{"id":"2177b65c-41e1-4f1e-99e5-f217c3e397a7","arxiv_id":"2412.06875","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"VQ4ALL builds a single universal codebook from the weight distributions of several networks, then learns per-network assignments to reconstruct low-bit weights while keeping accuracy close to the original models.","lead":"This paper describes VQ4ALL, a compression method that lets multiple neural networks share one universal codebook of weight patterns, reaching high compression with small accuracy loss. It matters because devices running several models could store a single shared code table instead of one per network, saving memory and simplifying hardware.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported '>16×' compression omits the 2–8 MB universal codebook; under standard accounting a 2-bit ResNet-18 is ~9×, not 16×.","rationale":"The reader's weakest assumption identifies precisely the load-bearing gap: the codebook is excluded from the per-network compression ratio without specifying the amortization. My independent reading confirms this is the most serious issue because the abstract and Table 3 make an unqualified quantitative claim, and the arithmetic shows the 16× and 32× ratios are not reproducible under standard accounting. The paper's hardware motivation (ROM sharing across multiple networks) can justify reporting marginal storage per additional network, but that requires explicit statement of the amortization assumption and aggregate numbers, which are absent. A secondary concern is that 'universal' is only tested on the architectures used to build the codebook (Table 6), so generalization to held-out architectures remains unverified; this reinforces the conditional verdict but is not needed to establish the accounting problem. The reader's CONDITIONAL verdict is appropriate: the method may be effective and the hardware story plausible, but the central compression claim must be restated or supported with full accounting. I therefore see no reason to change the verdict.","tokens_in":14822,"tokens_out":7990,"duration_ms":86160,"concrete_test":"Recalculate the compression ratios in Tables 2 and 3 using the full formula from Section 3.1: compressed size = sum of index bits + k × d × 32 bits for the universal codebook. Then, for the multi-network setting, report the smallest number N of networks for which (sum of full-precision sizes) / (sum of index sizes + codebook size) reaches the claimed 16×; if that N exceeds the number of evaluated networks or is unreachable, revise the headline claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim is built on compression ratios that count only index storage and exclude the universal codebook. In Section 5, the codebook is described as 'negligible' and 'stored as static code tables in the built-in ROM,' and the ratios in Table 3 are labeled as 'the average compression rate of compressed layers.' For 2-bit ResNet-18, the universal codebook has k=2^16, d=8, and 32-bit entries, so it occupies 2 MB. The compressed index stream is about 2 bits per weight, or roughly 2.9 MB for the compressed portion of ResNet-18, while the full-precision model is about 46.8 MB. Adding the codebook gives a total compressed size near 4.9 MB and a true compression ratio of about 9.5×, not 16×. At 1-bit, the codebook is 4 MB (k=2^16, d=16), and the reported 32× ratio drops below 9×. The paper never specifies the number of networks needed for the codebook to amortize to the claimed ratios, nor does it report aggregate storage across the evaluated network set. Without that accounting, the headline 'compression rates exceeding 16×' is not supported under the standard definition of model size for a single network.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces VQ4ALL, a vector quantization method that uses a single shared 'universal codebook' to represent many low-bit neural networks. The codebook is initialized by kernel density estimation over weight sub-vectors sampled from the networks to be compressed, and each network is then constructed by learning differentiable soft assignments to candidate codewords, with a Progressive Network Construction (PNC) strategy that gradually freezes assignments. Experiments on ResNet-18/50, MobileNet-V2, Mask R-CNN, and Stable Diffusion report state-of-the-art accuracy at low bit-widths and claim compression rates exceeding 16x. The central claim depends on counting only the compressed index storage and excluding the universal codebook from the per-network model size, with the codebook assumed to be stored once in ROM and shared across an unspecified number of networks.","tokens_in":15125,"tokens_out":4974,"duration_ms":52304,"significance":"If the compression accounting is stated correctly, VQ4ALL is a meaningful step toward efficient multi-network deployment: a single code table in ROM promises large savings in memory access, codebook count, and silicon area, and the experimental results are strong across classification, detection, and generation. The paper deserves credit for a clean ablation study (Table 5, Figure 3) showing the contribution of PNC and candidate assignments, and for Table 6, which demonstrates that the codebook is not overly sensitive to the exact combination of source networks. The method is also practical: it runs on a single GPU and does not require the massive compute reported for some baselines. However, the headline compression ratios and the 'universal' characterization are currently overstated relative to the evidence, primarily because the codebook size is omitted from per-network size and because the codebook is always sampled from the very networks being evaluated.","major_comments":[{"comment":"The headline claim of 'compression rates exceeding 16x' is not supported under the standard definition of model size for a single network, because the universal codebook is excluded from the reported per-network size. The paper states in Section 5.1 that the universal codebook is 'stored as static code tables in the built-in ROM' and that Table 3 reports 'the average compression rate of compressed layers,' but it never adds the codebook back into the deployed model size or provides an amortization analysis over the number of networks sharing it. Concrete recalculation: for 2-bit ResNet-18 (k=2^16, d=8, 32-bit entries), the universal codebook is 2 MB; the compressed index stream for the compressed portion is about 2.9 MB, and the full-precision model is about 46.8 MB, so the true single-network compression ratio is about 9.5x, not 16x. At 1 bit (k=2^16, d=16), the codebook is 4 MB and the reported 32x ratio drops below 9x. The paper must disclose the codebook size in every compression-ratio report, state how many networks are required for the amortized codebook cost to become negligible, and present a comparison with DKM and other baselines using the same accounting for both methods.","section":"Abstract, Section 5.1, Table 3"},{"comment":"The claim that the codebook is 'universal' is only tested on networks whose weights were used to build it. The method description says the KDE samples from 'the networks mentioned above,' and the experimental setup in Section 5 states that sub-vectors are extracted from exactly the networks evaluated in Sections 5.1-5.3. Table 6 varies the combination of these same four networks, which shows robustness to the composition of the source set but does not test generalization to an architecture or dataset not seen during codebook construction. To support the 'universal' and 'applies to most deep neural networks' claims, the paper should evaluate a held-out network (e.g., ResNet-34, ViT-Tiny, or a different detection backbone) using a codebook built only from the other networks and report accuracy and compression under the same accounting.","section":"Section 4.1, Section 5.4, Table 6"}],"minor_comments":[{"comment":"There is an internal contradiction about which layers are compressed in Table 3. The text says 'We do not compress the input layers ... The output layer is constructed using a small per-layer codebook derived from clustering its weights,' but later says 'Following the configuration of EWGS, none of the experiments in Table 3 compress the input and output layers of the network.' Please clarify which layers are actually included in the reported compression ratios.","section":"Section 5.1"},{"comment":"Equation (11) is typeset with unbalanced parentheses and an ambiguous denominator; it should be written as L_r = n * sum_{o,i/d,n} (r(1-r) / (o*i/d)) or equivalent, with a clear definition of the summation range.","section":"Equation (11)"},{"comment":"The update rule 'R ← R - u (∂/∂L, θ)' is not well-formed; it should explicitly state that R is updated with an optimizer u using the gradient ∂L/∂R and hyperparameters θ.","section":"Section 4.2"},{"comment":"In Table 1, the units of column 'C' are not specified (entries versus bytes), and the 'I/O' metric is not defined; please state how the number of codebook accesses is counted for the per-layer and universal variants.","section":"Table 1"},{"comment":"The text says that VQ4ALL 'continues to hold an accuracy close to 67%' when the compression ratio 'exceeds 40x,' but no table reports a 40x ratio; Figure 2 appears to show values in that range, so please label the figure axis clearly as applying to the compressed layers only and reconcile the numbers with the text.","section":"Section 5.1"}],"recommendation":"major_revision","confidential_remarks":"The paper is technically interesting and likely publishable after a revision that corrects the compression-ratio accounting and adds a genuine out-of-sample test for the universal codebook. The omitted codebook size is a load-bearing issue, but it is fixable by reporting total deployed size and amortization assumptions; I do not see evidence of a method flaw that would require rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here is the short version. VQ4ALL has a genuinely new idea—one frozen codebook shared across multiple network architectures, built from a kernel density estimate of the weights, with progressive hardening of soft assignments. The accuracy numbers are good, especially at 1–2 bits, and the ablation study is honest and informative. But the headline '>16× compression' only holds if you exclude the universal codebook from the model size. Under standard accounting, a 2-bit ResNet-18 comes out around 9.5×, not 16×, and at 1-bit it is below 9×. That is the paper's main weakness, and it is fixable.\n\nWhat is new: the shared codebook is a real departure from per-layer and per-network VQ methods like DKM and PQF. The KDE-based sampling is a nice trick—it avoids expensive k-means on the concatenated weights—and the PNC strategy does a good job of preventing the accuracy collapse that happens when you harden assignments all at once. The experiments are broad: ResNet-18/50, MobileNet-V2, Mask R-CNN, and Stable Diffusion, with reasonable baselines. I see no reason to doubt the accuracy results; they are internally consistent, and the comparison with DKM appears fair.\n\nThe soft spot is the compression-ratio accounting. The paper says the universal codebook is 'negligible' and Table 3 reports 'the average compression rate of compressed layers.' But the codebook is not negligible for a single network: with k=2^16 and d=8, it occupies 2 MB. For ResNet-18 the index stream is roughly 2.9 MB, so the total compressed size is about 4.9 MB versus 46.8 MB full precision—roughly 9.5×, not 16×. At 1-bit, d=16 makes the codebook 4 MB and the ratio drops further. The paper never says how many networks are needed for the codebook to amortize to the claimed ratios, nor does it report aggregate storage across the whole network set. That is a load-bearing assumption behind the headline.\n\nA smaller issue: 'universal' is only demonstrated on the networks used to build the codebook. Table 6 shows the method is fairly robust to which networks contribute to the KDE, but there is no test on a held-out architecture. That would strengthen the universality claim.\n\nBottom line: this is a serious paper with a useful method and solid evidence, but the headline compression ratio is overstated. It deserves a real review—the right response is a request for codebook-inclusive ratios, an amortization analysis, and a held-out architecture test, not a desk reject. If you work on low-bit VQ, read it and cite it, but quote the corrected ratios.","headline":"A genuinely novel shared-codebook VQ method with strong low-bit accuracy, but the '>16×' compression headline only survives if you don't count the codebook—standard accounting puts a 2-bit ResNet-18 near 9.5×.","tokens_in":15612,"tokens_out":3862,"would_cite":true,"duration_ms":34440,"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 single universal codebook compresses multiple network architectures past 16x while keeping accuracy near the floating-point original.","keywords":["vector quantization","universal codebook","network compression","kernel density estimation","progressive network construction","low-bit networks","knowledge distillation","image classification"],"falsifier":"Compute the deployed size of a single network as its assignment indices plus its share of the universal codebook, and compare with the floating-point model; for the paper's 2-bit ResNet-18 with a $2^{16}\\times 8$ codebook, the ratio drops from 16x to about 9x, so a direct measurement of single-network or few-network storage would settle whether the amortization assumption holds.","tokens_in":14636,"feed_emoji":"🗜️","tokens_out":8041,"duration_ms":75248,"temperature":0.7,"pith_summary":"VQ4ALL tries to establish that many different neural networks can be compressed from one shared, frozen codebook rather than from per-layer code tables. The paper builds a single codebook by kernel density estimation over weight sub-vectors sampled from several networks, then learns differentiable assignment ratios per network to reconstruct low-bit weights. If the claim is right, model fleets no longer multiply codebook storage: one static table can live in ROM, and switching tasks only swaps assignment indices. The reported support is compression above 16x with accuracies close to the floating-point baselines on ImageNet classification, COCO object detection, and text-to-image generation.","feed_headline":"One shared codebook compresses many networks past 16x","feed_subtitle":"One ROM-resident code table replaces per-layer codebooks, cutting memory traffic and silicon area.","key_machinery":"The carrying mechanism is the universal codebook $C$ of $k$ codewords of dimension $d$, sampled from the kernel density estimate $f(w)$ of all networks' weight sub-vectors and then frozen. Each sub-vector keeps the $n$ nearest codewords as candidates, with softmax ratios $R$ initialized inversely to Euclidean distance, so the reconstructed weight $\\widehat{W} = R C[A_c]$ is differentiable and gradients can move the ratios without touching the codebook. The Progressive Network Construction strategy then converts ratios above threshold $\\alpha$ (0.9999) into one-hot optimal assignments gradually, which prevents the accuracy collapse observed when all assignments are switched at once.","core_discovery":"The central claim is that universal vector quantization, meaning one codebook shared by all networks, can replace per-layer codebooks without sacrificing accuracy. VQ4ALL extracts the codebook from a kernel density estimate of the concatenated weight sub-vectors of ResNet-18/50, MobileNet-V2, Mask R-CNN, and Stable Diffusion, freezes it, and represents each sub-vector by a shortlist of candidate codewords with softmax ratios. The ratios are trained against task loss, block-wise knowledge distillation, and a regularization term; a Progressive Network Construction strategy freezes each assignment to a one-hot choice once its ratio crosses a high threshold. With the codebook excluded from per-network size, the method reports 11x, 16x, and 32x compression at 3-bit, 2-bit, and 1-bit for classification, 26.3x for Mask R-CNN, and usable 2-3 bit Stable Diffusion, with accuracy near each floating-point original.","pith_inferences":["Editorial inference: the per-network model size excludes the universal codebook, so the honest single-network ratio requires adding the codebook; the paper does not state how many networks are needed for the amortization to make that term negligible.","Editorial inference: if the codebook is truly reusable across architectures, the idea points toward a hardware primitive, one ROM code table plus per-network assignment storage, that would make the marginal storage for a new model very small.","Editorial inference: because only assignment learning is needed for a new network, the method could be tested as a continual-learning setting where new task models are added against a fixed codebook without retraining it.","Editorial inference: the same KDE-sampled universal codebook could be tried on other large-model families such as language models, where sub-vector statistics may also be shared across checkpoints."],"forward_implications":["A deployed fleet of compressed networks can be served from one ROM-resident code table, so adding another network adds only assignment indices, not another codebook.","Memory access and silicon area for codebooks stop growing with the number of networks, since the universal table is loaded once rather than per layer or per model.","Extremely low-bit compression remains usable: at 1-bit, VQ4ALL reports 68.0% Top-1 on ResNet-18, 74.7% on ResNet-50, and 60.4% on MobileNet-V2 at 32x compression.","The same codebook transfers across task families, with 26.3x compression on Mask R-CNN detection and segmentation and 2-3 bit Stable Diffusion generation staying close to the floating-point FID and CLIP scores."],"supporting_citations":[{"why":"Supplies the differentiable k-means formulation and the soft-to-hard assignment transition that VQ4ALL builds on and modifies.","marker":"[2]"},{"why":"The prior VQ4DiT assignment-search idea that this paper extends from diffusion transformers to a universal multi-network codebook.","marker":"[4]"},{"why":"Introduces vector quantization for network weight compression, the foundation for codebook-based representation.","marker":"[8]"},{"why":"The uniform-quantization baseline whose low-bit classification numbers VQ4ALL compares against in Table 3.","marker":"[14]"},{"why":"The per-layer VQ baseline on object detection and image generation that VQ4ALL compares against in Tables 2 and 4.","marker":"[23]"},{"why":"The product-quantization baseline on classification and detection used as a comparison point throughout the experiments.","marker":"[32]"}],"fun_headline_variants":["One universal codebook slashes model size 16x+","VQ4ALL: share one codebook, compress many nets","ROM codebook shrinks all networks past 16x","Common codebook yields 16x+ compression across nets"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The headline compression ratios count the universal codebook as free, assuming it is stored once in ROM and shared by so many networks that its size can be left out of each network's reported size.","fun_headline_variants_meta":{"raw":{"variants":["One universal codebook slashes model size 16x+","VQ4ALL: share one codebook, compress many nets","ROM codebook shrinks all networks past 16x","Common codebook yields 16x+ compression across nets"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000165,"raw_usage":{"total_tokens":1252,"prompt_tokens":951,"completion_tokens":301,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":567,"completion_tokens_details":{"reasoning_tokens":231}},"tokens_in":567,"tokens_out":301,"duration_ms":3426,"temperature":1.0,"reasoning_tokens":231,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T19:27:13.341015+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the deployed size of a single network as its assignment indices plus its share of the universal codebook, and compare with the floating-point model; for the paper's 2-bit ResNet-18 with a $2^{16}\\times 8$ codebook, the ratio drops from 16x to about 9x, so a direct measurement of single-network or few-network storage would settle whether the amortization assumption holds.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The uniform-quantization baseline whose low-bit classification numbers VQ4ALL compares against in Table 3."},{"cited_title":"Permute, quantize, and fine-tune: Efficient compression of neural networks","cited_arxiv_id":null,"evidence_quote":"The per-layer VQ baseline on object detection and image generation that VQ4ALL compares against in Tables 2 and 4."}],"review_version":1}