{"id":"217feac1-71c9-4fe8-9bc5-aa7a7f656a3f","arxiv_id":"2506.12737","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A unified feature codec for CNN and ViT features, built with format and value alignment, beats an architecture-specific baseline on ImageNet classification.","lead":"This paper proposes a single compression codec that can handle features from both CNNs and Transformers by reshaping them into the same 2D token format and then normalizing their value ranges. A smart generalist might care because it addresses a practical problem: real systems increasingly mix model architectures, so sharing one codec across them could cut storage and bandwidth.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed superiority over an architecture-specific baseline is not yet established: the ResNet comparison uses a DINOv2-oriented codec forced into the same 2D format, and part of the gain is attributed to training schedule rather than the proposed alignment.","rationale":"The reader correctly notes that the method relies on hand-chosen constants and a single architecture pair, making the 'universal' claim fragile. My reading goes one step further: even the narrow superiority claim over an architecture-specific baseline is not cleanly demonstrated, because the ResNet baseline is a large-model-oriented codec adapted to 2D rather than a native CNN feature codec, and the reported gains are partly attributed to training schedule choices. This does not disprove the method; it means the current evidence supports only a conditional acceptance pending a fairer baseline and a held-out evaluation. I therefore keep the reader's CONDITIONAL verdict and mark partial agreement, since the specific load-bearing weakness I identify is the comparison protocol rather than constant generalization alone.","tokens_in":11227,"tokens_out":6997,"duration_ms":81188,"concrete_test":"Train a true architecture-specific learned CNN codec on the original ResNet-50 C5 features (3D, 2048×7×7), using the same Hyperprior backbone, training image count, optimizer, and termination rule as the proposed method, and compare its rate-accuracy curve against the proposed universal codec on the held-out ImageNet classification features. Also compute the same comparison for DINOv2 against the [18] baseline. If the universal codec does not beat the native 3D ResNet codec or the advantage reverses when training protocols are matched, the central claim of superiority over architecture-specific coding fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (Abstract; Sec. 4) is that after format and value alignment, one Hyperprior codec beats an architecture-specific codec for both ResNet-50 and DINOv2 features. The evidence in Sec. 3.2-3.3 does not support this cleanly. For ResNet features, the baseline is not a CNN-specific codec: Sec. 3.2 says existing CNN feature coding methods require 3D inputs and are incompatible with DINOv2, so the authors compare against [18], a codec designed for large-model/Transformer features. Both methods then use shifted normalization; this forces the baseline to operate on the same 2D token format rather than testing a true architecture-specific codec on the native 2048×7×7 feature tensor. The observed gain on ResNet could reflect the baseline's format mismatch rather than an advantage of the universal alignment. Moreover, Sec. 3.3 attributes part of the DINOv2 improvement to 'optimized training strategies, particularly training data sampling and learning rate-based training termination criteria,' not to alignment; the lambda-to-sampling-ratio schedule (Sec. 2.3) is a training choice that is not part of the alignment pipeline. Without holding training protocol fixed or comparing against a properly trained 3D CNN feature codec, the rate-accuracy advantage cannot be assigned to the proposed method. The truncation and normalization constants are also selected on the same ImageNet feature distribution used for evaluation, so the 'universal' framing remains untested.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces cross-architecture universal feature coding (CAUFC), proposing a single Hyperprior-based codec that compresses both CNN and ViT features. The method first aligns formats by reshaping CNN activations from N×H×W into (H×W)×N 2D tokens, then aligns value ranges through truncation and a shifted normalization designed to make ResNet features resemble DINOv2 features. The codec is trained on mixed ResNet-50 and DINOv2 features from ImageNet, with bitrate-dependent sampling ratios and λ values. Experiments on image classification report rate-accuracy comparisons against the large-model feature codec of [18] for both feature types, claiming superior overall performance. The paper also includes an ablation on truncation ranges for DINOv2 features. The central claim is that, after these alignments, one codec can outperform an architecture-specific baseline for both ResNet-50 and DINOv2 features.","tokens_in":11528,"tokens_out":4866,"duration_ms":56911,"significance":"The problem formulation is timely and well motivated: real systems do mix CNN and Transformer backbones, and a universal feature codec would be practically useful. The proposed alignment steps are simple and intuitively reasonable, and the paper demonstrates, at least visually, that the shifted normalization makes the two feature distributions more similar. The paper is also honest about some limitations, explicitly noting that the DINOv2 medium-bitrate result is worse than baseline and attributing part of the gain to training strategy. However, the current evidence is insufficient to support the strong claim of superiority over an architecture-specific baseline. The comparison against [18] is not a clean architecture-specific baseline for CNN features, training protocol is not held fixed, and all alignment constants are selected on the same ImageNet test distribution used for evaluation. The evaluation covers only one task, one CNN, one ViT, and one test dataset, with no error bars. With additional controlled experiments, the central idea could be valuable, but as presented the claim is not yet established.","major_comments":[{"comment":"The claim that the proposed method beats an architecture-specific baseline is not supported by the current comparison. For ResNet features, the baseline [18] is designed for large-model/Transformer features, not for CNN feature coding; existing CNN feature codecs require 3D inputs and are not used. Both methods are forced into the same 2D token format with shifted normalization, so the reported ResNet gains may reflect the baseline's format mismatch rather than an advantage of the proposed alignment. Please compare against a CNN-specific codec operating on the native 2048×7×7 activation tensor, or at minimum against a 2D-adapted codec trained separately on ResNet features in the same preprocessed format, with the same training schedule and termination criteria.","section":"Sec. 3.2, 3.3, Fig. 5"},{"comment":"The attribution of the improvement to the proposed method is confounded by training protocol. The DINOv2 comparison uses the same preprocessing and Hyperprior architecture for both methods, and the paper explicitly credits the DINOv2 gains to 'optimized training strategies, particularly in terms of training data sampling and learning rate-based training termination criteria.' These are choices outside the proposed alignment pipeline (the sampling-ratio schedule is introduced in Sec. 2.3 as a training choice, not as part of format or value alignment). Please include an ablation that holds the training schedule and sampling ratio fixed and adds the alignment components one at a time, or report the baseline retrained under exactly the proposed schedule without alignment.","section":"Sec. 3.3, Sec. 2.3"},{"comment":"The 'universal' claim is currently untested across data distributions. The truncation ranges ([-5,5] for DINOv2 and [0,5] for ResNet), the shifted-normalization lower bound (set to the DINOv2 minimum), and the λ-to-sampling-ratio schedule (1:1, 1:2, 1:3, 1:5) are all hand-chosen on the ImageNet feature statistics used for both training and evaluation. No held-out feature distribution is used to select or validate these constants. To support the central generalization claim, please evaluate the same fixed constants on additional architectures (e.g., a different ViT or CNN) or on an out-of-domain dataset such as Places365, and report whether the rate-accuracy ordering is preserved.","section":"Sec. 2.2, Sec. 3.1, Sec. 3.3"},{"comment":"The evidence is quantitatively thin: there are no error bars, repeated runs, or integrated rate-accuracy metrics (e.g., BD-rate or area under the curve), and the DINOv2 comparison shows a medium-bitrate regime where the proposed method is marginally worse than baseline. Please report multiple training runs with standard deviations and provide a quantitative summary of the gain rather than relying on visual inspection of two curves. Also specify the size of the test set and the number of feature points per bitrate point, so that the significance of the accuracy differences can be assessed.","section":"Sec. 3.3, Fig. 5"},{"comment":"The semantic equivalence assumption underlying the proposed tokenization is load-bearing but not directly verified. The paper asserts that each CNN channel corresponds to a semantic attribute in the same way that each ViT token dimension does, and that sampling one attribute from each channel therefore preserves semantic alignment. This is plausible but not demonstrated. Please provide a quantitative or experimental check, for example by comparing the proposed channel-to-token reshaping against a flattening or a random permutation of channels in terms of downstream classification accuracy, or by measuring cross-architecture similarity between the resulting token dimensions.","section":"Sec. 2.1, Fig. 2"},{"comment":"The ablation in Sec. 3.4 studies truncation ranges for DINOv2 features but does not isolate the contributions of the three main components (format alignment, value alignment, and the shifted-normalization lower bound) in the final system. The rate-accuracy comparison in Fig. 4 is run with 'baseline Hyperprior models' and does not control for the training schedule. Please add a full ablation table, e.g., rows for (i) no alignment, (ii) format alignment only, (iii) standard normalization only, (iv) shifted normalization, (v) truncation only, and (vi) the full method, with the training protocol fixed across rows.","section":"Sec. 3.4"}],"minor_comments":[{"comment":"Several figures contain garbled text, with sequences such as '/uni00000013/uni00000015...' appearing in captions and axes. Please regenerate the figures with readable labels, including axis titles for the rate-accuracy plots.","section":"Figures 2-7"},{"comment":"The phrase 'lowerandupper' is missing a space; it should read 'lower and upper'.","section":"Sec. 2.2"},{"comment":"It is not stated which feature type corresponds to the first and second number in the 1:1, 1:2, 1:3, and 1:5 sampling ratios. Please state explicitly whether the ratio is DINOv2:ResNet or ResNet:DINOv2.","section":"Sec. 2.3, Sec. 3.1"},{"comment":"The test dataset is described only as 'the classification dataset proposed in [18]'. Please give the number of test images and the class composition, and clarify whether the original accuracies 86.4% and 100% are computed on clean images or under the same distortion conditions.","section":"Sec. 3.2"},{"comment":"The loss function L = BPFP + λ×||X − X_hat||^2 is not fully specified. Please define BPFP precisely (e.g., expected bits per feature point over the batch) and state how X and X_hat are preprocessed before the distortion term is computed.","section":"Sec. 2.3"},{"comment":"Reference [18] is an arXiv preprint by a largely overlapping set of authors. If a peer-reviewed version exists, it should be cited; otherwise, more detail on the baseline's architecture and training should be provided so that the comparison is reproducible.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The baseline [18] shares authors with the current submission. This is not itself a correctness issue, but it strengthens the need for a fully independent baseline and for access to the baseline implementation. The paper's strongest contribution is the problem formulation and the simple alignment idea; the current evaluation does not yet rule out the alternative explanation that the gains come from the training schedule and from comparing a universal codec against a Transformer-oriented codec on CNN features. I would urge the editor to require cross-architecture and cross-dataset generalization experiments before accepting the 'universal' framing."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the paper names a real gap and proposes a clean, simple fix, but the headline comparison does not yet support the claim of beating architecture-specific codecs.\n\nWhat is actually new: cross-architecture universal feature coding — one Hyperprior-based codec for both CNN and ViT features. The format alignment (reshaping CNN feature maps into tokens) and the shifted normalization are straightforward preprocessing moves, but the framing and the specific combination are new, and the paper is honest about being a first attempt. The distribution visualizations in Figs. 2 and 3 are useful evidence that the alignment does what it claims.\n\nThe main soft spot is the baseline. For ResNet features, the comparison is against the large-model/Transformer codec from [18], which the authors themselves note is incompatible with 3D CNN features. Forcing that baseline onto 2D tokens does not test whether a universal codec beats a truly architecture-specific CNN codec; it tests whether a Transformer-oriented codec works on CNN features. That is the reader's and the stress-test's point, and it lands. Second, in Sec. 3.3 the authors attribute part of the DINOv2 gain to training schedules — sampling ratios and learning-rate termination — not to alignment. Since the lambda-to-sampling-ratio schedule is a training choice outside the alignment pipeline, the gain is confounded. Third, the evaluation is one task, one dataset, one pair of architectures, and no error bars; the DINOv2 curve is worse than baseline at medium bitrates. The truncation constants are hand-selected on the same ImageNet feature distribution used for evaluation, so the word 'universal' is doing more work than the evidence can support.\n\nNone of these problems are disqualifying. The problem definition is worth having, the recipe is reproducible enough to test, and the ablations show the alignment has a real effect. What is missing is a proper 3D CNN baseline, a controlled training protocol comparison, more architectures/datasets, and code or data release.\n\nI would take this as a serious submission — the problem is important enough to warrant referee time — but I would expect heavy revision before accepting the central claim. For my own work, I would not cite it yet as evidence of rate-accuracy superiority; I would cite it as early problem framing once the experiments catch up.\n\nRecommendation: send to peer review, with a request for the missing baselines and controlled comparisons.","headline":"A plausible first step on cross-architecture feature coding, but the headline claim of beating architecture-specific codecs is not yet supported by the experiments.","tokens_in":12061,"tokens_out":3151,"would_cite":false,"duration_ms":28332,"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":"The paper claims that a single learned codec can compress CNN and Vision Transformer features for image classification after aligning their formats and value distributions, beating the architecture-specific baseline.","keywords":["cross-architecture universal feature coding","feature compression","coding for machines","distribution alignment","CNN features","Transformer features","Hyperprior","image classification"],"falsifier":"Run the same fixed alignment constants and codec on a held-out distribution such as Places365 or on different backbone pairs (for example Swin Transformer and ConvNeXt), and compare against separately trained architecture-specific codecs; if the unified codec no longer matches or beats those baselines, the universality claim fails.","tokens_in":1447,"feed_emoji":"🗜️","tokens_out":1558,"duration_ms":56038,"temperature":0.7,"pith_summary":"The paper introduces a new research problem, cross-architecture universal feature coding (CAUFC), and claims that a single learned codec can compress features from both CNNs and Vision Transformers after their formats and value distributions are aligned. The authors show this on image classification with ResNet-50 and DINOv2 features, using a Hyperprior model trained on a mixture of both feature types. They report that the unified codec achieves better rate-accuracy trade-offs than the existing architecture-specific baseline, with the largest gains at low bitrates. The point of this direction, if it holds, is that mixed-architecture systems can share one compression model instead of maintaining separate codecs per backbone.","feed_headline":"One codec compresses both CNN and ViT features","feed_subtitle":"Format and value alignment let a single Hyperprior model beat separate feature codecs on ImageNet classification.","key_machinery":"The load-bearing mechanism is a two-step distribution alignment that makes heterogeneous features look alike to a downstream entropy model. Format alignment reinterprets each CNN feature map location as a token spanning all channels, matching the token structure of ViTs; value alignment then clips extremes and maps both feature families into a shared [0,1] range, with a shifted lower bound for ResNet features to push their dominant mass toward DINOv2's. This preprocessing lets an unmodified Hyperprior model with one input channel serve as the unified codec, trained on a mixed dataset whose DINOv2-to-ResNet sampling ratio is scheduled by bitrate.","core_discovery":"The central claim is that cross-architecture universal feature coding is feasible: a single Hyperprior-based codec, once the two feature families are made statistically and structurally comparable, can compress ResNet-50 C5 features (2048×7×7) and DINOv2 layer-40 features (257×1536) for image classification with better rate-accuracy performance than a codec tailored to large-model features. Format alignment reshapes CNN feature maps into (H×W)×N tokens, treating each spatial location as a token whose dimensions are the channels, and leaves ViT tokens as M×L. Value alignment truncates DINOv2 features to [-5,5], ResNet features to [0,5], then normalizes both to [0,1], with ResNet's lower bound shifted to DINOv2's minimum so the two distributions occupy similar regions. Training mixes both feature types with sampling ratios that depend on the target bitrate. The paper reports consistent accuracy gains across bitrates for ResNet features and at low and high bitrates for DINOv2 features, with a small loss at medium bitrates.","pith_inferences":["The alignment constants were chosen on the same ImageNet test distribution used for evaluation; generalization to other datasets or backbones is untested, so 'universal' should be read as demonstrated for this particular pair of models and this task.","The semantic claim that CNN channels correspond to ViT token dimensions is assumed rather than proven; a direct probe or decoding experiment could test whether spatial-token alignment preserves task-relevant information across backbones.","A data-driven version that learns truncation bounds and normalization parameters from a small calibration set could replace the hand-set constants and likely transfer better.","The comparison to a single baseline leaves open whether separately trained, equally tuned architecture-specific codecs would close the gap; testing that pair would isolate the benefit of unification."],"forward_implications":["A single codec can serve mixed CNN/ViT deployments, removing the need for separate architecture-specific feature codecs.","Truncation range is an effective rate-accuracy control: wider ranges help at moderate-to-high bitrates, while narrower ranges can seriously degrade accuracy, especially for ResNet features.","Training-data sampling ratio between feature types is a practical lever: favoring DINOv2 at high bitrates and ResNet at low bitrates improves the overall rate-accuracy trade-off.","The same alignment recipe can in principle be carried to other feature-extraction backbones beyond ResNet-50 and DINOv2.","Replacing two codecs with one reduces storage and deployment overhead in collaborative intelligence systems."],"supporting_citations":[{"why":"Supplies the large-model feature coding benchmark, test conditions, and the architecture-specific baseline the proposed method must beat.","marker":"[18]"},{"why":"Provides the Hyperprior variational autoencoder used as the unified feature codec.","marker":"[19]"},{"why":"Defines the DINOv2 model and features that constitute the Transformer side of the experiments.","marker":"[16]"}],"fun_headline_variants":["One codec, two architectures: feature compression unified","Universal feature codec aligns CNN and ViT for better compression","Single codec for all features: alignment beats specialized models","Cross-architecture feature coding: one model compresses both","Feature codec that speaks both CNN and Transformer"],"cache_read_input_tokens":14080,"weakest_assumption_plain":"The load-bearing premise is that the hand-set alignment constants tuned on the ImageNet test distribution—truncation ranges, normalization bounds, and sampling ratios—generalize without retuning, and that treating CNN channels as equivalent to ViT token dimensions is valid.","fun_headline_variants_meta":{"raw":{"variants":["One codec, two architectures: feature compression unified","Universal feature codec aligns CNN and ViT for better compression","Single codec for all features: alignment beats specialized models","Cross-architecture feature coding: one model compresses both","Feature codec that speaks both CNN and Transformer"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000597,"raw_usage":{"total_tokens":2793,"prompt_tokens":942,"completion_tokens":1851,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":558,"completion_tokens_details":{"reasoning_tokens":1772}},"tokens_in":558,"tokens_out":1851,"duration_ms":13440,"temperature":1.0,"reasoning_tokens":1772,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T00:42:31.652223+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same fixed alignment constants and codec on a held-out distribution such as Places365 or on different backbone pairs (for example Swin Transformer and ConvNeXt), and compare against separately trained architecture-specific codecs; if the unified codec no longer matches or beats those baselines, the universality claim fails.","supporting_citations":[{"cited_title":"Cross-architecture universal feature coding via distribution alignment","cited_arxiv_id":"2506.12737","evidence_quote":"Supplies the large-model feature coding benchmark, test conditions, and the architecture-specific baseline the proposed method must beat."},{"cited_title":"IMOFC: identity-level metric optimized feature compression for identification tasks,","cited_arxiv_id":null,"evidence_quote":"Provides the Hyperprior variational autoencoder used as the unified feature codec."},{"cited_title":"Overview of the MPEG-CDVS standard,","cited_arxiv_id":null,"evidence_quote":"Defines the DINOv2 model and features that constitute the Transformer side of the experiments."}],"review_version":1}