{"id":"ae137b7b-ed45-401c-a760-623c9d904fe9","arxiv_id":"2412.08979","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Wander is a low-rank sequence adapter that fuses token-level features across an arbitrary number of modalities with far fewer parameters than full fine-tuning.","lead":"The paper introduces Wander, an adapter module that fuses multiple modalities at the level of individual tokens using low-rank tensor decomposition, and reports that it improves accuracy while using fewer trainable parameters than existing adapters on four multimodal benchmarks. A generalist reader might care because parameter-efficient fine-tuning of large multimodal models is a practical bottleneck, and the method claims to handle any number of modalities.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline 'consistent SOTA outperformance' rests on single-run point estimates; reported margins are small enough that the claim may not survive seed-to-seed noise.","rationale":"I agree with the CONDITIONAL verdict but not exactly with the reader's stated weakest assumption. The rank-8 CP assumption is a legitimate concern, and the paper provides only empirical support for it; however, the more decisive gap is that the empirical support itself has no statistical grounding. Without error bars, neither the headline comparisons nor the rank-sensitivity plots can be reliably interpreted. This is why I set agreement_with_reader to 'partial'. I do not see an internal mathematical contradiction in the main derivation: Eq. 13 is a plausible factorized form of the sequence-level interaction, and the parameter-count reduction follows from it, modulo some loose notation about outer products of matrices. The missing LMF baseline is also relevant and should be added, but the single check that would settle whether the central claim lands is repeated-seed evaluation with significance testing. If the claimed margins survive that test, the verdict should stand; if not, the claim needs to be weakened.","tokens_in":13231,"tokens_out":10480,"duration_ms":107755,"concrete_test":"Run each configuration in Tables 1-4 with at least five independent random seeds, using identical frozen backbones, evaluation scripts, and hyperparameters; report mean and standard deviation, and apply a paired significance test (paired bootstrap or Wilcoxon) comparing Wander against the strongest baseline in each table (e.g., UniAdapter on UPMC-Food 101, MaPLe on CMU-MOSI, full fine-tuning on IEMOCAP/MSRVTT). If the differences driving the headline are not significant at p<0.05 or the 95% confidence interval includes zero, the 'consistent outperformance' claim is not established. Releasing the training and evaluation code would make this check auditable.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Tables 1-4 are the entire empirical basis for the central claim that Wander 'outperforms state-of-the-art efficient transfer learning methods consistently.' Every number is a single point estimate: no random seeds, confidence intervals, or significance tests are reported. Many of the decisive margins are small. On UPMC-Food 101 (Table 1), Wander(d=64) is 91.1 vs UniAdapter 90.8 ACC, and Wander(d=128) ties full fine-tuning at 91.8. On CMU-MOSI (Table 3), Wander(d=64) leads ACC-2 83.2 vs MaPLe 82.1, but F1 is 82.9 vs 82.0 and ACC-7 33.6 vs 31.2; on MSRVTT (Table 4), Wander(d=16) ties or edges full fine-tuning by 0.1-0.2. On IEMOCAP (Table 2), Wander(d=64) ACC is 74.7 vs full fine-tuning 74.8, a 0.1 deficit. For datasets of this size, run-to-run standard deviations of 0.5-1.5 points are typical, so none of these margins is clearly outside noise. The rank sensitivity analysis in Figure 5 suffers from the same problem: apparent differences between ranks of 4-12 may be within noise. The absence of released code compounds the issue by preventing independent reproduction. The CP rank-8 approximation is not the most load-bearing worry; even if rank-8 is adequate, the empirical validation currently cannot distinguish the method from a lucky run.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes Wander, a parameter-efficient adapter for fine-tuning multimodal models with an arbitrary number of modalities. The authors first observe that existing multimodal adapters are limited to vision-language pairs and to vector-level fusion. They then propose a sequence-level fusion that applies an outer product to token sequences and factorizes the resulting high-order interaction tensor with a low-rank decomposition. The core derivation in Section 3.3, Eq. (13), shows that the materialization of the full outer product can be avoided, yielding a parameter count that grows linearly in the number of modalities. Experiments on UPMC-Food 101, CMU-MOSI, IEMOCAP, and MSRVTT report that Wander outperforms several efficient transfer learning baselines and, in some cases, matches or exceeds full fine-tuning.","tokens_in":13546,"tokens_out":14385,"duration_ms":137328,"significance":"If the results are reproducible, the paper makes a useful contribution: it generalizes multimodal efficient transfer learning beyond two modalities, introduces a token-level interaction mechanism that is more fine-grained than vector-level fusion, and provides a factorized form (Eq. 13) that avoids explicitly constructing the high-order tensor. The algebraic reduction is a genuine strength, and the paper evaluates on datasets with 2, 3, and 7 modalities, which is broader than many adapter papers. The parameter-efficiency argument is plausible for large numbers of modalities. However, the empirical support for the headline claim is currently weakened by the lack of repeated runs and by inconsistencies in the reported parameter counts, and the terminology 'CP decomposition' needs to be aligned with the actual factorization used.","major_comments":[{"comment":"The method is described as using CP decomposition, but the factors w^r_{h,m} in R^{dh x dm} are matrices, not vectors, and the equality W_h = sum_r tensorprod_m w^r_{h,m} does not correspond to a standard CP decomposition of the (M+1)-way tensor W_h. The construction is a per-output-slice rank-R factorization, with parameter count R * dh * sum_m d_m rather than the CP count R * (dh + sum_m d_m). Please correct the terminology and complexity statements, and clarify the meaning of Rh and Rt in Figure 5 in light of this distinction.","section":"Section 3.3, Eqs. (5)-(6)"},{"comment":"The reported tunable-parameter counts for the same Wander configuration are inconsistent. Table 3 lists Wander(d=16) with 0.3M parameters on CMU-MOSI, whereas Table 5 lists Wander(d=16) with 0.1M; Table 5 also lists Wander(d=32) with 0.1M, equal to d=16, and the '-nonlinearity' and '-residual' rows with 0.2M, which is larger than the complete model. Please unify the counting convention (adapter-only vs. including the prediction head) and correct the numbers, since the parameter-efficiency claim depends on them.","section":"Section 4.2, Table 3, and Section 4.3, Table 5"},{"comment":"The claim that Wander 'outperforms state-of-the-art efficient transfer learning methods consistently' rests on single point estimates. Several decisive margins are small: on UPMC-Food 101, Wander(d=64) is 91.1 vs. UniAdapter 90.8; on IEMOCAP, Wander(d=64) is 74.7 vs. full fine-tuning 74.8; on MSRVTT, Wander(d=16) ties full fine-tuning on Text-to-Video R@5 at 57.2. For datasets of these sizes, 0.1-1.0 point differences are typically within seed-to-seed noise. Please report means and standard deviations over multiple seeds, or confidence intervals, for at least the main tables, and temper the consistency claim accordingly.","section":"Section 4.2, Tables 1-4"},{"comment":"The rank sensitivity analysis is the only evidence for the default choice Rh=Rt=8, but it is based on single runs and appears to show non-negligible variation between ranks 4 and 12, particularly for Rt. The statement that 'the value of the rank will not affect the performance significantly' is not supported by the data as presented. Please provide repeated runs for the rank sweep and, ideally, a measure of the approximation error of the factorized interaction tensor, to justify the default rank.","section":"Section 4.5, Figure 5"}],"minor_comments":[{"comment":"The tensor contraction and the 'transpose' of a multi-index expression are not defined; please specify the index order for the tensor product and for the transposition to make the derivation unambiguous.","section":"Section 3.3, Eq. (13)"},{"comment":"SF-VF is described as 'the vector fusion form of sequence fusion in Equation 8', but Equation 8 is the per-token vector fusion; please clarify what sequence-level aggregation is used for the SF-VF baseline.","section":"Section 4.3 and Table 5"},{"comment":"The x-axis labels appear garbled and the values are not legible; please replace the figure with one having clear axis labels and units.","section":"Figure 4"},{"comment":"The modifications to UniAdapter, MaPLe, and PMF for the non-vision-language backbones and the more-than-two-modality settings are described only briefly; please provide details on how each baseline was adapted to ensure a fair comparison.","section":"Section 4.2"},{"comment":"It is unclear whether the GPU time and memory figures are for the reduced-dimension settings described in Section 4.3, and the FLOPs unit is not stated; please clarify the experimental conditions for the cost comparison.","section":"Section 4.4 and Table 6"},{"comment":"The paper does not mention code release; given the parameter-count inconsistencies and the small performance margins, releasing code and training configurations would substantially aid reproducibility.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The core idea is interesting and the algebraic derivation in Eq. (13) is largely sound, but the empirical evaluation needs substantial strengthening (multiple seeds, error bars, consistent parameter counting). I recommend major revision. Please also have the tensor-factorization terminology checked by someone familiar with CP decomposition, since the factorization in Eq. (6) is not a standard CP decomposition of the full weight tensor."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe genuinely new piece here is the sequence-level low-rank fusion in Equation 13—applying the LMF trick from Liu et al. 2018 to token sequences instead of single vectors, so the full multimodal outer product never has to be built. The derivation is clean and the method is simple, and the paper is upfront about its LMF roots. That is a real, if incremental, contribution to parameter-efficient multimodal transfer learning.\n\nWhat it does well: the algebra in Section 3.3 checks out, the ablations in Table 7 and Table 5 are sensible, and the cost analysis in Section 4.4 is honest about the parameter savings. The fact that it handles any number of modalities without redesigning the coupling function is a legitimate practical advantage over the vision-language-only adapters.\n\nThe problem is empirical. Every number in Tables 1–4 is a single run. No seeds, no confidence intervals, no significance tests. And the decisive margins are small: 91.1 vs 90.8 on UPMC-Food 101, 83.2 vs 82.1 on CMU-MOSI, a 0.1 deficit on IEMOCAP. For datasets of that size, run-to-run noise of 0.5–1.5 points is typical, so the central claim of consistent SOTA outperformance is not established. The stress-test note is on point. There is also an internal inconsistency: Wander(d=16) shows 0.3M parameters in Table 3 but 0.1M in Table 5, with no explanation. And for a paper whose vector fusion is exactly LMF, the absence of a direct LMF baseline on the same tasks is a glaring omission. No code is released either, which makes independent verification harder.\n\nThese are all fixable. I don't see a load-bearing flaw in the method itself. The rank-8 CP approximation might be fine; the problem is we can't tell from the current experiments. So: send it to peer review, but require multiple seeds with variance, a proper LMF comparison, a reconciliation of the parameter counts, and ideally code. If the authors deliver that, this could become a useful baseline in the multimodal adapter literature.\n\nIt's worth a reading-group discussion, though I wouldn't rely on the empirical numbers yet.","headline":"A clean extension of LMF to sequence-level fusion, but the empirical evidence is too thin to support the 'consistent SOTA' claim.","tokens_in":14076,"tokens_out":3707,"would_cite":true,"duration_ms":34804,"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":"Wander, a low-rank sequence multimodal adapter, achieves token-level fusion across any number of modalities while training far fewer parameters than existing adapters, matching or beating full fine-tuning on benchmarks with two, three…","keywords":["parameter-efficient transfer learning","multimodal adapter","sequence fusion","CP decomposition","low-rank factorization","outer product fusion","multimodal models"],"falsifier":"On a three-modality benchmark, compute the exact sequence-interaction tensor of Equation 9 with reduced dimensions and measure the relative Frobenius-norm error of its best rank-8 CP approximation; if the error is large while downstream accuracy stays high, the method would be relying on something other than faithful low-rank reconstruction of the interaction tensor, and if the error is large and accuracy drops sharply at higher ranks, the low-rank assumption itself would fail.","tokens_in":13025,"feed_emoji":"⚡","tokens_out":12249,"duration_ms":94708,"temperature":0.7,"pith_summary":"The paper introduces Wander, an adapter module that lets a pre-trained multimodal backbone be fine-tuned for a new task by training only a small set of added parameters. Its central claim is that token-level interactions between sequences of different modalities can be modeled efficiently by factoring the high-dimensional interaction tensor through CP decomposition, so the full outer product never has to be computed. If the claim holds, multimodal transfer learning can scale to any number of modalities—not just vision and language—while using far fewer trainable parameters than existing adapters and matching or exceeding full fine-tuning accuracy. The paper supports this with experiments on four benchmarks spanning two, three, and seven modalities.","feed_headline":"Low-rank adapter fuses 2 to 7 modalities with 100x fewer parameters","feed_subtitle":"Token-level fusion via CP decomposition lets frozen backbones handle any number of modalities with far fewer parameters.","key_machinery":"The central identity is Equation 13, which factorizes the sequence-level interaction tensor through CP (CANDECOMP/PARAFAC) decomposition. Instead of materializing the full M-modal outer product tensor, the output is expressed as an element-wise (Hadamard) product across modalities of sums of rank-one terms, each term being a per-modality low-rank matrix multiplied with the sequence features. This turns an exponential parameter and memory cost into one that is linear in the number of modalities, and it is the mechanism by which Wander achieves token-level interactions among any number of modalities with few trainable parameters.","core_discovery":"The paper claims that Wander performs fine-grained, token-level fusion of multimodal sequences in a parameter-efficient way, and that this generalizes to more than two modalities. Starting from the outer-product fusion of unimodal representations, Wander factorizes both the feature-projection weight and the sequence-interaction weight via CP decomposition into sums of rank-one tensors. The key result is Equation 13, which rewrites the fused output as an element-wise (Hadamard) product across modalities of sums of low-rank terms, eliminating the need to materialize the exponential interaction tensor. Empirically, the paper reports that Wander outperforms state-of-the-art efficient transfer learning methods on UPMC-Food 101 (2 modalities), CMU-MOSI and IEMOCAP (3 modalities), and MSRVTT (7 modalities), with fewer trainable parameters than competing adapters and, on CMU-MOSI and MSRVTT, better accuracy than full fine-tuning.","pith_inferences":["The same CP-factorized Hadamard-product identity could be applied to other high-order interaction layers in Transformers, such as factorizing attention across multiple input sequences, not just multimodal fusion.","A natural next step is a formal bound on the approximation error of the rank-R factorization as a function of R and the modality dimensions; the paper only provides empirical rank-sensitivity results.","Because the fusion treats all modalities symmetrically, it may suit settings with missing modalities or dynamically changing input types, which the paper does not test.","The reported gains on the seven-modality MSRVTT benchmark suggest the efficiency advantage grows with modality count, so a systematic scaling study would quantify the savings beyond seven modalities."],"forward_implications":["Wander can be added to any frozen Transformer-based multimodal backbone and trained with only the task head, regardless of the number of input modalities.","On the tested benchmarks, Wander matches or beats full fine-tuning while using roughly one to two orders of magnitude fewer trainable parameters, for example 0.9 million versus 80 million on CMU-MOSI.","Because the interaction tensor is never materialized, the method avoids the exponential memory blow-up of explicit outer-product fusion, making it feasible for many modalities.","The CP rank can be kept small (the default is 8) with little performance change, so the parameter savings persist as the number of modalities grows."],"supporting_citations":[{"why":"Introduces the adapter module with down-projection, nonlinearity, and up-projection that Wander generalizes by removing the explicit up-projection.","marker":"(Houlsby et al. 2019)"},{"why":"LoRA, the standard low-rank reparameterization baseline that serves as a principal comparison point.","marker":"(Hu et al. 2021)"},{"why":"Supplies the efficient low-rank multimodal fusion technique that Wander extends to sequence-level, token-level fusion.","marker":"(Liu et al. 2018)"},{"why":"Establishes the outer-product fusion of unimodal representations that Wander inherits and makes parameter-efficient.","marker":"(Zadeh et al. 2017)"},{"why":"Provides the MMT backbone and the seven-modality feature set used for MSRVTT, the dataset demonstrating extension beyond two modalities.","marker":"(Gabeur et al. 2020)"},{"why":"UniAdapter, a recent vision-language adapter baseline that Wander claims to outperform while also supporting more than two modalities.","marker":"(Lu et al. 2024)"},{"why":"PMF, the prompt-based multimodal fusion baseline that Wander compares against on all datasets.","marker":"(Li et al. 2023b)"}],"fun_headline_variants":["Low-rank adapter fuses 2-7 modalities at token level","Wander adapter: token-level fusion for 2-7 modalities","CP-decomposed adapter cuts multimodal fine-tuning cost","Efficient adapter fuses modalities without huge tensors"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes the information needed for downstream tasks lives in a low-rank subspace of the full inter-modality token-interaction tensor, so a rank-8 CP factorization preserves it.","fun_headline_variants_meta":{"raw":{"variants":["Low-rank adapter fuses 2-7 modalities at token level","Wander adapter: token-level fusion for 2-7 modalities","CP-decomposed adapter cuts multimodal fine-tuning cost","Efficient adapter fuses modalities without huge tensors"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000148,"raw_usage":{"total_tokens":1183,"prompt_tokens":936,"completion_tokens":247,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":552,"completion_tokens_details":{"reasoning_tokens":178}},"tokens_in":552,"tokens_out":247,"duration_ms":3029,"temperature":1.0,"reasoning_tokens":178,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T17:20:36.720575+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a three-modality benchmark, compute the exact sequence-interaction tensor of Equation 9 with reduced dimensions and measure the relative Frobenius-norm error of its best rank-8 CP approximation; if the error is large while downstream accuracy stays high, the method would be relying on something other than faithful low-rank reconstruction of the interaction tensor, and if the error is large and accuracy drops sharply at higher ranks, the low-rank assumption itself would fail.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the MMT backbone and the seven-modality feature set used for MSRVTT, the dataset demonstrating extension beyond two modalities."}],"review_version":1}