{"id":"7ed24991-7e0d-4fb8-8df9-65854ffbd6e5","arxiv_id":"2411.19297","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"FreqFit is a frequency-domain filter module that, when inserted between ViT blocks, improves the accuracy of existing PEFT methods on most but not all evaluated benchmarks.","lead":"This paper adds a learnable frequency-domain filter, called FreqFit, between layers of a vision transformer to boost the accuracy of parameter-efficient fine-tuning methods. The authors report gains of 1 to 16 percent across 24 image classification benchmarks, but the strength of that claim depends on how the comparison baselines were tuned.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"FreqFit's headline +16.4% Adapter gain rests on an implausibly weak Adapter-64 baseline (56.0 vs ~72+ for other PEFTs); a plain scale-shift recovers +14.8, so the claimed benefit is a baseline artifact.","rationale":"The reader's weakest assumption identifies the same load-bearing concern: the Adapter-64 baseline appears undertuned, and the central claim collapses if it is. I agree with that assessment after independently checking the tables. The strongest quantitative support for the abstract's claim is Table 1's mean gains, and the single largest gain (+16.4%) is driven entirely by an implausible baseline: Adapter-64 at 56.0 mean accuracy, while every other PEFT method in the same table sits between 71 and 73. The paper's own supplementary material (Table 6) makes the artifact concrete: a trivial scale-shift transform on the same baseline yields +14.8 of the +16.4 gain, leaving only +1.6 for the frequency filter. That means the headline result does not demonstrate frequency-based adaptation; it demonstrates that the Adapter baseline was not configured to a competitive level. The 'consistently improves' claim is also internally contradicted by the authors' own tables (e.g., FreqFit-Bias on MAE at -0.8 in Table 2, and multiple negative per-task entries in Table 1), and most results lack error bars or seed variance. The theoretical support in Section 8 would not rescue the empirical claim even if it were fully correct, and it is itself problematic (O(1) parameter claim vs. an H*W*D filter). In summary, the central claim is not supported as stated. A revision that fixes the Adapter baseline, narrows the claim to modest gains on some methods, and reports variance could be reconsidered, but the present paper should be rejected for the claim it makes.","tokens_in":24828,"tokens_out":5783,"duration_ms":47822,"concrete_test":"Recompute the VTAB-1k results in Table 1 for both Adapter-64 and FreqFit-Adapter-64 under identical conditions, but tune the Adapter baseline properly: search reduction factors {16, 64, 256}, learning rates, and insertion positions on each task's validation split (or use the official VPT/AdaptFormer configurations), and report the mean over at least 3 seeds. If the corrected Adapter-64 mean rises from 56.0 toward the ~70 level that scale-shift already achieves, the claimed +16.4% gain is a baseline artifact and the central claim fails; if Adapter-64 remains near 56 under a fair hyperparameter search, the result instead indicates a serious implementation bug in the baseline that must be fixed before any comparison is meaningful.","verdict_should_be":"REJECT","load_bearing_attack":"The abstract's central claim ('FreqFit consistently improves performance over the original PEFT methods with gains ranging from 1% to 16%') is carried almost entirely by Table 1. The largest entry, FreqFit-Adapter-64's +16.4% mean gain, rests on an Adapter-64 baseline of 56.0 mean accuracy on VTAB-1k, about 16 points below every other PEFT row in the same table (VPT 72.0, LoRA 73.1, BOFT 71.2, VeRA 72.8, FourierFT 72.8). A baseline that far below its peers is not representative of a properly configured Adapter, and the paper provides no evidence that this Adapter-64 configuration was tuned to a comparable level; Section 4 explicitly says a single default configuration was used across all experiments, with no reduction-factor search for Adapter. The suspicion is confirmed within the paper's own supplementary Table 6: adding a plain scale-shift transform to the same Adapter-64 baseline raises it from 56.0 to 70.8, a +14.8 gain that leaves FreqFit's advantage over scale-shift at only +1.6. Thus the headline +16.4% is mostly an artifact of an undertuned baseline, not evidence that frequency-based adaptation is the cause. Moreover, the 'consistently improves' wording is contradicted by the paper's own tables: FreqFit-Bias on MAE degrades by -0.8 in Table 2, and several per-task entries in Table 1 are negative. Once the Adapter baseline is corrected, the advertised improvement range collapses to roughly 0-2%, and the frequency-specific benefit over scale-shift is not established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FreqFit, a plug-in module inserted between ViT blocks that transforms features with an FFT, a learnable frequency-domain filter, an inverse FFT, and a per-channel scale-and-shift, followed by a residual connection. The module is designed to be combined with existing PEFT methods such as Linear probing, Bias tuning, Adapter, VPT, LoRA, BOFT, VeRA, and FourierFT. The authors report extensive experiments on VTAB-1k and FGVC tasks with ImageNet-21K, MAE, MoCo, and CLIP pretrained ViT-B/16 models, claiming consistent improvements of 1% to 16% over the base PEFT methods, with the largest gain being +16.4% for FreqFit-Adapter-64 on VTAB-1k. The paper also provides two theorems intended to show that FreqFit performs transformations that spatial-domain PEFT methods cannot replicate.","tokens_in":25229,"tokens_out":8107,"duration_ms":63260,"significance":"If the central claims held, FreqFit would be a simple, architecture-agnostic module with broad applicability, and the paper would be a useful contribution to the PEFT literature. The paper does make a genuine effort to cover many datasets, multiple pretraining strategies, and several PEFT families, and it provides source code. However, the empirical evidence as presented does not support the headline gain: the largest reported improvement is an artifact of an implausibly weak Adapter baseline, the abstract's 'consistently improves' claim is contradicted by the paper's own tables, and the theoretical argument contains a clear error about parameter complexity. The comparison against a simple scale-shift control, which the paper itself includes, reduces the claimed benefit of the frequency-filtering mechanism to roughly 1-2% on average, with no error bars on the main tables.","major_comments":[{"comment":"Table 1 reports an Adapter-64 baseline of 56.0 mean accuracy on VTAB-1k with ImageNet-21K weights, which is about 16 points below every other PEFT row in the same table (VPT 72.0, LoRA 73.1, BOFT 71.2, VeRA 72.8, FourierFT 72.8) and well below published Adapter results on this benchmark. Section 4 states that a single default configuration was used with no search over the Adapter reduction factor. The headline gain of +16.4% for FreqFit-Adapter is therefore likely an artifact of an undertuned baseline rather than evidence for the method. The supplementary Table 6 confirms this concern: adding a plain scale-shift to the same Adapter-64 baseline yields +14.8, leaving FreqFit's edge over scale-shift at only +1.6.","section":"Table 1 and Section 4"},{"comment":"The abstract claims that FreqFit 'consistently improves performance' with gains ranging from 1% to 16%. This is contradicted by the paper's own results: Table 2 shows FreqFit-Bias on MAE degrades to 53.8 from 54.6 (-0.8), and Table 1 contains multiple negative per-task differences, such as FreqFit-Linear on Retinopathy at 72.9 versus 74.0. The word 'consistently' is not supported by the data, and the lower bound of the claimed gain range is not respected by these entries.","section":"Abstract vs. Tables 1 and 2"},{"comment":"Theorem 1 and the surrounding text state that FreqFit uses O(1) parameters. This is incorrect: the learnable filter K in Eq. (5) has H x W x D trainable entries, and Algorithm 1 in the appendix defines a parameter tensor of shape h x w x dim x 2, which scales with the feature map and embedding dimensions. The claim that the filter can be 'parameterized efficiently ... regardless the input dimensions' is not reflected in the implementation. This error undermines the theoretical contribution as stated.","section":"Section 3 and Appendix, Theorem 1"},{"comment":"The paper's own comparison against scale-shift shows that most of the gain over the raw baselines is captured by the scale-shift components alone. In supplementary Table 6, scale-shift raises Adapter-64 from 56.0 to 70.8, while FreqFit-Adapter reaches 72.4; the incremental gains of FreqFit over scale-shift are 1.6 for Adapter, 2.6 for LoRA, 1.4 for VPT, and 0.1 for Bias. Since scale-shift (alpha, beta) is the final stage of FreqFit in Eq. (7), the experiments do not isolate the effect of frequency-domain modulation, and the central mechanism of the paper is not established as the cause of the reported improvements.","section":"Fig. 4 and supplementary Table 6"},{"comment":"Tables 1 and 2 report no error bars or number of seeds for the VTAB-1k results. Several claimed gains are small (FreqFit-LoRA +1.6, FreqFit-VeRA +0.8, FreqFit-FourierFT +1.4), and without variance estimates these differences cannot be distinguished from training noise. Only Table 3 reports standard deviations, for a subset of FGVC experiments, and the main benchmark tables do not allow the reader to assess the stability of the improvements.","section":"Tables 1 and 2"}],"minor_comments":[{"comment":"There is a typo: 'consituents' should be 'constituents' in the paragraph on the Fourier transform.","section":"Section 3"},{"comment":"Algorithm 1 initializes a parameter named 'self.filter_weight' but then uses 'self.complex_weight' in the forward pass; the code as written would not run.","section":"Appendix, Algorithm 1"},{"comment":"The column header 'Sun395' should be 'Sun397' to match the VTAB-1k dataset naming used elsewhere in the paper.","section":"Table 2"},{"comment":"The paper alternates between the spellings 'FreqFit' and 'FreqFiT' (e.g., the title uses 'FreqFit' while Sections 5 and 6 use 'FreqFiT'); please use one consistent spelling.","section":"Throughout"},{"comment":"In the paragraph on Adapter, the text says the improvement is with 'Imagenet-1K pre-trained weights,' but Table 1 is based on ImageNet-21K; this inconsistency should be corrected.","section":"Section 5"},{"comment":"The proof notation in Eqs. (15)-(17) is confusing: the inverse transform uses U_H^H and U_W^H but the text calls them conjugate transposes, and the claim that F is a 'full-rank diagonal matrix' is inconsistent with K being a general 3D filter; the proof would benefit from clearer definitions.","section":"Appendix, Theorem 1 proof"}],"recommendation":"reject","confidential_remarks":"The manuscript's main quantitative claim rests on an Adapter baseline that appears to be non-representative, and the paper's own scale-shift control reduces the effective gain to about 1-2%. The theoretical 'O(1) parameters' claim is false as stated. These are load-bearing issues that would require reworking the central claims and rerunning experiments with properly tuned baselines and error bars; even then, the contribution would be much weaker than advertised. I do not see a path within the current manuscript's scope to support the abstract's claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the FreqFit module—an FFT-based filter plus scale-shift inserted between ViT blocks—is a clean plug-in and the paper runs a lot of experiments, but the headline '+16.4% on VTAB-1k' is mostly an artifact of an undertuned Adapter baseline. What survives is a small but fairly consistent gain over a plain scale-shift, which might be real but isn't proven at the scale claimed.\n\nThe genuinely useful part is the comparison with scale-shift (Supplementary Table 6, Figure 4). Averaged over methods, FreqFit beats SSF by about 1.2 points on the same backbone and eval setup. That is a modest, credible effect and the right control experiment. They also report seed variance for the FGVC runs, re-run Gated-VPT with matched seeds, and ship code. Those are good practices.\n\nThe problem is the central claim 'consistently improves 1% to 16%.' Table 1's Adapter-64 baseline sits at 56.0 mean, roughly 16 points below the other PEFT rows in the same table. With a default configuration and no reduction-factor search, that baseline is simply weak, and the +16.4 is not a FreqFit effect. The paper's own Table 6 confirms this: adding plain scale-shift to that same Adapter gets +14.8, leaving FreqFit's advantage over scale-shift at +1.6. Meanwhile FreqFit-Bias on MAE is -0.8, which contradicts the 'consistently improves' wording. Most tables have no error bars, so the 1-2% gains are within noise.\n\nThe theory section has a load-bearing internal inconsistency: Theorem 1 claims O(1) parameters while defining K in C^{H x W x D} and implementing it as an H x W x D tensor in the released code. That is not a proof-sketch detail; it is the basis for the 'missing piece' argument. The theorems as stated also don't establish the empirical claims—'cannot replicate' is true for any two different linear operations and doesn't translate to a performance guarantee.\n\nWho is this for? Someone working on PEFT for ViTs might get a useful data point: frequency-domain feature modulation adds a little over scale-shift, and the negative results (Bias, self-supervised MAE) are worth knowing. A reader should not take the 1-16% headline at face value.\n\nMy recommendation: send it to review with a request for major revision. The authors need to fix the Adapter baseline, report variance across seeds, and narrow the claim to 'a small consistent gain over scale-shift.' The empirical idea is coherent enough to deserve refereeing, but the current version overstates what is known.","headline":"The headline 16% Adapter gain is a baseline artifact; the paper's honest contribution is a ~1-2% gain over scale-shift, undermined by an internally inconsistent theory section.","tokens_in":25750,"tokens_out":3723,"would_cite":false,"duration_ms":32061,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Frequency filter lifts fine-tuning accuracy by 1 to 16 points: the paper claims inserting a learnable FFT-based module between frozen ViT blocks lets parameter-efficient methods capture high-frequency image details, with gains from 1% to…","keywords":["frequency-domain adaptation","parameter-efficient fine-tuning","vision transformers","fast Fourier transform","learnable spectral filter","visual prompt tuning","low-rank adaptation","VTAB-1k"],"falsifier":"Re-run Table 1 on VTAB-1k with ImageNet-21K weights while tuning the Adapter baseline's reduction factor per published practice, then add FreqFit and measure the gain; if the re-tuned Adapter baseline rises to around 70 mean accuracy and the FreqFit-Adapter gain shrinks to the roughly 1-2% seen with LoRA, the strongest reported result fails.","tokens_in":24608,"feed_emoji":"📈","tokens_out":8843,"duration_ms":70648,"temperature":0.7,"pith_summary":"This paper claims that parameter-efficient fine-tuning methods act in the spatial domain and therefore miss the high-frequency image structure that many downstream tasks need. To fix it, the authors introduce FreqFit, a plug-in between ViT blocks that takes a feature map, applies a 2D fast Fourier transform, multiplies the spectrum by a learnable complex filter, transforms it back, applies a learned scale-and-shift, and adds the result residually. Across 24 datasets, four pre-training regimes, and seven existing PEFT methods, the paper reports consistent accuracy gains of 1% to 16% over each method alone. A method-agnostic frequency filter would be a practically attractive addition to any existing fine-tuning recipe.","feed_headline":"Frequency filter boosts fine-tuning accuracy by up to 16 points","feed_subtitle":"A learnable FFT filter between frozen ViT blocks helps LoRA, Adapter, and VPT see subtle image detail.","key_machinery":"The central object is the FreqFit layer: a learnable complex filter $K \\in \\mathbb{C}^{H \\times W \\times D}$ that is multiplied elementwise with the 2D fast Fourier transform of the feature map $X \\in \\mathbb{R}^{H \\times W \\times D}$ along the spatial dimensions, after which an inverse FFT, a per-channel scale $\\alpha$ and shift $\\beta$, and a residual connection return the result to the spatial domain. The layer does spectral modulation: because the FFT mixes information across all spatial positions, a single filter can amplify or suppress whole frequency bands in a way that per-token spatial updates cannot. The paper's Theorem 1 claims that FreqFit with $O(1)$ parameters per channel produces transformations that spatial-domain PEFT methods cannot replicate, and Theorem 2 claims FreqFit and spatial PEFT are complementary; both proofs rely on the contrast between FreqFit's within-channel, across-token mixing and LoRA-style within-token, across-channel updates.","core_discovery":"FreqFit is a frequency fine-tuning module inserted between ViT blocks, computing $\\hat{X} = X + \\alpha \\odot F^{-1}(F(X) \\odot K) + \\beta$, where $F$ is the 2D fast Fourier transform along the spatial dimensions, $K \\in \\mathbb{C}^{H \\times W \\times D}$ is a learnable filter with one complex weight per frequency per channel, and $\\alpha, \\beta$ are learned per-channel scale and shift. The paper's central claim is that this module, with $O(1)$ parameters per channel, changes feature maps in ways spatial-domain PEFT methods cannot, because the FFT couples all spatial positions while a low-rank or per-token update does not. Combining FreqFit with methods such as LoRA, Adapter, VPT, BOFT, VeRA, and FourierFT is claimed to give consistent accuracy gains on 24 datasets, with mean gains of 0.8% to 16.4% on VTAB-1k with ImageNet-21K weights and a gain above 10% for FreqFit-LoRA on CIFAR-100. The largest reported gains occur with the weakest spatial baselines and the smallest with already-strong methods, consistent with the paper's complementarity argument.","pith_inferences":["Because the proofs hinge on the FFT's global across-token mixing, the same complementarity should hold for any global linear token-mixing operator, such as a learned global convolution, so the frequency framing is plausibly one instance of a broader global-versus-token-local decomposition of PEFT design space.","The near-parameter-free cost of the filter suggests a testable extension where FreqFit is added to fully fine-tuned or heavily tuned models; the paper only tests frozen-backbone PEFT settings.","The above-10% CIFAR-100 gain for FreqFit-LoRA without Mixup or strong augmentation hints that spectral modulation may substitute for part of what augmentation provides, predicting that FreqFit's gain shrinks when strong augmentation is used.","Since the paper reports FreqFit combined with CLIP, a natural unstated extension is applying the module to dense prediction and multi-modal heads, where high-frequency detail matters for segmentation and detection."],"forward_implications":["Any spatial-domain PEFT method can be upgraded by inserting FreqFit before each ViT block, with the largest relative gains where the base method is weakest (Adapter and Linear) and smaller gains where it is already strong (LoRA, VeRA).","FreqFit used alone with a linear head rivals or beats full fine-tuning on some VTAB-1k tasks, meaning frequency modulation can substitute for updating the backbone.","FreqFit can be placed inside self-attention after the QKV projection, giving gains up to 6% on KITTI-Dist with no attention-weight updates.","The learned filters specialize per task and per layer into high-pass, low-pass, and band-pass patterns, supporting the claim that the model adapts its frequency response to the data.","FreqFit outperforms scale-and-shift feature transformation by about 1.2% mean accuracy across the same methods, so the frequency modulation, not merely scaling and shifting, carries the benefit."],"supporting_citations":[{"why":"LoRA is the running example in the proofs of Theorems 1 and 2 and a primary baseline across all tables.","marker":"[22]"},{"why":"The Adapter baseline that yields the largest reported FreqFit gain (+16.4% mean on VTAB-1k).","marker":"[21]"},{"why":"VPT is a baseline method and the source of the prompt-length configuration reused for FreqFit-VPT.","marker":"[23]"},{"why":"FourierFT is the closest prior frequency-domain PEFT method; FreqFit is positioned against it and combined with it as FreqFit-FourierFT.","marker":"[17]"},{"why":"SSF is the scale-and-shift baseline that FreqFit is compared against in Sec. 6, attesting that the gain comes from frequency modulation rather than scaling alone.","marker":"[27]"},{"why":"Defines the VTAB-1k benchmark and its 19 tasks on which the reported mean gains are calculated.","marker":"[54]"},{"why":"Gated-VPT supplies the prompt-length settings and the reproducibility comparison for FreqFit-VPT on MAE and MoCo in Table 3.","marker":"[52]"},{"why":"Cited for the claim that high-frequency features are vital for performance, the premise motivating the frequency-based design.","marker":"[19]"}],"fun_headline_variants":["FreqFit FFT module boosts PEFT accuracy up to 16%","Frequency tuning adds up to 16% to ViT fine-tuning gains","Learnable FFT bridge lifts LoRA, Adapter, VPT performance","Fast Fourier module gives PEFT up to 16% boost"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported gains assume the baseline PEFT methods, especially the Adapter-64 configuration that scores 56.0 mean accuracy on VTAB-1k, are implemented and tuned at a representative level; if that baseline is undertuned, the +16.4% headline gain comes from the comparison rather than from FreqFit.","fun_headline_variants_meta":{"raw":{"variants":["FreqFit FFT module boosts PEFT accuracy up to 16%","Frequency tuning adds up to 16% to ViT fine-tuning gains","Learnable FFT bridge lifts LoRA, Adapter, VPT performance","Fast Fourier module gives PEFT up to 16% boost"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001294,"raw_usage":{"total_tokens":5342,"prompt_tokens":1068,"completion_tokens":4274,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":684,"completion_tokens_details":{"reasoning_tokens":4195}},"tokens_in":684,"tokens_out":4274,"duration_ms":26136,"temperature":1.0,"reasoning_tokens":4195,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T10:19:18.342055+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run Table 1 on VTAB-1k with ImageNet-21K weights while tuning the Adapter baseline's reduction factor per published practice, then add FreqFit and measure the gain; if the re-tuned Adapter baseline rises to around 70 mean accuracy and the FreqFit-Adapter gain shrinks to the roughly 1-2% seen with LoRA, the strongest reported result fails.","supporting_citations":[{"cited_title":"Parameter-efficient transfe r learning for nlp","cited_arxiv_id":null,"evidence_quote":"The Adapter baseline that yields the largest reported FreqFit gain (+16.4% mean on VTAB-1k)."},{"cited_title":"Vi- sual prompt tuning","cited_arxiv_id":null,"evidence_quote":"VPT is a baseline method and the source of the prompt-length configuration reused for FreqFit-VPT."}],"review_version":1}