{"id":"2d4f1db6-4eb1-4cb1-8cc2-3978e586d2b3","arxiv_id":"2507.17347","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Swin-TUNA inserts layer-dependent depthwise-convolution adapters into a frozen Swin-L backbone and reports 50.56 mIoU on FoodSeg103 and 74.94 mIoU on UECFoodPix Complete with 8.13M trainable parameters.","lead":"This paper introduces Swin-TUNA, a parameter-efficient adapter that adds trainable multi-scale modules to a frozen Swin Transformer for food image segmentation. It reports state-of-the-art accuracy on two food segmentation benchmarks while training about 8 million parameters instead of hundreds of millions.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline claim against FoodSAM is not controlled: full-fine-tuning baselines are quoted at different crop sizes and protocols, and no same-protocol full fine-tuning of the Swin-L pipeline is reported; until that baseline is run, 'PEFT surpasses full fine-tuning' is unsupported.","rationale":"The reader's weakest assumption is the right one, and I sharpen it: the decisive comparison is not just to FoodSAM but to full fine-tuning of the same backbone under the same protocol. The parameter-efficiency part of the claim is internally consistent for training cost, since only adapters and the head are trained and 8.13M trainable parameters is small. But the accuracy claim is only as strong as the full-fine-tuning baseline it beats. Quoting FoodSAM's published 46.42 mIoU at 768×768 while training Swin-TUNA at 640×640 introduces a protocol confound. The fact that linear probing, with no backbone training, already exceeds FoodSAM strongly suggests the baseline is not a like-for-like measurement. A same-protocol full fine-tuning run of Swin-L would directly resolve whether the adapter truly beats full fine-tuning; in strong segmentation backbones this is a high bar. I also note the paper selects the final architecture using FoodSeg103 test-set results in the ablation (Table 3) and reports no repeated-run variance, which are secondary but reinforce the need for an independent control. The code link and matched PEFT comparisons are genuine strengths, and there is no sign of fabrication; the missing control is checkable, so the appropriate response is to keep the CONDITIONAL verdict rather than reject or accept.","tokens_in":12355,"tokens_out":10617,"duration_ms":110323,"concrete_test":"Using the released code or MMSegmentation, train the same Swin-Large backbone and segmentation head from ImageNet-22K with all parameters unfrozen under the paper's exact recipe on FoodSeg103: AdamW lr=1e-4, weight decay=0.01, 100k iterations, batch size 4, random crop 640×640, cosine schedule, and the paper's mIoU evaluation. If this full fine-tuning baseline reaches or exceeds 50.56 mIoU, the claim that Swin-TUNA outperforms full fine-tuning is falsified. Repeat the analogous run at 512×512 with 90k iterations on UECFoodPix Complete and compare with 74.94; optionally, rerun FoodSAM at the same crop sizes to confirm the headline comparison.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is tested in Section 4.2, Tables 1 and 2, where the only full-fine-tuning comparators, FoodSAM and FDSNet, are quoted from papers using 768×768 crops, whereas Swin-TUNA is trained at 640×640 on FoodSeg103 and 512×512 on UECFoodPix Complete. Those baselines also use different backbones and heads; FoodSAM is a SAM-based model, not a fine-tuned Swin-L. The PEFT baselines appear to be matched, but no full fine-tuning run of the paper's own Swin-L/segmentation-head pipeline under the same recipe is reported. This matters because the quoted FoodSAM number is already below the paper's own linear probing result (46.42 vs 47.85 on FoodSeg103), which is only plausible if the baselines are not protocol-matched. Unless a same-protocol full fine-tuning baseline is run and shown to stay below 50.56/74.94, the headline 'surpassing full fine-tuning' is not supported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Swin-TUNA, a parameter-efficient fine-tuning (PEFT) method for food image segmentation. The method inserts tunable adapter modules into a frozen Swin-L backbone, with stage-dependent depthwise convolution kernel sizes and bottleneck dimensions, plus two trainable scaling factors s1 and s2 to balance frozen and adapted features. Experiments on FoodSeg103 and UECFoodPix Complete report mIoU values of 50.56% and 74.94%, respectively, with 8.13M trainable parameters, and the paper claims that this surpasses the fully parameterized FoodSAM model while using only about 4% of the parameters. The paper also reports faster convergence and better low-data generalization than other PEFT baselines, and the code is made publicly available.","tokens_in":12590,"tokens_out":4464,"duration_ms":46482,"significance":"If the central comparison were properly controlled, the result would be practically significant: achieving state-of-the-art food segmentation accuracy with a small fraction of trainable parameters would be a useful contribution, and the hierarchical adapter design is a plausible, easily transferable idea. The paper has real strengths: it provides public code, evaluates on two established benchmarks, and the PEFT baselines in Tables 1 and 2 appear to be matched in crop size and batch size. However, the main advertised claim, that PEFT surpasses full fine-tuning, currently rests on comparisons to published FoodSAM and FDSNet numbers that were obtained under different crop sizes and with different architectures. The significance is therefore conditional on the authors either providing a same-protocol full fine-tuning baseline or reframing the claim.","major_comments":[{"comment":"The headline claim that Swin-TUNA surpasses full fine-tuning is not supported by a controlled comparison. FoodSAM and FDSNet are quoted from earlier papers trained at 768×768 crops, whereas Swin-TUNA is trained at 640×640 on FoodSeg103 and 512×512 on UECFoodPix Complete, and those baselines use different backbones and segmentation heads (FoodSAM is SAM-based, not a fine-tuned Swin-L). No full fine-tuning of the paper's own Swin-L/segmentation-head pipeline under the same recipe is reported. This matters because the quoted FoodSAM number on FoodSeg103 (46.42) is already below the paper's own linear probing result (47.85), which strongly suggests protocol mismatch rather than a true advantage of the adapter. I request a same-protocol full fine-tuning baseline, or a clear restatement of the claim as 'outperforms published FoodSAM/FDSNet numbers' rather than 'surpassing full fine-tuning.'","section":"§4.2, Tables 1 and 2"},{"comment":"The final configuration of Swin-TUNA—kernel sizes [7,5,5,3] and bottleneck widths [64,64,96,192]—is selected by comparing FoodSeg103 test-set mIoU values, and no validation split or model-selection protocol is described. Using the test set for hyperparameter selection can inflate the reported test mIoU relative to an independent evaluation, and it makes it difficult to judge whether the advantage over Mona, AdaptFormer, and BitFit is robust. Please report a validation-based selection procedure, or explicitly evaluate the selected configuration on a held-out test split that was not used for any model choice.","section":"§4.4, Table 3"},{"comment":"No repeated runs, seeds, or confidence intervals are reported. The decisive differences are small—for example, 50.56 vs. 49.39 for BitFit and 49.31 for Mona on FoodSeg103, and 74.94 vs. 75.89 for FDSNet on UECFoodPix—so single-run results cannot establish that the ranking is stable. The convergence claims in Figure 4 are also based on single curves. I request at least multiple seeds (or a statement that these are single runs) and variance estimates for the main comparisons, so that the reader can assess whether the observed gaps are within run-to-run noise.","section":"Tables 1, 2, and Figure 4"}],"minor_comments":[{"comment":"The sentence 'Swin-TUNA outperforms the other PEFT methods on FoodSeg103 and eventually outperforms all other methods' is contradicted by Table 2, where FDSNet reports 75.89 mIoU versus Swin-TUNA's 74.94; please correct this sentence.","section":"§4.2"},{"comment":"The abstract and conclusion say that only 4% of parameters are updated, while the text in §4.2 says 'with a mere 8% of the parameters trained'; please check which number is intended and make it consistent.","section":"Abstract and §4.2"},{"comment":"Equation (5) contains 'oplus' in the running text and the residual term appears to add z^{l-1} twice; the notation should be cleaned up so the reader can follow the exact computation.","section":"Eq. (5)"},{"comment":"The caption for Figure 5(a) describes both structures as 'parallel'; the right side is described in the text as 'sequential', so the caption should be fixed.","section":"Figure 5(a)"},{"comment":"The table headers '%' are unclear; use labels such as 'fixed' vs. 'adaptive' for the two factors instead.","section":"Table 3"},{"comment":"The final sentence, 'providing an efficient solution for assembling lightweight food image,' is incomplete and should be rewritten.","section":"Section 5"}],"recommendation":"major_revision","confidential_remarks":"The central issue is the uncontrolled comparison with full fine-tuning baselines. If the authors add a same-protocol full fine-tuning baseline and address the test-set hyperparameter selection, the paper's main claim could become defensible. The small gaps between the best PEFT methods also warrant repeated runs. I do not see a circularity problem; the result is empirical and the reported mIoU is not an identity or restatement of an input."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the paper has one sound contribution—a concrete adapter variant that modestly beats matched PEFT baselines on two food segmentation datasets—and one inflated claim: that it surpasses full fine-tuning (FoodSAM). I agree with the reader's conditional verdict; the full fine-tuning comparison is not controlled, and the numbers as presented do not support the headline.\n\nWhat is new: the adapter uses layer-dependent depthwise convolution kernel sizes and bottleneck widths. That is a sensible, incremental design variation over AdaptFormer and Mona, and the paper tests it systematically in the ablation. The PEFT baselines (Mona, BitFit, AdaptFormer, linear probing) are run under the same crop size and batch size, which is the right way to compare. Code is linked, and the method is described concretely enough to reproduce.\n\nSoft spots, in proportion. The stress-test note is on target: the headline \"surpasses full fine-tuning\" rests on FoodSAM and FDSNet numbers quoted from different papers, trained at 768x768 crops, while the PEFT methods run at 640x640 (FoodSeg103) and 512x512 (UECFoodPix). The paper's own linear probing result (47.85) already beats the quoted FoodSAM (46.42) on FoodSeg103. That tells you the baselines are not protocol-matched. A same-recipe full fine-tuning of the Swin-L pipeline is not reported, so the claim is unsupported. The paper would be nearly as useful without that claim.\n\nThe margins over the closest PEFT baselines are small: 0.24% over Mona on FoodSeg103 and 0.24% on UECFoodPix. There are no error bars or repeated runs, so these gains could be noise. The ablation on FoodSeg103 reports numbers on the test set and the best configuration is selected on that basis. That is a leak, even if the differences are tiny.\n\nTwo small internal inconsistencies: the text says kernels are \"smaller in shallow layers and larger in deeper layers,\" but the listed schedule [7,5,5,3] does the opposite (larger in shallow, smaller in deep)—the underlying design rationale is correct, the sentence is backwards. Also the conclusion says \"8% of the parameters trained\" while the abstract says \"4%\"; 8.13M is about 4% of a Swin-L pipeline.\n\nThere is also heavy self-citation (roughly a third of the references), but the relevant citations (FDSNet, FoodCSWin) are real prior work from the same group, so I don't see that as a manipulation.\n\nBottom line: this is a solid incremental PEFT paper with a poorly supported headline. It deserves a serious referee, who should ask for matched full fine-tuning baselines, error bars, and test-set-based selection to be fixed. I would not cite the FoodSAM comparison, but the adapter design is worth knowing for anyone working on PEFT for segmentation.","headline":"A sound incremental adapter design with matched PEFT baselines, but the headline claim of beating full fine-tuning is unsupported because the full fine-tuning baselines are protocol-mismatched and even linear probing beats them.","tokens_in":13150,"tokens_out":9731,"would_cite":false,"duration_ms":78571,"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":"A frozen Swin Transformer with 8.13M trainable adapter parameters beats the fully fine-tuned FoodSAM on two food segmentation benchmarks.","keywords":["food image segmentation","parameter-efficient fine-tuning","Swin Transformer","adapter module","semantic segmentation","hierarchical feature adaptation","FoodSeg103","UECFoodPix Complete"],"falsifier":"Retrain FoodSAM under Swin-TUNA's exact 640x640 crops on FoodSeg103 and 512x512 crops on UECFoodPix, matching batch size and iteration count, and compare mIoU; if FoodSAM then reaches 50.56% or higher on FoodSeg103, or 74.94% or higher on UECFoodPix Complete, the claim that an 8.13M-parameter adapter surpasses the 632.75M-parameter FoodSAM would fail.","tokens_in":12124,"feed_emoji":"🍽️","tokens_out":12763,"duration_ms":116115,"temperature":0.7,"pith_summary":"This paper claims that a frozen Swin Transformer equipped with small per-layer adapter modules—a parameter-efficient fine-tuning (PEFT) approach—can segment food images more accurately than FoodSAM, a fully fine-tuned model with 632.75 million parameters. On the FoodSeg103 benchmark the method reaches 50.56 mIoU, and on UECFoodPix Complete it reaches 74.94 mIoU, both above FoodSAM's published scores, while training only about 4% of the network's parameters (8.13 million). The practical stake is that precise food segmentation for dietary monitoring or industrial inspection currently depends on heavy Transformer models that are costly to train and slow to run. If the claim holds, lightweight deployment and quick adaptation to new food domains become realistic with existing frozen backbones.","feed_headline":"8.13M trainable params beat FoodSAM's 632M on food segmentation","feed_subtitle":"Swin-TUNA updates only 4% of weights yet beats FoodSAM's full-size model on two food benchmarks.","key_machinery":"The load-bearing mechanism is the TUNA module, a parallel branch inside each Swin Transformer block consisting of a down-projection, a depthwise convolution, a 1x1 convolution, an up-projection, GeLU activation, and dropout, with a residual connection feeding back into the block. Its named innovation is Hierarchical Feature Adaptation: the depthwise kernel size and the mapping dimension are chosen per stage, with kernel sizes [7,5,5,3] and bottleneck dimensions [64,64,96,192], so shallow and deep layers receive different adapter structures. Two trainable scale parameters, s1 and s2, multiply the frozen task-agnostic features and the adapter's task-specific features, letting the network learn how much of each contributes to the output.","core_discovery":"On its own terms, the discovery is that a hierarchy-aware PEFT adapter can exceed full fine-tuning on a fine-grained dense prediction task. Swin-TUNA freezes a Swin-L backbone and inserts a TUNA module in parallel with each Swin block, training only the adapters and the segmentation head. With kernel sizes and bottleneck widths that vary across the four stages (kernel sizes [7,5,5,3] and dimensions [64,64,96,192]), the module treats shallow texture features and deep semantic features differently, and two trainable scalars balance the frozen task-agnostic branch against the adapter's task-specific branch. The measured outcome is 50.56 mIoU on FoodSeg103 and 74.94 mIoU on UECFoodPix Complete, ahead of every PEFT baseline tested and, on both datasets, ahead of the fully fine-tuned FoodSAM.","pith_inferences":["A fair head-to-head test would rerun FoodSAM and FDSNet under Swin-TUNA's exact crop sizes, batch sizes, and iteration budgets; the published baselines used 768x768 crops while Swin-TUNA used 640x640 on FoodSeg103 and 512x512 on UECFoodPix, so the reported margins might shrink or invert under matched protocols.","The 98.7% parameter reduction compares Swin-TUNA's trainable count with FoodSAM's total count, so a more direct efficiency claim would report total inference-time parameters and FLOPs, where the frozen Swin-L backbone still contributes substantially.","The layer-wise recipe of small kernels and narrow bottlenecks in early stages and larger kernels and wider bottlenecks deeper could be tested on other transformer backbones or small-sample segmentation tasks such as medical imaging, since the paper's stated mechanism is not food-specific."],"forward_implications":["A frozen large backbone can be reused for a new dense prediction task with a small trainable branch, so food segmentation models can run on cheaper hardware with less memory than full fine-tuning.","Adapter tuning preserves pre-training knowledge, which the paper argues explains why Swin-TUNA converges faster and degrades less in low-data settings than full fine-tuning.","Because Swin-TUNA outperforms AdaptFormer, Mona, BitFit, and linear probing on both benchmarks, the result implies that layer-dependent adapter design matters more than raw trainable parameter count.","On FoodSeg103 the method also surpasses FDSNet, a full fine-tuning hybrid baseline, supporting the paper's positioning of PEFT as a path to competitive accuracy rather than a compromise."],"supporting_citations":[{"why":"FoodSAM is the fully fine-tuned baseline whose published mIoU scores Swin-TUNA claims to surpass on both datasets.","marker":"[8]"},{"why":"FDSNet is the full fine-tuning hybrid baseline that Swin-TUNA beats on FoodSeg103 and trails by 0.95 mIoU on UECFoodPix Complete.","marker":"[34]"},{"why":"FoodSeg103 supplies the dataset and benchmark that is used for the main comparison, ablations, and convergence curves.","marker":"[17]"},{"why":"UECFoodPix Complete supplies the second benchmark dataset for the transfer experiment.","marker":"[18]"},{"why":"Swin Transformer is the frozen backbone into which the TUNA adapters are injected.","marker":"[16]"},{"why":"The original Adapter formulation defines the down-projection, nonlinearity, and up-projection paradigm that TUNA follows.","marker":"[39]"},{"why":"AdaptFormer is the adapter-based PEFT baseline that Swin-TUNA outperforms on both datasets.","marker":"[45]"},{"why":"Mona is the multi-cognitive-filter PEFT baseline that Swin-TUNA outperforms on both datasets.","marker":"[46]"},{"why":"BitFit is the bias-only PEFT baseline that Swin-TUNA outperforms on both datasets.","marker":"[42]"},{"why":"This cited existing PEFT attempt for SAM supports the paper's claim that prior PEFT methods had not surpassed full fine-tuning in segmentation accuracy.","marker":"[15]"}],"fun_headline_variants":["4% of weights beat FoodSAM on food segmentation","Swin-TUNA: 8.13M params outscore FoodSAM's 632M","Cut 98.7% params, still beat FoodSAM","PEFT with 4% trainable beats FoodSAM fully tuned","Hierarchy-aware adapter beats full fine-tuning"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that the published FoodSAM and FDSNet results are comparable baselines, even though they were produced under different crop sizes and training protocols than Swin-TUNA.","fun_headline_variants_meta":{"raw":{"variants":["4% of weights beat FoodSAM on food segmentation","Swin-TUNA: 8.13M params outscore FoodSAM's 632M","Cut 98.7% params, still beat FoodSAM","PEFT with 4% trainable beats FoodSAM fully tuned","Hierarchy-aware adapter beats full fine-tuning"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000196,"raw_usage":{"total_tokens":1375,"prompt_tokens":971,"completion_tokens":404,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":587,"completion_tokens_details":{"reasoning_tokens":315}},"tokens_in":587,"tokens_out":404,"duration_ms":4545,"temperature":1.0,"reasoning_tokens":315,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T14:49:48.090329+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain FoodSAM under Swin-TUNA's exact 640x640 crops on FoodSeg103 and 512x512 crops on UECFoodPix, matching batch size and iteration count, and compare mIoU; if FoodSAM then reaches 50.56% or higher on FoodSeg103, or 74.94% or higher on UECFoodPix Complete, the claim that an 8.13M-parameter adapter surpasses the 632.75M-parameter FoodSAM would fail.","supporting_citations":[{"cited_title":"FoodSAM: Any Food Segmentation","cited_arxiv_id":"2308.05938","evidence_quote":"FoodSAM is the fully fine-tuned baseline whose published mIoU scores Swin-TUNA claims to surpass on both datasets."},{"cited_title":"Food image segmentation based on deep and shallow dual-branch network","cited_arxiv_id":null,"evidence_quote":"FDSNet is the full fine-tuning hybrid baseline that Swin-TUNA beats on FoodSeg103 and trails by 0.95 mIoU on UECFoodPix Complete."},{"cited_title":"A large-scale benchmark for food image segmentation","cited_arxiv_id":null,"evidence_quote":"FoodSeg103 supplies the dataset and benchmark that is used for the main comparison, ablations, and convergence curves."},{"cited_title":"UEC-FoodPIX Complete: A large-scale food image segmentation dataset","cited_arxiv_id":null,"evidence_quote":"UECFoodPix Complete supplies the second benchmark dataset for the transfer experiment."},{"cited_title":"Parameter-efficient transfer learning for nlp, 2019","cited_arxiv_id":null,"evidence_quote":"The original Adapter formulation defines the down-projection, nonlinearity, and up-projection paradigm that TUNA follows."},{"cited_title":"5%>100%: Breaking performance shackles of full fine-tuning on visual recognition tasks, 2024","cited_arxiv_id":null,"evidence_quote":"Mona is the multi-cognitive-filter PEFT baseline that Swin-TUNA outperforms on both datasets."},{"cited_title":"Bitfit: Simple parameter-efficient fine-tuning for transformer-based masked language-models, 2022","cited_arxiv_id":null,"evidence_quote":"BitFit is the bias-only PEFT baseline that Swin-TUNA outperforms on both datasets."},{"cited_title":"Mopeft: A mixture-of-pefts for the segment anything model, 2024","cited_arxiv_id":null,"evidence_quote":"This cited existing PEFT attempt for SAM supports the paper's claim that prior PEFT methods had not surpassed full fine-tuning in segmentation accuracy."}],"review_version":1}