in §3.4.2 it is substituted for the real time-difference vector D of §3.3.1, where it feeds Softplus in eq. (6), mask construction in eqs. (7)–(8), and the SSD output in eq. (9). LayerNorm and FFN in eq. (19) are then applied to the resulting M. No real-part, magnitude, or learned projection is specified, and none of these operations is defined for complex tensors as written. Please state the exact real-extraction rule and verify that all subsequent operations are real-valued.","section":"§3.4.1–3.4.2, Eq. (16)"},{"comment":"With Δ defined as an L-vector in eq. (17), the expression Δ̂^T X has shape 1×N, not L×N, so the equation is dimensionally inconsistent even if all quantities are real. Please clarify whether the intended operation is elementwise gating (for example Δ̂[:, None] * X) or a different row-wise scaling, and correct the formula accordingly.","section":"§3.4.2, Eq. (18)"},{"comment":"The fine-tuning loss denominator sums exp(<u_j, i_j> / τ) over j=1..|I|; full-corpus ranking for user k should instead sum exp(<u_k, i> / τ) over candidate items i ∈ I. As written, the denominator pairs item j with user j and is not a proper softmax over candidates for u_k. This typo changes the training objective and must be fixed.","section":"§3.5.3, Eq. (22)"},{"comment":"The optional item modality bias is implemented as two matrices E_t, E_v ∈ R^{|I|×N} added to item features; this is exactly an ID embedding table of size |I|×N per modality, so the statement that the module 'bypasses the dimensionality explosion of explicit ID features' is not supported. Please revise this claim or clarify what dimensionality is actually being bypassed.","section":"§3.2.3"}],"minor_comments":[{"comment":"D is defined as [0, d_1, ..., d_{T-1}] with d_l = t_{l+1} - t_l for l ∈ [1, T]; a sequence of length L has only L-1 gaps, so the index range should be stated consistently as l ∈ [1, L-1] to avoid an off-by-one ambiguity.","section":"§3.3.1, Eq. (5)"},{"comment":"The notation Q_t = P_t * ω uses Q_t and P_t that are not defined; please define the causal convolution input and output variables explicitly.","section":"§3.3.1, Eq. (4)"},{"comment":"The heading 'Comparasion' is a typo and should read 'Comparison'.","section":"§4.2 heading"},{"comment":"In the Movies row, the number of items is printed as '59.203' with a period instead of a comma; please correct the formatting.","section":"Table 1"},{"comment":"The annotations 'Score =2.4%' and 'Score =3.6%' in the caption are not explained in the text; please clarify what these percentage values represent.","section":"Fig. 2"},{"comment":"Since §3.2.3 describes the item modality bias as optional, please state explicitly whether the main results in Tables 2 and 3 use the bias or the ID-removed variant, and how the switch between the two is handled in the reported numbers.","section":"Tables 2 and 3"},{"comment":"Reference [7] is the authors' own TiM4Rec paper; the temporal backbone is therefore not independently introduced in this manuscript and should be attributed more prominently in §3.3.1.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The empirical study is extensive and the convergence-efficiency comparison is a useful addition, but the undefined TiCoSSD fusion module and the dimension mismatch in eq. (18) prevent acceptance in the current form. The fixes are local: specify a real projection after the inverse FFT, correct eq. (18) and eq. (22), and revise the claim in §3.2.3. I would be willing to review a revised version."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a solid empirical paper with a real efficiency story, but the central Cross-SSD fusion block is not mathematically well-defined as printed, and until that is fixed the reported SOTA numbers are attached to an unreproducible model.\n\nWhat's genuinely new here is the specific combination: weight-shared TiSSD alignment followed by Cross-SSD fusion with dual-channel Fourier filtering, all trained with a single cross-entropy objective. That combination does not appear in the cited prior work, and the paper takes care to compare against Mamba-based multi-modal baselines (M3Rec, HM4SR, ATHWE) that it beats on several metrics. The experiments are extensive: five pre-training domains, five downstream domains, ID-removed variants, a sensible ablation study, convergence curves in the appendix, and a clean efficiency table. If the model works as described, the 10x faster convergence claim is plausible and practically useful.\n\nThe soft spots are real and need attention. The load-bearing one: eq. (16) produces a complex-valued bD_f, and eq. (18) says to substitute it for the real D in the TiSSD computations of Section 3.3.1. That substitution is undefined. Equations (5)-(7) would make the time-gap, the discretization step, and the state matrices complex, and eq. (19) then feeds a complex tensor into LayerNorm and FFN. There is no stated real-extraction rule. On top of that, eq. (18) has a dimension inconsistency: with Δ an L-vector and X an L×N matrix, Δ^T X is 1×N, not L×N. Both are fixable — e.g., define bD_f as the real part of the inverse FFT and fix the transpose — but they have to be stated and verified.\n\nTwo smaller issues. The fine-tuning loss in eq. (22) has a denominator that sums exp(<u_j, i_j>/τ) over j, which is not the standard cross-entropy negative log-likelihood; likely a typo, but as written it is a different objective. And the paper says it retains modality-missing items following MISSRec, but never specifies how the model handles an item with no image. That matters because Kindle has 0% image coverage. The significance claims would also be stronger with error bars, but that's minor.\n\nOverall: the architecture idea is reasonable, the experiments are above average for this area, and the flaws are mostly omissions and typos rather than deep conceptual errors. This paper deserves a serious referee, but only with a request for major revision on the math. I would not cite it in its current form; I would revisit after the authors fix the fusion equations.","headline":"A promising but not-yet-reproducible multi-modal sequential recommender whose core fusion equation feeds a complex signal into real-valued operations without a projection rule; the empirical work is otherwise solid and the fix looks mechanical.","tokens_in":23265,"tokens_out":2246,"would_cite":false,"duration_ms":26920,"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":"One architecture uses state-space decay and shared projections to make multi-modal sequential recommendation transfer about ten times faster.","keywords":["multi-modal sequential recommendation","transfer learning","state space duality","state space models","time-aware modeling","Fourier filtering","cross-modal alignment","cross-entropy objective"],"falsifier":"In the released code, inspect the tensor feeding the Cross-SSD equation: if the complex-valued fused signal from the inverse FFT is never converted to a real value before being used in the time-difference equations, the central fusion equation is not reproducible; additionally, re-running fine-tuning from the same pretrained checkpoint while replacing that fused signal with its real part would change the convergence curve if the Fourier fusion is genuinely load-bearing.","tokens_in":22152,"feed_emoji":"⚡","tokens_out":7607,"duration_ms":76500,"temperature":0.7,"pith_summary":"The paper tries to show that transferable multi-modal sequential recommendation can be made dramatically cheaper by replacing complex training objectives with algebraic constraints inside the model. Its proposed framework, MMM4Rec, uses State Space Duality (SSD)—a structured sequence model whose state-decay mask acts like linear attention biased toward recent items—to align image and text features and then fuse them temporally. Both pre-training and fine-tuning use only a plain cross-entropy loss, with no contrastive or auxiliary losses. On five Amazon domains, the authors report state-of-the-art retrieval and roughly 10 times faster average convergence when transferring to large downstream datasets. The argument is that SR-compatible structural constraints, not optimization gymnastics, are what make multi-modal transfer efficient.","feed_headline":"One recommender cuts fine-tuning time to a tenth","feed_subtitle":"A single cross-entropy loss plus state-space decay aligns images and text for fast transfer to new domains.","key_machinery":"The load-bearing object is the Time-aware Cross SSD (TiCoSSD) module—a reconfiguration of State Space Duality (SSD), a sequence model equivalent to linear attention with a state-decaying mask, in which the structured mask is built from inter-item time differences, the C matrix comes from the visual sequence, the B and X matrices come from the text sequence, and a pair of Fourier-domain filters (an adaptive per-modality filter and a shared learnable filter) merge the two modalities' time signals before the mask is applied. A companion Time-aware SSD (TiSSD) with shared weights performs sequence-level alignment before fusion. These modules together impose the algebraic constraints: temporal decay for recency bias, a refined time-aware mask to keep early critical items, and shared projections for cross-modal semantic consistency.","core_discovery":"MMM4Rec's central claim is that a two-stage alignment-then-fusion design, constrained by weight sharing and SSD's temporal-decay mask, is enough to make multi-modal sequential recommenders transfer efficiently. In the alignment stage, visual and textual sequences pass through the same time-aware SSD module with shared projection weights, pushing both modalities into a common recommendation semantic space. In the fusion stage, a Cross-SSD block derives the query-like matrix C from one modality and the key/value-like matrices B and X from the other, while a dual-channel Fourier filter combines the two modalities' time-difference signals into one fused signal that enters the mask. The same cross-entropy loss is used for pre-training and fine-tuning, and the authors report that this recipe beats prior transferable recommenders in retrieval and reaches strong downstream performance in a fraction of the epochs.","pith_inferences":["If the convergence-speed result generalizes beyond these five domains, the design suggests that architectural priors such as temporal decay and weight tying could substitute for elaborate pretraining objectives in other transferable recommender systems.","The Fourier-domain fusion of time-difference signals is a modular idea that could be lifted into other state-space or linear-attention recommenders that already use time-dependent masks.","A natural next test is whether the same alignment-then-fusion recipe extends to more than two modalities, since the paper only combines image and text.","One could probe the boundary of the claim by applying MMM4Rec to domains with very short interaction sequences, where SSD's recency bias provides less advantage and the Fourier fusion must carry more of the load."],"forward_implications":["The model reaches state-of-the-art retrieval on the five downstream domains, with the largest reported gain being a 31.78% NDCG@10 improvement on the Pantry domain.","Fine-tuning converges about 10 times faster on average than the two main transferable baselines, requiring fewer epochs and less time per epoch on item-rich domains.","Transfer works with a single cross-entropy objective in both phases, so downstream adaptation needs no contrastive sampling or auxiliary losses.","The ID-removed variant stays competitive, meaning the learned multi-modal sequence representation itself carries transferable signal without item identifiers.","Ablation results tie every component—time awareness, shared-weight alignment, adaptive filter, learnable filter—to the reported performance, since removing any one degrades the metrics."],"supporting_citations":[{"why":"Supplies the State Space Duality formalism, including the structured mask and the linear-attention interpretation that MMM4Rec's TiSSD and TiCoSSD modules are built on.","marker":"[3]"},{"why":"Provides the Time-aware SSD variant whose equations MMM4Rec adopts for temporal state-space sequence modeling.","marker":"[7]"},{"why":"Supplies the causal convolution and zero-order-hold discretization steps used in the TiSSD and TiCoSSD transforms.","marker":"[9]"},{"why":"MISSRec is the main transferable multi-modal baseline and the parameter-efficient adapter design MMM4Rec follows and compares against in convergence and accuracy.","marker":"[41]"},{"why":"MMSRec is one of the two transfer-efficiency baselines in the efficiency table, representing contrastive multi-modal transfer that MMM4Rec claims to outperform.","marker":"[33]"},{"why":"Provides the frozen cross-modal image/text encoder that generates the multi-modal features used as input to the modality adapters.","marker":"[46]"},{"why":"Amazon Reviews is the source of the ten domains used for pre-training and downstream transfer evaluation.","marker":"[26]"}],"fun_headline_variants":["State-space duality makes multi-modal recommender transfer 10x faster","One loss, two stages: multi-modal recommender transfers 10x quicker","Cross-modal alignment plus state-space decay speeds fine-tuning 10x","Same loss for pretrain and finetune: multi-modal transfer 10x faster"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire fusion stage depends on treating the complex-valued cross-modal time signal from the inverse Fourier transform as a real time-difference input to the state-space mask, but the paper states no rule for extracting that real signal.","fun_headline_variants_meta":{"raw":{"variants":["State-space duality makes multi-modal recommender transfer 10x faster","One loss, two stages: multi-modal recommender transfers 10x quicker","Cross-modal alignment plus state-space decay speeds fine-tuning 10x","Same loss for pretrain and finetune: multi-modal transfer 10x faster"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000719,"raw_usage":{"total_tokens":3238,"prompt_tokens":961,"completion_tokens":2277,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":577,"completion_tokens_details":{"reasoning_tokens":2196}},"tokens_in":577,"tokens_out":2277,"duration_ms":16201,"temperature":1.0,"reasoning_tokens":2196,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T11:13:32.103739+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"In the released code, inspect the tensor feeding the Cross-SSD equation: if the complex-valued fused signal from the inverse FFT is never converted to a real value before being used in the time-difference equations, the central fusion equation is not reproducible; additionally, re-running fine-tuning from the same pretrained checkpoint while replacing that fused signal with its real part would change the convergence curve if the Fourier fusion is genuinely load-bearing.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the State Space Duality formalism, including the structured mask and the linear-attention interpretation that MMM4Rec's TiSSD and TiCoSSD modules are built on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Time-aware SSD variant whose equations MMM4Rec adopts for temporal state-space sequence modeling."},{"cited_title":"https://openreview.net/forum? id=SygXPaEYvH","cited_arxiv_id":null,"evidence_quote":"MISSRec is the main transferable multi-modal baseline and the parameter-efficient adapter design MMM4Rec follows and compares against in convergence and accuracy."}],"review_version":1}