{"id":"c178a2d2-4136-47e6-a446-d04313e8feff","arxiv_id":"2412.14633","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A progressive fine-to-coarse reconstruction schedule with a two-stage optimization strategy improves 3-bit and 4-bit post-training quantization accuracy for vision transformers, achieving 75.61% Top-1 on ViT-B at 3 bits.","lead":"This paper proposes a progressive fine-to-coarse reconstruction method for low-bit post-training quantization of vision transformers, claiming state-of-the-art accuracy including 75.61% Top-1 on 3-bit ViT-B. It matters because accurate low-bit quantization could enable efficient deployment of vision transformers on memory-limited devices.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 3's block-wise baseline (36.58%) is ~38 points below the same paper's I&S-ViT (74.87%) for 4-bit ViT-S, so the claimed 37.45% PFCR gain is inflated and the central SOTA claim needs qualification.","rationale":"Read in good faith: the paper proposes a curriculum of reconstruction granularities for PTQ of ViTs, and reports large gains over SOTA in Table 1 (e.g., 75.61% 3-bit ViT-B). For this claim to hold, two things must be true: (a) the reported Table 1 numbers are reproducible, and (b) the ablation in Table 3 fairly attributes the gain to PFCR/POS. The weakest place is (b): Table 3's block-wise baseline is 36.58% for 4-bit ViT-S, while the same paper lists I&S-ViT — a block-wise method — at 74.87%. Since the ablation is conducted under the authors' 'same setting' (default 300 iterations), it likely does not represent a strong or SOTA block-wise baseline. A fair comparison would be against I&S-ViT's published numbers, where the PFCR gain is only ~1.4 points. This does not refute the method's plausibility, but it undercuts the quantitative magnitude claim. Additionally, the abstract's 'best' claim is contradicted by two 6-bit DeiT entries in Table 1, a limitation the body partially acknowledges in §4.2. The reader's weakest_assumption identifies the same baseline issue, and we agree. A concrete test: rerun the block-wise baseline with I&S-ViT's official settings or code; if it recovers ~74.87%, Table 3's 37.45% improvement is an artifact. The verdict should remain CONDITIONAL, pending fair baselines and code release.","tokens_in":17110,"tokens_out":9869,"duration_ms":60294,"concrete_test":"Run the same block-wise reconstruction (BRECQ-style, no PFCR/POS) on 4-bit ViT-S under the I&S-ViT hyperparameters from Table 1 (or with the official I&S-ViT code). If the baseline reaches ~74.87% rather than 36.58%, the 37.45% improvement in Table 3 is inflated; re-report Table 3 with a fair baseline and recompute the claimed gain. Also verify whether 6-bit DeiT-S/DeiT-B remain non-SOTA.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.4 (Table 3) reports a 'Block-wise' 4-bit ViT-S baseline of 36.58% and claims PFCR improves it by 37.45%. But Table 1 lists I&S-ViT, a block-wise reconstruction method, at 74.87% for the same model and bit-width. The reproduced baseline therefore appears to be a weak or under-trained implementation (using the paper's default 300 iterations and 4e-5 learning rate), not a representative block-wise baseline. Against a strong block-wise baseline (I&S-ViT), PFCR's gain in Table 1 is only 76.31 vs 74.87 = +1.44. Thus the paper's magnitude claim for the benefit of progressive fine-to-coarse reconstruction is unsupported. In addition, the abstract's unqualified 'best Top-1 accuracy among SOTA' is contradicted by Table 1 at 6-bit DeiT-S (79.34 vs Outlier-Aware 79.50) and 6-bit DeiT-B (81.65 vs I&S-ViT 81.68). The core idea is plausible, but the quantitative support is fragile without a fair baseline.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Progressive Fine-to-Coarse Reconstruction (PFCR) for post-training quantization of Vision Transformers. The method defines MHSA and MLP units with shortcuts as the finest reconstruction granularity, then progressively combines them into coarser blocks and reconstructs at each granularity level, with a two-stage Progressive Optimization Strategy (POS). Experiments on ImageNet and COCO report accuracy improvements over prior PTQ methods, particularly at 3-bit and 4-bit quantization (e.g., 75.61% top-1 for 3-bit ViT-B).","tokens_in":17427,"tokens_out":6530,"duration_ms":44190,"significance":"If the reported results hold, the paper offers a simple and practical improvement for low-bit PTQ of ViTs, with the progressive fine-to-coarse idea being conceptually clear and easy to integrate into existing reconstruction frameworks. The paper includes extensive ablations, efficiency comparisons, loss distribution analysis, and qualitative visualizations. The main weakness is that the quantitative support for the central claim currently rests on an ablation baseline that appears far weaker than the published block-wise method it is compared against, and the abstract overstates the SOTA claim.","major_comments":[{"comment":"The block-wise baseline of 36.58% top-1 for 4-bit ViT-S is far below the 74.87% reported for the block-wise method I&S-ViT in Table 1 under the same model and bit-width. The paper states this baseline was reproduced 'under the same setting,' but with the default 300 iterations and 4e-5 learning rate it appears severely under-trained. Consequently, the claimed 37.45% improvement from PFCR over block-wise reconstruction is not a fair measure; against the I&S-ViT block-wise baseline, the improvement in Table 1 is only 1.44 points. This is load-bearing because the Introduction and Section 4.4 use this comparison to motivate the central claim that progressive fine-to-coarse reconstruction is superior to single-granularity block-wise reconstruction. The ablation should either use a properly tuned block-wise baseline or report the exact hyperparameter settings and explain why the reproduced baseline is so much weaker than the published number.","section":"Section 4.4, Table 3"},{"comment":"The abstract states that the method 'achieves the best Top-1 accuracy among state-of-the-art methods,' but Table 1 shows it is second-best for 6-bit DeiT-S (79.34 vs Outlier-Aware 79.50) and 6-bit DeiT-B (81.65 vs I&S-ViT 81.68). The body of the paper acknowledges this in Section 4.2, so the abstract overstates the result. The claim should be qualified to 'best in most settings, particularly at 3-bit and 4-bit' or similar.","section":"Abstract, Section 4.2, Table 1"},{"comment":"The recursive definition of R_g appears to contain an indexing error. For g>1 the formula reads R_g = R_{g-1}(R_{g-1}(X_{l-2^{g-1}+1})), but for a coarse unit spanning 2^g blocks ending at index l, the input should be X_{l-2^g+1}, not X_{l-2^{g-1}+1}. As written, the recurrence does not match the description of combining two adjacent R_{g-1} units and is not well-defined. Since Eq. (17) is the formal definition of the proposed reconstruction units, this needs to be corrected for reproducibility.","section":"Eq. (17), Section 3.2"},{"comment":"The main comparison tables report results from a single run without variance estimates. Given that calibration and reconstruction use randomly sampled 64 and 1024 images, and that several claimed improvements are as small as 0.07-0.3 points, reporting mean plus/minus standard deviation over multiple runs (as done for Figure 4) would be needed to substantiate the SOTA claims.","section":"Tables 1 and 2"}],"minor_comments":[{"comment":"The section title reads 'Progessive Fine-to-Course Reconstruction'; it should be 'Progressive Fine-to-Coarse Reconstruction.' There is also a typo in the same section: 'coaserest' should be 'coarsest.'","section":"Section 3.2 title"},{"comment":"The notation uses the rounding operator ⌊·⌉ inconsistently with the floor operator ⌊·⌋ used in Eq. (3). For a power-of-two test, floor and rounding happen to coincide for the given examples, but the notation should be unified for clarity.","section":"Eq. (15)"},{"comment":"The caption states that 'We test three times to compute the mean and std,' but the figure does not show error bars or report the numerical mean and standard deviation values. Either add error bars or remove the statement.","section":"Figure 4"},{"comment":"The sentence 'APQ-ViT achieves the best box average precision with the Swin-S backbone' is immediately followed by 'Nevertheless, it suffers significant performance degeneration with the Swin-S backbone,' which is contradictory. The second clause should presumably refer to the Swin-T backbone, for which APQ-ViT is indeed poor.","section":"Section 4.3, Table 2"},{"comment":"The asterisk marker '∗' is used for RepQ-ViT results but is not explained in the caption. Please add a note that these are reproduced with the official code.","section":"Table 2 caption"},{"comment":"The sentence 'boosting their classification performance by 38.67%. and 2.28% severally' has a misplaced period and should read 'by 38.67% and 2.28%, respectively.'","section":"Section 4.4.2"}],"recommendation":"major_revision","confidential_remarks":"The core idea is plausible and the 3/4-bit results against strong baselines are interesting, but the unfairness of the Table 3 baseline and the overstatement in the abstract need to be fixed before the paper can be considered for publication. I would also encourage the editor to ask the authors to release code, since the method's reproducibility would be improved."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this paper right away. The core idea is a progressive fine-to-coarse reconstruction for low-bit ViT PTQ: reconstruct attention and MLP modules first, then combine them into blocks, then into pairs of blocks, and so on. That is a sensible extension of existing granularity-based methods, and the authors also add a two-stage smooth-then-rugged optimization strategy (POS). The reported numbers are strong on their face, especially 3-bit ViT-B at 75.61% and 4-bit ViT-B at 83.19%. But the paper's central evidence that progressive reconstruction beats ordinary block-wise reconstruction is undercut by the ablation table. The \"Block-wise\" baseline in Table 3 is 36.58% for 4-bit ViT-S, while the same paper lists I&S-ViT, a block-wise method, at 74.87% for that setting. That is a ~38-point gap. Either the baseline is not representative or something else is off. Against the published I&S-ViT number, the gain is about 1.4 points, not 37 points.\n\nWhat the paper does well: the fine-to-coarse idea is clearly explained, the equations are straightforward, and the authors are honest about some limitations. They acknowledge the second-best results at 6-bit on DeiT-S and DeiT-B, and they discuss the added memory cost in Table 6. The COCO experiments are a reasonable bonus. I also appreciate the loss-distribution analysis in Figure 3, which nicely illustrates why fine-first reconstruction might control error accumulation.\n\nSoft spots, in order of size. First, the unfair baseline is a real problem. The 37.45% improvement claim in Section 4.4 is inflated. Second, the abstract's \"best Top-1 accuracy among state-of-the-art methods\" is contradicted by their own Table 1 in at least two 6-bit settings. Third, there are no error bars, and no code or data are provided, so we cannot check the reproducibility. These are fixable, but they matter.\n\nWho is this for? Researchers working on PTQ for vision transformers will find the progressive granularity idea worth examining. It's a plausible approach, and the reported numbers, if confirmed, would be useful. The paper deserves a serious referee, but not as-is. My recommendation: send it to peer review and require major revisions. Ask for a fair block-wise baseline (ideally using the published implementation), a more precise SOTA statement, and either code or a detailed reproducibility appendix.\n\nI'd hesitate to cite the specific numbers until I see the baseline issue resolved.","headline":"Plausible progressive reconstruction idea, but the headline results rest on a weak reproduced baseline and the SOTA claim overreaches.","tokens_in":17914,"tokens_out":2971,"would_cite":false,"duration_ms":20120,"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 progressive fine-to-coarse reconstruction schedule for low-bit post-training quantization of vision transformers reports 75.61% Top-1 accuracy on a 3-bit ViT-B and brings 4-bit ViT-B within 1.35% of full precision.","keywords":["Post-Training Quantization","Vision Transformers","Reconstruction Granularity","Image Classification","Object Detection","Instance Segmentation","Low-bit Quantization","Model Compression"],"falsifier":"Re-run the Table 3 ablation for 4-bit ViT-S under the paper's stated calibration data and hyperparameters, but replace the reproduced block-wise baseline with a published strong block-wise reconstruction such as I&S-ViT; if that baseline already achieves roughly 74.9% Top-1, then the 37.45% gain attributed to PFCR is an artifact of the weak baseline, and the per-block reconstruction-loss explosion reported for block-wise methods would not replicate.","tokens_in":16966,"feed_emoji":"🎯","tokens_out":11358,"duration_ms":81839,"temperature":0.7,"pith_summary":"The paper is trying to establish that low-bit post-training quantization of vision transformers can be made much more accurate by reconstructing the quantized model at progressively coarser granularities, rather than at a single fixed granularity. The method starts from the two finest reconstruction units — the multi-head self-attention module and the MLP, each with its shortcut — and after optimizing them, combines pairs to form blocks, then pairs of blocks, and so on, with each coarser stage inheriting the optimized weights of the finer stage. This progressive fine-to-coarse reconstruction is paired with a two-stage optimization schedule (POS) that first quantizes only activations and then both weights and activations, which the paper says eases optimization. On ImageNet, the paper reports 75.61% Top-1 accuracy for a 3-bit quantized ViT-B and 4-bit ViT-B within 1.35% of full precision. The same method also improves object detection and instance segmentation on COCO, indicating the benefit is not limited to classification.","feed_headline":"Progressive reconstruction lifts 3-bit ViT-B to 75.61 percent","feed_subtitle":"A fine-to-coarse schedule closes most of the accuracy gap for 3- and 4-bit vision transformers on ImageNet.","key_machinery":"The central object is the progressive fine-to-coarse reconstruction unit. The finest units are defined as $R^A_0(X_l)=X_l+\\mathrm{MHSA}(\\mathrm{LN}(X_l))$ and $R^P_0(Y_l)=Y_l+\\mathrm{MLP}(\\mathrm{LN}(Y_l))$, i.e., the attention and MLP modules with their shortcuts; these are optimized with mean-square-error losses, then combined into coarser units $R_g$ at granularity levels $g=1,\\dots,G$ by the recursive composition in Eq. (16)-(17). The second mechanism is the Progressive Optimization Strategy (POS), a two-stage schedule: stage one quantizes only activations, keeps weights full-precision, and uses $G=1$; stage two quantizes both weights and activations with the full $G$, while diminishing learning rates and increasing iterations with $g$ according to Eq. (18)-(19). The combination of progressive inheritance of weights from finer to coarser granularities and the two-stage quantization schedule is what the paper identifies as the source of the accuracy gains.","core_discovery":"The central claim is that the reconstruction granularity used in post-training quantization should be progressively increased rather than fixed in advance. The paper shows that when finer units (attention and MLP with shortcuts) are reconstructed first, they provide a better initialization for reconstructing the coarser blocks built from them, and the coarser reconstruction in turn refines the parameters of the finer units, so that iterating this process suppresses accumulated reconstruction error. The paper substantiates this claim with per-block reconstruction-loss curves that stay flat under PFCR but explode under block-wise reconstruction, and with ImageNet accuracy gains that grow as the coarsest granularity level G increases (e.g., 3-bit ViT-S improves from 19.23% to 49.83% as G goes from 0 to 3 with fixed 250-iteration training). The reported headline result is 75.61% Top-1 accuracy for 3-bit ViT-B, outperforming prior state-of-the-art methods on the same setting.","pith_inferences":["The same fine-to-coarse schedule could likely be applied to other transformer-based architectures, including large language models, if the memory cost of the coarser granularity levels is reduced; the conclusion itself notes that direction as future work.","A validation-set-driven choice of the coarsest granularity level $G$ could avoid the mild 6-bit overfitting while preserving the 3-4 bit gains, since the paper fixes $G$ by a power-of-two formula rather than by data.","The two-stage POS suggests that activation-only quantization could serve as a universal warm start for any weight-and-activation PTQ method, not only for PFCR; this could be tested by pairing POS with existing single-granularity reconstruction methods.","The reported accuracy gains at 3 bits indicate that even finer reconstruction units (e.g., LayerNorm outputs or the post-softmax attention) might push useful PTQ down to 2 bits, where current ViT quantization methods typically collapse."],"forward_implications":["If the reported results hold, 3-bit and 4-bit ViTs can be deployed with accuracy close to full precision, making transformer-based vision models practical on memory-constrained devices.","The method extends to object detection and instance segmentation, so the reconstruction technique is not tied to the image-classification objective.","Reconstruction time drops by roughly a factor of seven compared with Outlier-Aware (about 19 minutes vs 130 minutes for 4-bit ViT-S) while improving accuracy, so low-bit PTQ becomes more practical.","The ablation shows that increasing the number of granularity levels G is itself the main driver of accuracy in low-bit settings, which indicates that the progressive schedule, not just more training, is responsible for the gains.","At 6-bit quantization the progressive schedule can slightly hurt accuracy due to overfitting on the small calibration set, so the method's best operating range is at 3-4 bits."],"supporting_citations":[{"why":"Introduces block-wise reconstruction for PTQ, the single-granularity paradigm that PFCR extends by making granularity progressive.","marker":"Li et al. (2021)"},{"why":"I&S-ViT, the main state-of-the-art baseline PFCR claims to beat, and the source of the rugged-loss-landscape observation behind POS.","marker":"Zhong et al. (2023)"},{"why":"Outlier-Aware, which first defines reconstruction granularity and is the principal 4-bit competitor plus the efficiency/memory comparison of Table 6.","marker":"Ma et al. (2024)"},{"why":"PD-Quant, a block-wise reconstruction baseline with prediction-difference loss, compared in Tables 1 and 2.","marker":"Liu et al. (2023a)"},{"why":"RepQ-ViT, a scale-reparameterization PTQ method whose 4-bit ImageNet results serve as a comparison baseline.","marker":"Li et al. (2023)"}],"fun_headline_variants":["Fine-to-coarse PTQ hits 75.61% on 3-bit ViT-B","Progressive reconstruction boosts 3-bit ViT-B to 75.61%","3-bit ViT-B hits 75.61% with progressive PTQ","Progressive fine-to-coarse PTQ yields accurate 3-bit ViTs","Fine-to-coarse reconstruction: 3-bit ViT-B at 75.61%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the reproduced block-wise reconstruction baseline used in the ablation (36.58% Top-1 for 4-bit ViT-S) is a fair and representative implementation of single-granularity reconstruction; if a properly tuned block-wise method already reaches about 74.9% on the same setting, the claimed 37.45% improvement from progressive fine-to-coarse reconstruction is inflated and the central superiority claim is not established.","fun_headline_variants_meta":{"raw":{"variants":["Fine-to-coarse PTQ hits 75.61% on 3-bit ViT-B","Progressive reconstruction boosts 3-bit ViT-B to 75.61%","3-bit ViT-B hits 75.61% with progressive PTQ","Progressive fine-to-coarse PTQ yields accurate 3-bit ViTs","Fine-to-coarse reconstruction: 3-bit ViT-B at 75.61%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000652,"raw_usage":{"total_tokens":3046,"prompt_tokens":1057,"completion_tokens":1989,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":673,"completion_tokens_details":{"reasoning_tokens":1880}},"tokens_in":673,"tokens_out":1989,"duration_ms":28428,"temperature":1.0,"reasoning_tokens":1880,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T12:02:50.466914+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the Table 3 ablation for 4-bit ViT-S under the paper's stated calibration data and hyperparameters, but replace the reproduced block-wise baseline with a published strong block-wise reconstruction such as I&S-ViT; if that baseline already achieves roughly 74.9% Top-1, then the 37.45% gain attributed to PFCR is an artifact of the weak baseline, and the per-block reconstruction-loss explosion reported for block-wise methods would not replicate.","supporting_citations":[],"review_version":1}