{"id":"77a2ffc9-8818-40b4-beab-0964dba728a3","arxiv_id":"2507.17768","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Using a relative-entropy coreset and layer-wise feature alignment, QuaRC improves 2-bit quantization-aware training accuracy by up to 5.72 points on ImageNet-1K when training on 1% of the data.","lead":"QuaRC is a method for retraining quantized neural networks on a small fraction of a dataset, aimed at phones and drones. It picks the most informative images by measuring how much the low-precision model's predictions diverge from the full-precision model, then aligns internal features during training.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline ImageNet gain is not decomposed: CIFAR ablations show CLC, not RES, drives most of QuaRC's advantage, so the coreset-selection claim may be unsupported on the flagship result.","rationale":"The reader identifies the RES heuristic as the weakest assumption, and I agree the RES component is the fragile part of the argument. However, the sharper concern is not only the small 10-point correlation in Fig. 2(a); the paper's own CIFAR-100 ablations show that CLC produces the large majority of QuaRC's improvement over ACS. On the flagship ImageNet result, no decomposition is provided, so it is unknown whether RES contributes anything at scale. This does not invalidate the overall empirical claim that the full QuaRC pipeline outperforms ACS, but it does mean the paper's central novelty as stated in the title and contributions is not yet established on the headline benchmark. The appropriate remedy is a targeted ablation plus error bars, not rejection. I also note a smaller internal inconsistency: Fig. 4(b) reports 67.18% Top-1 at 20% data while Table VI reports 65.86% for the same condition; this reinforces the need for released code and reproducible seeds. The verdict should remain CONDITIONAL because the main tables are plausible and the CLC component appears effective, but the RES contribution requires verification before the paper's central claim can be accepted at face value.","tokens_in":20477,"tokens_out":7002,"duration_ms":85779,"concrete_test":"Run 2w2a ResNet-18 on ImageNet-1K with a 1% coreset, using the same ResNet-101 teacher and at least three seeds for each of three configs: (a) ACS selection + KD only, (b) ACS selection + CLC without RES, and (c) full QuaRC (ACS+RES+CLC). If config (b) approximately matches config (c) while both clearly exceed config (a), the reported 5.72% gain should be attributed to CLC rather than RES, and the RES claim on ImageNet is unsupported. If instead config (c) exceeds config (b) by more than the seed-level noise, RES is validated at scale.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section V-C, Table III shows that on 2w32a MobileNetV2 with a 1% CIFAR-100 coreset, baseline ACS is 46.84, +RES is 48.77, +CLC is 55.25, and +RES+CLC is 56.36. Thus CLC alone contributes +8.41 points, while RES contributes at most +1.93 points, or +1.11 points when added on top of CLC. The paper's headline result, a 5.72% Top-1 improvement over ACS on 2w2a ResNet-18 with a 1% ImageNet-1K subset, is reported only for the full QuaRC pipeline; no ablation separates the contribution of RES from that of CLC and the ResNet-101 teacher. Since the abstract and contributions frame RES as a central novelty for capturing quantization errors during coreset selection, the current evidence leaves open that the ImageNet gain is driven almost entirely by CLC plus teacher distillation rather than by the RES coreset. If RES degenerates to near-random or ACS selection at ImageNet scale, the paper's core contribution is not supported on its flagship benchmark.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes QuaRC, a two-phase framework for quantization-aware training (QAT) on edge devices with small coresets. In the selection phase, QuaRC introduces a Relative Entropy Score (RES) that ranks training samples by the KL divergence between the outputs of the full-precision and quantized models. In the training phase, it adds a Cascaded Layer Correction (CLC) loss that aligns intermediate-layer outputs of the quantized student with those of the full-precision teacher. Experiments on CIFAR-100 and ImageNet-1K with MobileNetV2 and ResNet-18 across 2/3/4-bit settings report consistent gains over a coreset-based baseline, with a headline result of 5.72% Top-1 improvement over ACS on a 1% ImageNet-1K subset for 2-bit ResNet-18. The paper also gives a complexity analysis, a convergence proof for SGD on a fixed coreset, an ablation study, and a UAV crack-detection case study.","tokens_in":20716,"tokens_out":6290,"duration_ms":73478,"significance":"If the reported gains hold, QuaRC is practically valuable: it targets a real bottleneck (QAT on edge devices with scarce data) and combines a quantization-error-aware selection criterion with a layer-wise correction loss. The CLC component is shown to be consistently beneficial across coreset selection methods (Table V), and the efficiency-accuracy analysis (Fig. 4) supports the practical motivation. The authors provide a clear ablation table on CIFAR-100 and a real-world case study. However, the central RES contribution is not isolated on the flagship ImageNet benchmark, the supporting correlation analysis is statistically weak, the convergence proof does not cover the actual algorithm, and all experimental results lack variance estimates, so the strength of the central claim exceeds what the current evidence supports.","major_comments":[{"comment":"The flagship ImageNet claim (5.72% Top-1 improvement over ACS at 1% subset, 2w2a ResNet-18) is never decomposed. In the CIFAR-100 ablation (Table III), CLC alone contributes +8.41 Top-1 over ACS, while RES alone contributes +1.93, and RES on top of CLC contributes +1.11. Because the ImageNet setup also uses a stronger ResNet-101 teacher, the reported 5.72% gain cannot be attributed to the relative-entropy coreset selection. Please provide an ImageNet ablation (ACS, +RES, +CLC, +RES+CLC) with the same teacher and training schedule, or explicitly report the RES-only and CLC-only contributions on this benchmark.","section":"§V-B and §V-C (Table II vs. Table III)"},{"comment":"The core premise of RES is validated with a Spearman correlation computed over only 10 coresets (rho=0.867, p=0.00117), all from one architecture (MobileNetV2), one bitwidth (2w32a), and one dataset fraction (1% of CIFAR-100), with no repeated runs or confidence intervals. With n=10, this correlation estimate is highly unstable, and the later ablations show that RES's incremental accuracy gain is small. To make the relative-entropy selection rule credible, report correlations over more coresets (e.g., 30 or more) and across at least one additional architecture/bitwidth, or provide bootstrap confidence intervals.","section":"§III-B, Fig. 2(a)"},{"comment":"The convergence analysis is for vanilla SGD on a fixed coreset DS with a Lipschitz-smooth, bounded-gradient loss; it does not model the actual QuaRC pipeline. In QuaRC, the coreset is re-selected every R epochs so DS changes over time, the training loss includes L_KD and beta*L_CLC, and gradients are computed through the straight-through estimator with quantization error. Consequently, Theorem 1 does not establish convergence of QuaRC, only a known property of SGD for a generic smooth loss. The authors should either state this limitation explicitly or provide an analysis that accounts for the changing coreset and the quantization/staleness error.","section":"§IV-B, Theorem 1"},{"comment":"All reported accuracies are single-run numbers with no error bars or standard deviations. Given the extremely small subset fractions (1%), run-to-run variation from coreset selection and training initialization can be substantial. For the headline 5.72% ImageNet improvement, without variance estimates it is not possible to assess whether the difference from ACS is significant. Please report mean and standard deviation over at least three seeds for the main comparisons, or justify why a single seed is sufficient.","section":"§V-A, Tables I-II"}],"minor_comments":[{"comment":"The text states that the dRES-only row achieves 55.42% Top-1, but the table reports 55.41%; please reconcile.","section":"§V-C, Table IV"},{"comment":"The efficiency inequality is stated as holding when S is sufficiently small, but the second term O(2*F*N*T/R) also depends on the selection interval R. For small R (e.g., R=1), the inequality may not hold. Please state the explicit condition on S and R, or include R in the comparison.","section":"§IV-A, Eq. (10)"},{"comment":"The notation in L_CLC is confusing: the summation index is c (layers) while the probabilities use subscript i (pi_Q, pi_F). Please use a consistent layer index and define pi_Q/pi_F as the intermediate outputs at layer c.","section":"§III-C, Eq. (8)"},{"comment":"The hyperparameters beta and R are set manually for each dataset (beta=1e5, R=50 for CIFAR-100; beta=3e3, R=10 for ImageNet) with no sensitivity analysis. Please include a brief study of how performance varies with beta and R, since the claimed robustness depends on these choices.","section":"§V-A and Algorithm 1"},{"comment":"The input list includes 'Initial coreset DS(t)' but t is the loop variable and DS(t) is defined during the loop; please rename this to DS(0) or clarify.","section":"Algorithm 1"}],"recommendation":"major_revision","confidential_remarks":"The paper does not provide code or checkpoints, which would help verify the single-run numbers. The CIFAR-100 ablations are the strongest evidence, but the headline ImageNet result needs a proper component ablation before the central RES contribution is credible. The convergence theorem is generic and could be removed or reframed as a baseline guarantee rather than a proof for QuaRC."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take for your inbox. The QuaRC paper adds two things to QAT-with-coresets: a relative-entropy score (RES) for picking samples and a cascaded layer correction (CLC) loss that aligns intermediate features with a teacher. The combination is new for this task family, and the empirical sweep is respectable: multiple bitwidths, fractions, datasets, plus a real UAV crack-detection case study. The authors also clearly state limitations, which I appreciate. Citations look appropriate; they lean on ACS, which is the right baseline.\n\nBut the central claim needs a caveat. On the CIFAR-100 ablation (Table III), CLC alone lifts ACS from 46.84 to 55.25 Top-1, while RES alone moves it to 48.77 and RES on top of CLC adds about one point. The headline ImageNet number (5.72% over ACS at 1% data) is only reported for the full pipeline, so we don't know whether RES contributes anything there. Given RES is framed as the paper's main novelty, that's a load-bearing gap. The stress-test note is right: the current evidence leaves open that the flagship result is almost entirely CLC plus a stronger teacher.\n\nOther soft spots: the RES heuristic rests on a Spearman correlation over 10 coresets with no error bars; the convergence proof is the standard SGD story and never touches coreset approximation or quantization error; and there is no code or multiple seeds. These are fixable, but they cap the paper's evidentiary weight right now.\n\nWhat's genuinely good: the problem is real, the CLC idea is simple and seems to help uniformly across coreset methods (Table V), and the UAV case study is a nice sanity check even if it's one run.\n\nFor a reader: if you work on QAT or data selection, worth a skim for CLC. I wouldn't cite it as evidence yet, but it deserves a serious referee. With a decomposed ImageNet ablation and multi-seed numbers, it could be a solid TMLR/ECCV-type contribution. My recommendation: engage with it, but ask for the missing ablation before endorsing.","headline":"The ImageNet headline is carried by CLC, not the RES coreset; a useful empirical paper that needs a decomposed ablation and error bars before its central claim is supported.","tokens_in":21212,"tokens_out":2829,"would_cite":false,"duration_ms":32464,"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":"QuaRC claims that relative-entropy coreset selection plus cascaded layer correction lets 1% data QAT beat state-of-the-art coreset methods.","keywords":["quantization-aware training","coreset selection","relative entropy","low-bit quantization","edge computing","knowledge distillation","layer-wise correction","data-efficient training"],"falsifier":"Run RES-guided coreset selection on a new architecture or dataset where the ranking of per-sample relative entropies is measured against final accuracy over many seeds; if the Spearman correlation is not consistently positive, or if coresets chosen by low relative entropy match or beat high-entropy coresets, the central claim fails.","tokens_in":20290,"feed_emoji":"⚡","tokens_out":4754,"duration_ms":48051,"temperature":0.7,"pith_summary":"QuaRC is a framework for making quantization-aware training (QAT) practical on edge devices when only a small fraction of the data can be used. It argues that the best coreset for retraining a quantized model is not the most diverse or hardest samples, but the samples whose output distributions diverge most between the full-precision and quantized models, measured by relative entropy. On top of this selection, it adds a Cascaded Layer Correction loss that aligns intermediate layer outputs of the quantized model with the full-precision model, reducing error accumulation that ordinary knowledge distillation misses on small data. If correct, low-bit quantized models can be retrained on-device with 10% or less of the data while approaching full-data accuracy.","feed_headline":"1% of data, 5.72% accuracy gain: QuaRC retrains quantized models","feed_subtitle":"Relative-entropy coreset selection plus layer-wise correction lets edge devices run QAT on tiny subsets.","key_machinery":"The two load-bearing mechanisms are the Relative Entropy Score (RES) and Cascaded Layer Correction (CLC). RES is the KL divergence $d_{\\mathrm{RES}} = \\sum_m p_Q^{(m)}(w_q,x_q)\\log\\left(p_Q^{(m)}(w_q,x_q)/p_F^{(m)}(w_r,x_r)\\right)$, computed per sample between quantized and full-precision output logits, used to rank and select the coreset. CLC adds the loss $\\mathcal{L}_{\\mathrm{CLC}} = \\sum_c p^i_Q(w_q,x_q)\\log\\left(p^i_Q(w_q,x_q)/p^i_F(w_r,x_r)\\right)$, aligning chosen intermediate-layer distributions of the quantized model to the full-precision model, with total loss $\\mathcal{L}_{\\mathrm{TOTAL}} = \\mathcal{L}_{\\mathrm{KD}} + \\beta \\mathcal{L}_{\\mathrm{CLC}}$.","core_discovery":"The paper's central claim is that quantization error, not generic data difficulty, should drive coreset selection for QAT. It defines the Relative Entropy Score (RES) as the KL divergence between the output probability distributions of the full-precision and quantized models for each sample, and selects the top S% of samples by this score combined with gradient-based scores. It reports that this selection alone improves 2-bit MobileNetV2 on 1% CIFAR-100 from 46.84% to 48.77% Top-1 accuracy over the ACS baseline, and that the Cascaded Layer Correction strategy, which minimizes KL divergence between intermediate-layer outputs, contributes the larger gain (to 55.25%), with the full QuaRC reaching 56.36%. On ImageNet-1K with 2-bit ResNet-18 and a 1% subset, QuaRC achieves 46.34% Top-1, a 5.72% absolute improvement over ACS. The paper also reports a Spearman correlation of 0.867 between the mean relative entropy of coresets and the resulting Top-1 accuracy as evidence for the selection principle.","pith_inferences":["If the relative-entropy ranking is a stable property of the quantization process, RES could be computed once per calibration set and reused across bitwidths, avoiding repeated full-dataset forward passes; the paper does not test this.","The CLC idea generalizes beyond classifiers to any hierarchical network, including Transformers, but the paper only evaluates convolutional classifiers, so that extension is speculative.","A testable extension is to compare RES against uncertainty-based or loss-based selection to see whether relative entropy captures quantization-specific signal or just general sample difficulty.","Because RES depends on the current quantized model, its benefit may shrink as the model improves during training; periodic reselection every R epochs is assumed but the selection interval is not ablated."],"forward_implications":["With 10% or less of the training data, QuaRC can cut QAT training time while keeping accuracy close to full-data training; on CIFAR-100, the gain over ACS grows as the subset shrinks.","The Cascaded Layer Correction loss improves accuracy for every coreset selection method tested, including random sampling, so intermediate-layer alignment addresses a source of error that is independent of which samples are chosen.","On the UAV crack-detection case study, coreset QAT finishes in 2.55 minutes versus 7.02x longer for full-data QAT, while losing less than 0.5% Top-1 accuracy, making on-device retraining practical.","RES and CLC are complementary: together they outperform either alone at every bitwidth tested (2/3/4-bit)."],"supporting_citations":[{"why":"Supplies the adaptive gradient-based coreset baseline (ACS) that QuaRC is compared against and whose loss and selection schedule QuaRC extends.","marker":"[16]"},{"why":"Provides the LSQ+ quantization scheme used for all experiments, defining the quantization and dequantization operations in the fake quantization nodes.","marker":"[34]"},{"why":"Provides the CIFAR-100 dataset used for the primary MobileNetV2 experiments and the ablation studies.","marker":"[30]"},{"why":"Provides the ImageNet-1K dataset used for the ResNet-18 experiments, where the headline 5.72% improvement is reported.","marker":"[31]"},{"why":"Defines the MobileNetV2 architecture used as the student model on CIFAR-100 and in the UAV case study.","marker":"[5]"},{"why":"Defines the ResNet architecture used as the student model on ImageNet-1K and as the teacher in the teacher-model comparison.","marker":"[33]"},{"why":"Provides the Forgetting coreset baseline, which QuaRC claims performs worse than random sampling under QAT.","marker":"[27]"},{"why":"Provides the Moderate geometry-based coreset baseline, which QuaRC claims underperforms under low-bit quantization.","marker":"[25]"},{"why":"Provides the Contextual Diversity baseline, a clustering-based selection method that QuaRC compares against.","marker":"[26]"}],"fun_headline_variants":["QuaRC: 1% data, 5.72% accuracy boost for edge QAT","Relative entropy picks coresets for edge QAT, +5.72% Top-1","Edge QAT on tiny subsets: QuaRC's two-step boost","QuaRC: smarter coresets and layer correction for edge QAT","5.72% accuracy gain with 1% data: QuaRC for edge"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The selection rule assumes that samples whose outputs differ most between the quantized and full-precision models are the best ones to train on, an assumption supported mainly by a single Spearman correlation of 0.867 computed over 10 coresets.","fun_headline_variants_meta":{"raw":{"variants":["QuaRC: 1% data, 5.72% accuracy boost for edge QAT","Relative entropy picks coresets for edge QAT, +5.72% Top-1","Edge QAT on tiny subsets: QuaRC's two-step boost","QuaRC: smarter coresets and layer correction for edge QAT","5.72% accuracy gain with 1% data: QuaRC for edge"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000743,"raw_usage":{"total_tokens":3391,"prompt_tokens":1095,"completion_tokens":2296,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":711,"completion_tokens_details":{"reasoning_tokens":2187}},"tokens_in":711,"tokens_out":2296,"duration_ms":18944,"temperature":1.0,"reasoning_tokens":2187,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T16:40:05.483664+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run RES-guided coreset selection on a new architecture or dataset where the ranking of per-sample relative entropies is measured against final accuracy over many seeds; if the Spearman correlation is not consistently positive, or if coresets chosen by low relative entropy match or beat high-entropy coresets, the central claim fails.","supporting_citations":[{"cited_title":"Robust and efficient quantization-aware training via coreset selection,","cited_arxiv_id":null,"evidence_quote":"Supplies the adaptive gradient-based coreset baseline (ACS) that QuaRC is compared against and whose loss and selection schedule QuaRC extends."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Forgetting coreset baseline, which QuaRC claims performs worse than random sampling under QAT."},{"cited_title":"Moderate coreset: A universal method of data selection for real-world data- efficient deep learning,","cited_arxiv_id":null,"evidence_quote":"Provides the Moderate geometry-based coreset baseline, which QuaRC claims underperforms under low-bit quantization."},{"cited_title":"Contextual diversity for active learning,","cited_arxiv_id":null,"evidence_quote":"Provides the Contextual Diversity baseline, a clustering-based selection method that QuaRC compares against."}],"review_version":1}