{"id":"24a3ba79-9e65-461e-aed1-4ea317949d73","arxiv_id":"2607.11359","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Pre-tuning a full-precision model with quantization-error noise and weight averaging before PTQ consistently improves low-bit (2-4 bit) quantized accuracy across CNN architectures and segmentation.","lead":"This paper introduces ETBQ, a fine-tuning stage that adds simulated weight- and activation-quantization noise to a full-precision model before it is compressed, then averages late-training weights to settle in a flatter loss basin. The pre-conditioned model loses less accuracy under 2- and 4-bit post-training quantization, improving results on image classification and semantic segmentation benchmarks.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The mechanism claim—quantization-aligned perturbations, not generic fine-tuning/noise, drive ETBQ gains—is untested: no isotropic-noise or plain-fine-tuning control isolates the quantization-aligned component.","rationale":"The reader's verdict is CONDITIONAL based on the missing plain-fine-tuning control and concerns about the Gaussian surrogate. I agree that a control is missing, but I identify the more precise experimental hole: there is no comparison between quantization-aligned perturbations and generic isotropic perturbations of matched magnitude. The Gaussian-surrogate fidelity is not the central risk—even a crude Gaussian approximation could preserve the alignment that matters, and in fact the paper's Appendix A shows the Gaussian fit is quite good at 4-bit. The risk is that the entire benefit of WQN/AQN could be reproduced by any form of noise-conditioned fine-tuning (or even by plain fine-tuning plus SWA), which would invalidate the causal claim that conditioning on the target quantizer's error distribution is what makes the FP model quantization-tolerant. The paper's component ablations actually hint at this: SWA alone accounts for almost half the total gain, and WQN+AQN without SWA underperforms WQN alone, suggesting that the perturbation statistics are not straightforwardly responsible for the final result. Because this concern is concrete and addressable with a single additional ablation, it does not overturn the current CONDITIONAL verdict; it sharpens the condition under which the central claim would be accepted. I therefore leave the verdict unchanged but emphasize the need for the isotropic-noise control in the revision.","tokens_in":18505,"tokens_out":4408,"duration_ms":40823,"concrete_test":"Run ETBQ on ResNet-18/CIFAR-100 W2A4 under two additional controls: (1) plain fine-tuning for 120 epochs with the same SGD schedule and SWA but no perturbations (already approximated by the SWA-only row, but with explicit no-SWA variant); (2) the same ETBQ pipeline but with WQN perturbations replaced by channel-wise zero-mean isotropic Gaussian noise whose per-channel variance equals the empirical weight-quantization-error variance, and AQN perturbations replaced by per-tensor zero-mean isotropic Gaussian noise with the same variance and the same salience-aware mask. Compare W2A4 accuracies. If the isotropic-noise control achieves 78.06 ± 0.3 pp, the quantization-aligned mechanism is not supported; if it falls below ~77.3 pp, alignment is load-bearing.","verdict_should_be":"UNCHANGED","load_bearing_attack":"ETBQ's central claim is that pre-conditioning with perturbations sampled from the target quantizer's error distribution (WQN/AQN) is what improves downstream PTQ, rather than merely additional fine-tuning, SWA, or generic noise regularization. The paper does not provide the decisive control: fine-tuning with the same schedule, SWA, and isotropic Gaussian noise of matched magnitude. Table II includes 'SWA only' (which does improve over baseline by 0.75 pp) but this is a no-noise fine-tuning control, not a generic-noise control. Appendix C-A compares 'naive additive' vs 'differential' injection, but both use quantization-aligned errors; there is no comparison to an unaligned Gaussian perturbation with the same variance. Consequently, the observed gains—e.g., full ETBQ improving W2A4 from 76.47 to 78.06 on CIFAR-100—could be explained by any noise-injection regularizer or even by the extra 120 epochs of fine-tuning plus SWA. The Gaussian-surrogate concern raised by the reader is secondary: if alignment is what matters, a moment-matched Gaussian surrogate may suffice; but if generic noise reproduces the gains, the mechanism claim collapses regardless of Gaussian fidelity. The paper's own ablation shows WQN+AQN without SWA (77.43) is worse than WQN alone (77.67), indicating that the perturbation schedule is not a simple additive benefit. Thus the load-bearing uncertainty is not whether the Gaussian approximation is exact, but whether the quantizer-aligned structure of the perturbations is causally necessary for the reported improvements.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"ETBQ proposes an upstream full-precision (FP) pre-conditioning stage for SGD-trained models before low-bit post-training quantization (PTQ). During tuning, it injects Weight Quantization Noise (WQN), modeled as channel-wise Gaussians with temporal differencing (Eqs. 5–9), and Activation Quantization Noise (AQN), modeled as per-tensor Gaussians with salience-aware stochastic masking (Eqs. 10–15), followed by Stochastic Weight Averaging (SWA). The resulting FP model is then fed to standard PTQ backends such as QDrop and QEP. Experiments on CIFAR-100, Tiny-ImageNet, ImageNet, and Cityscapes report consistent improvements, e.g., W2A4 top-1 on Tiny-ImageNet ResNet-18 rises from 63.71 to 65.85, and Cityscapes U-Net mIoU from 57.72 to 63.52. The paper argues that ETBQ moves the FP model to a flatter loss basin, making it more tolerant to quantization errors, and contrasts the method with calibration-only PTQ and full QAT.","tokens_in":18905,"tokens_out":4073,"duration_ms":39540,"significance":"If the claims hold, ETBQ is a simple, backend-agnostic way to improve low-bit PTQ by improving the FP starting point, with compute positioned between calibration-only PTQ and full QAT (Table VII). The paper's strengths are its broad empirical coverage (five architectures, four datasets, two PTQ backends), released code, and honest reporting of component ablations and limitations, including the explicit statement that the flatness evidence is correlational and the acknowledgment that the Gaussian surrogate degrades at W2A2. The core empirical result — that the integrated ETBQ pipeline improves downstream PTQ across diverse settings — appears internally consistent. The main gap is that the mechanism claim (quantization-aligned perturbations, not generic fine-tuning/noise, drive the gains) is not isolated experimentally: there is no plain-fine-tuning control and no isotropic-noise control. The central empirical contribution is credible, but the attribution of the effect to quantization-aligned structure is under-tested.","major_comments":[{"comment":"The ablation table has no plain-fine-tuning control. Table II includes WQN-only, AQN-only, SWA-only, and combinations, but no row with the same 120-epoch SGD schedule, cosine annealing, label smoothing, and SWA but without any WQN/AQN noise injection. Since the full ETBQ gain over direct QDrop is 1.59 pp (78.06 vs 76.47), and SWA alone already gives 77.22, the incremental contribution of quantization-aligned perturbations could be much smaller than the contribution of extra training plus SWA. Add a control row: same schedule, same optimizer, same SWA, no noise injection. Without it, the paper's central claim that pre-conditioning with quantization-error samples is what improves PTQ is not established.","section":"Section V-B, Table II"},{"comment":"The mechanism claim distinguishes WQN from 'generic random perturbation' and 'generic anti-correlated injection,' but the experiments do not include an isotropic-noise control. Table IX compares naive additive vs differential injection, but both use quantization-aligned errors. A decisive control is to inject isotropic Gaussian noise with the same per-channel variance and the same temporal differencing, but with directions not aligned to the target quantizer's error distribution. If this control reproduces the W2A4 gains, the quantization-aligned component of WQN is not necessary. This is load-bearing because the paper's stated contribution is alignment with the target quantizer's error distribution, not merely noise-induced flatness.","section":"Section IV-A, Appendix C-A"},{"comment":"The hyperparameter rho (target AQN mask density) is never reported, and lambda_max is tuned per dataset without sensitivity analysis. Table VIII lists lambda_max = 1.0, 0.8, 0.05, 0.8 for CIFAR-100, Tiny-ImageNet, ImageNet, and Cityscapes, but no ablation shows how sensitive results are to this choice. Since Eq. (13) uses rho to control how many channels receive activation perturbation, omitting its value makes the method hard to reproduce and leaves open the possibility that results depend on finely tuned noise intensity. Report rho for every dataset and provide a small sensitivity study (e.g., lambda_max in {0.2, 0.5, 1.0} for ResNet-18/CIFAR-100 W2A4).","section":"Table VIII, Eq. (13)"},{"comment":"The paper acknowledges that the Gaussian surrogate becomes inaccurate under 2-bit quantization (NRMSE_QQ increases to about 0.10) and that W2A2 gains are smaller. This is not fatal, but it weakens the statement in Section IV-A that WQN/AQN model 'the genuine quantization error' rather than a generic perturbation. In the very regime emphasized in the title (2-bit), the moment-matched Gaussian is a loose approximation. Either soften the modeling claim, or add a comparison where the perturbation is sampled directly from the empirical error distribution on the calibration set for W2A2, to test whether the Gaussian approximation, rather than the alignment principle, limits the method.","section":"Section VI, Appendix A-A"}],"minor_comments":[{"comment":"Typo: 'WQA+AQN+SWA' should be 'WQN+AQN+SWA'.","section":"Section V-B, Table II text"},{"comment":"The paragraph beginning 'Transferred activation quantization error overwhelms the weight quantization one.' is duplicated verbatim within the same section; remove the repetition.","section":"Section IV-A"},{"comment":"The initial learning rate entry for Tiny-ImageNet reads '0.0012×10^{-5}', which appears to be a formatting corruption. Please clarify the intended value.","section":"Table VIII"},{"comment":"The Hessian spectral density comparison is used as motivational evidence for flatness, but the text does not specify how the Hessian is estimated (e.g., full Hessian, Hutchinson trace estimate, top-k power iteration). Since the flatness claim is central to the narrative, add one or two sentences on the estimation procedure (in the main text or appendix).","section":"Fig. 1, Section I"},{"comment":"The algorithm says 'Refresh WQN and AQN statistics on D_c' every epoch, but Section IV-B says AQN statistics are updated via EMA with beta=0.9. Clarify whether the refresh is a full recomputation or an EMA update, and at what frequency (per epoch or per iteration).","section":"Algorithm 1, line 4"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope and the integrated pipeline results are likely to be useful to the community. The missing plain-fine-tuning and isotropic-noise controls are not fatal to the empirical claim, but they are load-bearing for the mechanism claim and should be addressed before publication. I would like to see the authors either add the controls or explicitly reframe the contribution as 'the full ETBQ pipeline including extra tuning and SWA improves PTQ' rather than claiming the quantization-aligned perturbation structure is necessary. No citation or novelty concerns from my side."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the empirical result is real, the mechanism story is not proven. ETBQ—an FP-only pre-conditioning stage with quantizer-aligned noise plus SWA, applied before QDrop/QEP—consistently improves low-bit PTQ across CIFAR-100, Tiny-ImageNet, ImageNet, and Cityscapes, and the gains on lightweight models are large enough to matter. The paper is also honest: it reports drops relative to each method's own FP32, acknowledges the Gaussian surrogate degrades at W2A2, and labels the Hessian evidence correlational.\n\nWhat's genuinely new is the positioning: a backend-agnostic, pre-PTQ conditioning stage that treats quantization tolerance of the FP model as an upstream problem. The component ablations are informative, and the appendix separates the temporal-differencing and salience-mask design choices. That is real work, not a repackaging.\n\nWhere it gets soft is the causal claim. There is no plain-fine-tuning control in Table II—'SWA only' is not 'fine-tune only'—and there is no control with isotropic Gaussian noise of matched magnitude. The comparison in Table IX is between two quantizer-aligned injection schemes, so it doesn't tell you whether alignment is what matters. A generic noise regularizer or simply extra fine-tuning could reproduce much of the gain; the SWA-only row already buys 0.75pp. Second, every result is a single run. No seeds, no error bars; the +8.53pp on MobileNet-V2 W2A4 is the kind of number that makes me want variance before I quote it. Third, rho is never given a value, and lambda_max is tuned per dataset with no sensitivity analysis. Fourth, the code is linked but without a commit hash; for a methods paper that's an easy fix. The W2A2 weakness is acknowledged and consistent with the limitations, so I don't hold that against them.\n\nNone of this sinks the paper. The integrated pipeline is a useful practical contribution that survives even if the mechanism story is overstated. A serious referee could reasonably ask for the missing controls and error bars and return a solid revision.\n\nRecommendation: send it to review. It's the kind of paper that needs a knowledgeable referee rather than a desk reject.","headline":"A serious, honestly written empirical paper that shows an upstream fine-tuning stage can improve low-bit PTQ; the quantizer-alignment mechanism is plausible but untested—worth peer review with requested controls.","tokens_in":19394,"tokens_out":3075,"would_cite":true,"duration_ms":29142,"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":"Low-bit post-training quantization accuracy is capped by the full-precision model's tolerance to quantization error, and ETBQ shows that tolerance can be improved beforehand, yielding 2.14-point top-1 gains on Tiny-ImageNet and 5.80-point m","keywords":["post-training quantization","low-bit quantization","quantization-aware tuning","quantization error tolerance","loss landscape flatness","stochastic weight averaging","weight quantization noise","activation quantization noise"],"falsifier":"Run an otherwise identical full-precision fine-tuning control that uses the same epochs, data, and compute but omits WQN and AQN perturbations; if that control matches ETBQ's W2A4 accuracy, the perturbation mechanism is not the cause. Additionally, at W2A2, compare the actual per-layer weight and activation quantization errors against the Gaussian model; large tail deviation would predict the small gains observed.","tokens_in":18412,"feed_emoji":"⚙️","tokens_out":3899,"duration_ms":36573,"temperature":0.7,"pith_summary":"The paper argues that low-bit post-training quantization (PTQ) accuracy is limited not only by how well a PTQ backend reconstructs the full-precision model, but also by how sensitive the full-precision model itself is to quantization errors. To address this, it introduces ETBQ, a full-precision tuning stage placed before PTQ: the model is optimized under weight and activation quantization-error perturbations sampled from statistics estimated on a calibration set, then averaged with stochastic weight averaging. The resulting full-precision model is used by an unchanged reconstruction-based PTQ backend, and the paper reports consistent gains across CIFAR-100, Tiny-ImageNet, ImageNet, and Cityscapes, with the largest improvements at W2A4. If the claim holds, it means quantization tolerance can be built into the full-precision model upstream rather than repaired only after quantization.","feed_headline":"Pre-quantization tuning lifts low-bit accuracy by up to 8.5 points","feed_subtitle":"A full-precision pre-tuning stage makes SGD-trained models more tolerant to 2-bit and 4-bit post-training quantization.","key_machinery":"The core mechanism is calibration-driven perturbation conditioning. WQN simulates the target channel-wise weight quantizer, estimates per-channel Gaussian error statistics, and forms a differenced perturbation P_t = δ_t − δ_{t−1} so that non-zero-mean quantization error does not accumulate in the FP weights; the task loss is evaluated at the perturbed weights while the stored weights remain clean. AQN estimates per-tensor Gaussian statistics for activation quantization error and samples an error tensor, then applies a salience-aware Bernoulli mask so that only channels with larger spatial absolute responses receive the perturbation. SWA then averages the late-training weights, and the paper","core_discovery":"The central claim is that a full-precision model can be pre-conditioned, before any PTQ backend runs, to be more tolerant of low-bit quantization error. ETBQ does this by injecting two calibration-estimated perturbation types during an SGD fine-tuning stage: Weight Quantization Noise, which samples from the channel-wise weight-quantization error distribution and uses temporal differencing to control drift, and Activation Quantization Noise, which samples from per-tensor activation-quantization error and applies a salience-aware stochastic mask. Stochastic weight averaging is then used to stabilize the trajectory, and BatchNorm statistics are recomputed. The output is a full-precision model t","pith_inferences":["Editorial inference: If flatness of the loss landscape is the operative cause, a sharpness-penalizing fine-tuning scheme that does not require calibration-set error statistics should reproduce much of ETBQ's gain; this is directly testable.","Editorial inference: The Gaussian surrogate for quantization error is the method's fragile link, and the small W2A2 gains give a natural place to test whether heavy-tailed or structured quantizer error violates the model's assumptions.","Editorial inference: Because ETBQ also improves FP32 accuracy slightly, part of the gain may be regularization rather than quantization-specific conditioning; a control that fine-tunes with the same compute but without WQN/AQN would separate these contributions."],"forward_implications":["If the claim is correct, PTQ users can improve low-bit accuracy without modifying the quantizer or training a fake-quantized model, simply by pre-tuning the full-precision model.","The benefit is strongest at W2A4, where quantization error is severe but the model has not collapsed; gains at W2A2 are smaller, suggesting a robustness boundary at extremely low bit-widths.","ETBQ sits between calibration-only PTQ and full quantization-aware training in cost, using a full-data fine-tuning stage of about 80 epochs on ImageNet while leaving quantizer optimization to the downstream PTQ backend.","The preconditioned model is backend-agnostic: the paper reports gains with two reconstruction-based PTQ backends across classification and semantic segmentation, indicating the improvement is not tied to one calibration method."],"fun_headline_variants":["ETBQ: pre-tune to rescue 2-bit and 4-bit PTQ","Pre-PTQ tuning lifts low-bit accuracy, no QAT","Tune before quantize: robust low-bit models","Pre-conditioning boosts post-training quantization","FP pre-tuning improves 2-bit and 4-bit PTQ"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The central claim depends on the assumption that Gaussian noise sampled from calibration-set error statistics faithfully represents the errors the downstream quantizer will actually introduce at low bit-widths, and that this specific conditioning—not ordinary fine-tuning—is what improves downstream PTQ accuracy.","fun_headline_variants_meta":{"raw":{"variants":["ETBQ: pre-tune to rescue 2-bit and 4-bit PTQ","Pre-PTQ tuning lifts low-bit accuracy, no QAT","Tune before quantize: robust low-bit models","Pre-conditioning boosts post-training quantization","FP pre-tuning improves 2-bit and 4-bit PTQ"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000269,"raw_usage":{"total_tokens":1507,"prompt_tokens":842,"completion_tokens":665,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":586,"completion_tokens_details":{"reasoning_tokens":578}},"tokens_in":586,"tokens_out":665,"duration_ms":6389,"temperature":1.0,"reasoning_tokens":578,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T06:54:48.589655+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run an otherwise identical full-precision fine-tuning control that uses the same epochs, data, and compute but omits WQN and AQN perturbations; if that control matches ETBQ's W2A4 accuracy, the perturbation mechanism is not the cause. Additionally, at W2A2, compare the actual per-layer weight and activation quantization errors against the Gaussian model; large tail deviation would predict the small gains observed.","supporting_citations":[],"review_version":2}