{"id":"57eccf28-c84c-48e2-9693-a220df0790e5","arxiv_id":"2608.03919","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Updating only normalization affine parameters under the target fake-quantization graph recovers severely degraded low-bit quantized models while training less than 1.43% of the parameters.","lead":"The paper shows that retuning only the tiny set of scale and shift values inside normalization layers can rescue low-bit quantized neural networks that usually collapse to near-random accuracy. The result matters because it offers a cheap, targeted alternative to retraining all parameters, though it currently needs the full training set to work on ImageNet.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theoretical cancellation proof (Eqs. 13–16) assumes per-channel affine error, but the implemented activation quantizer is per-tensor; the central explanatory claim is unsupported.","rationale":"The reader's weakest assumption correctly identifies the mismatch between the per-channel affine model in Eq. (13) and the per-tensor activation quantizer used in experiments. This is the most load-bearing concern because it directly invalidates a stated mathematical guarantee that anchors the paper's conceptual contribution. The empirical results—NAP recovering W4A4 to 66.11%—may still be true, but the paper's central explanation of *why* the affine subspace is high-leverage is unsupported. Without a valid mechanism, the work risks being interpreted as a specialized fine-tuning trick rather than a principled subspace-optimization principle. The paper is otherwise transparent about limitations (data dependence, backend sensitivity, non-universal gains), which supports a CONDITIONAL rather than REJECT verdict: the empirical claims deserve scrutiny but are not fatally undercut by this concern. A direct test of the theory's applicability would settle whether the explanation is salvageable.","tokens_in":12446,"tokens_out":5298,"duration_ms":57356,"concrete_test":"Run the NAP procedure on ImageNet MobileNetV2 W4A4 with a per-channel activation quantizer instead of the per-tensor quantizer actually used. If the theory is correct, the per-channel version should achieve accuracy at least as high as the per-tensor one, and the affine residual e_c = Q(a_c) − (α_c a_c + δ_c) should be small. Independently, on the real per-tensor graph, fit the affine model (13) to calibration data and compute the explained variance per channel; if the explained variance is low (e.g., <80%), the cancellation guarantee is not operative and the theory must be rejected or replaced.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's core explanatory claim is that BN affine parameters 'fully cancel the channel-wise affine component of quantization distortion' (Eq. 16). This relies on decomposition (13), which models the quantized pre-BN activation as Q(a_c) = α_c a_c + δ_c + e_c with channel-wise constants α_c and δ_c. However, the experimental protocol explicitly states that activations are quantized 'asymmetrically and per-tensor' (Experimental Protocol). For a per-tensor quantizer, a single scale and zero-point are shared across all channels, so the quantization error of channel c cannot be represented as α_c a_c + δ_c with independent channel-wise constants: the affine distortion is global, not channel-wise. Therefore, the derivation of the optimal BN parameters in Eqs. (14)–(16) does not apply to the actual graph used in the experiments. The method may still work through a different mechanism—e.g., BN gains/shifts rescale channels to reduce the dynamic range presented to the per-tensor quantizer—but the paper's promised 'theoretical analysis confirms...' is invalid as stated. This is load-bearing because it removes the principled justification for the 'response-control subspace' novelty, leaving the impressive empirical recoveries without a validated explanation.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Normalization Affine Preconditioning (NAP), which freezes backbone weights and tunes only normalization affine parameters (BN γ/β or RMSNorm γ) under the target fake-quantization graph. The central claim is that these parameters form a low-dimensional, high-leverage response-control subspace that can recover collapsed low-bit accuracy. Experiments on MobileNetV2/ImageNet report W4A4 accuracy rising from 0.33% (RTN) to 66.11% with NAP alone, outperforming QDrop; further experiments on CIFAR-100 QAT checkpoints, Cityscapes U-Net, and Qwen2.5-3B language modeling show improvements. The paper also provides a theoretical analysis (Eqs. 13–16) claiming that BN affine parameters cancel the channel-wise affine component of quantization distortion, with nonlinear residuals forming an irreducible boundary. Ablations examine parameter-subset comparisons, knowledge distillation, backend alignment, and data dependence.","tokens_in":12702,"tokens_out":5912,"duration_ms":62391,"significance":"If the empirical results hold, the observation that a tiny parameter subset (≈34K parameters) can rescue extremely low-bit models is practically valuable and conceptually interesting, especially because NAP adds no inference branch and can precede or follow existing PTQ/QAT methods. The paper's strengths include controlled equal-size parameter-subset comparisons (Table 5), a matched-versus-mismatched backend ablation (Table 6), honest reporting of failure cases (e.g., W2A4 OOQ-Freeze degradation in Table 2, small-data collapse in Table 7), and consistent results across classification, segmentation, and language modeling. However, the paper does not ship code, does not report error bars or multiple seeds, and its main theoretical justification does not match the implemented quantizer granularity. The empirical evidence is suggestive but currently incomplete; the theoretical explanation is not established.","major_comments":[{"comment":"The cancellation derivation assumes a per-channel affine decomposition Q(a_c)=α_c a_c+δ_c+e_c with channel-wise constants α_c, δ_c. However, the Experimental Protocol states that activations are quantized asymmetrically and per-tensor. A per-tensor quantizer uses one scale and zero-point shared by all channels, so the quantization error of channel c cannot be written with independent channel-wise affine constants. Consequently, Eqs. (14)–(16) do not apply to the implemented graph, and the claim that BN affine parameters 'fully cancel the channel-wise affine component of quantization distortion' is unsupported as stated. The 'irreducible error boundary' is likewise a restatement of the assumption that e_c lies outside the affine subspace, not an independently derived bound. This is load-bearing because it is the paper's stated theoretical confirmation of the method's principle. Please eit","section":"Method, 'Why Is the Affine Subspace High-Leverage?' Eqs. (13)–(16); Experimental Protocol"},{"comment":"Algorithm 1 fixes the quantization parameters Ω0 after calibration, but the protocol says 'observers are recalibrated after each tuning epoch.' These statements conflict. If quantization scales are updated during NAP, the comparison is not purely affine-parameter adaptation; quantizer-grid recalibration may by itself contribute to the observed gains. This confounds the attribution of improvements to the normalization affine subspace. Please clarify which procedure was actually used. If recalibration occurs, report an ablation with fixed Ω0 and another with recalibrated Ω, and state how observers are updated relative to the frozen BN running statistics.","section":"Algorithm 1 vs 'CNN optimization' in Experimental Protocol"},{"comment":"The headline ImageNet comparison is not matched in data budget. QDrop uses 20K reconstruction iterations on a calibration set, while NAP is tuned for 20 epochs on the full ImageNet training set. Table 7 shows that NAP with 10K tuning images reaches only 3.88% W4A4, versus 66.11% with full data. Thus the comparison to QDrop conflates method with data access, and the paper's own ablation shows the current ImageNet procedure is not a calibration-only or few-shot PTQ method. Please report results under equal data budgets—for example, NAP with QDrop's calibration data, and QDrop with full-data tuning—or reposition NAP explicitly as a full-training-data post-processing/pre-conditioning method.","section":"Table 1 vs Table 7 and Experimental Protocol"},{"comment":"Every table entry is obtained from a single run with random seed 42. Given randomness in initialization, data order, stochastic QAT, and reconstruction procedures, the main numerical claims (e.g., 66.11 vs 60.71 in Table 1) cannot be assessed for statistical significance. The paper states that code will be released upon publication, but for a paper whose evidence is primarily empirical, the absence of code or repeated-seed results is a major limitation. Please provide mean ± std over at least three seeds for the key tables, or release code as part of the revision.","section":"Experimental Protocol: Hardware, software, and runs"}],"minor_comments":[{"comment":"The provided text has many missing spaces (e.g., 'networks,rooted', 'thea', 'Wealso') suggesting source corruption or broken line-breaking. Please fix the formatting in the camera-ready version.","section":"Throughout"},{"comment":"Norm Tweaking is identified as the closest prior work for RMSNorm-based LLM quantization, but it is not compared in Table 4. A direct comparison with Norm Tweaking under the same protocol would strengthen the positioning and clarify the claimed distinction.","section":"Related Work and Table 4"},{"comment":"The 'matched extra QAT' and 'quantizer-only tuning' baselines are not described in detail. Please specify the number of parameters updated, hyperparameters, and number of epochs for these branches so that the diagnostic is reproducible.","section":"Figure 1"},{"comment":"The W3A3 rows show 0.09–0.11 for 5K/10K samples versus 40.21 for full data; this complete collapse is informative but should be discussed explicitly. Is it due to clipping failure, optimization instability, or something else?","section":"Table 7"}],"recommendation":"major_revision","confidential_remarks":"The manuscript presents striking empirical results, but the theoretical explanation is currently mismatched with the implementation, and the main comparisons mix different data budgets. The single-run evaluations and unreleased code make it impossible to verify robustness. These issues are addressable in a revision: reframe or fix the theory, clarify the calibration protocol, add equal-data comparisons, and release code or repeated-seed numbers. I would not accept the paper in its current form, but I do not see an unfixable fundamental error in the empirical claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The thing you should know first: this paper finds that updating only BN/RMSNorm affine parameters under the fake-quantized graph can recover collapsed low-bit accuracy. On ImageNet MobileNetV2 W4A4, NAP goes from 0.33% to 66.11%, beating QDrop (60.71%) without updating any conv or linear weights. That is a large effect, and it is backed by matched-subspace controls (random/Fisher subsets stay under 18%) and careful ablations of backend alignment and data coverage. If it is real, it is a genuinely useful knob for PTQ and post-QAT.\n\nWhat is actually new: Norm Tweaking already tuned normalization weights for LLM PTQ, and the paper explicitly acknowledges that. The new pieces are the BN-focused affine subspace formulation, the target-graph alignment, the alternating NAP-QAT for RMSNorm LLMs, and the controlled evidence that the subspace matters beyond parameter count. The Cityscapes U-Net result suggests the effect transfers across tasks, which is a good sign.\n\nNow the soft spots, in proportion. The theory in Eqs. 13–16 assumes a per-channel affine error model, but the implementation uses per-tensor activation quantization. So the \"fully cancel\" claim does not actually apply to the experimental graph. That does not kill the method—the gains may come from reducing dynamic range exposed to the per-tensor quantizer—but the abstract and theory section overstate what the analysis shows. The empirical results are also single runs with no error bars, and code is not released. On its own that would make me cautious, but here it is compounded by the ImageNet recipe needing 20 epochs on the full training set. That stretches the \"PTQ\" label; Table 7 shows 10K images drop accuracy to 3.88%. Credit where due: the paper is unusually honest about this in the body, including the failure cases (NAP hurts OOQ-Freeze at W2A4; NAP+QDrop is not uniformly better than NAP alone). Those caveats are not in the abstract, though.\n\nWho is this for? People working on low-bit PTQ/QAT for compact CNNs and LLMs. It is a plausible high-leverage adaptation idea that deserves a serious look, but it needs code, multiple seeds, and a revised theory section that either matches the quantizer or is framed as motivation rather than proof. I would send it to peer review, and I would ask for major revisions before trusting the headline.\n\nRecommendation: engage with it, but treat the empirical claim as promising rather than established until the reproducibility and theory gaps are closed.","headline":"A clearly-written paper with a striking empirical claim about quantization-friendly BN affine tuning; the theory is neat but mismatched to the implementation, and the evaluations are single-run, so the central claim needs more evidence before I'd trust it.","tokens_in":13209,"tokens_out":1903,"would_cite":false,"duration_ms":20475,"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":"Retuning only the normalization layer's per-channel scale and shift, all other weights frozen, restores collapsed 4-bit MobileNetV2 from 0.33% to 66.11% top-1 on ImageNet, beating a reconstruction-based quantizer.","keywords":["normalization affine parameters","low-bit quantization","post-training quantization","quantization-aware training","subspace optimization","batch normalization","RMSNorm","knowledge distillation"],"falsifier":"Disable the affine parameters entirely (fix γ=1, β=0 in every batch-norm layer) and rerun the NAP pipeline at W4A4 on ImageNet MobileNetV2; if accuracy still approaches 66%, the affine response subspace is not the causal mechanism. Complementarily, switch activation quantization from per-tensor to per-channel: the theorem predicts NAP's gain should grow if channel-wise affine error dominates, while flat or lower accuracy would refute the cancellation story.","tokens_in":12276,"feed_emoji":"🔢","tokens_out":15859,"duration_ms":127143,"temperature":0.7,"pith_summary":"Most low-bit quantization pipelines adapt every parameter at once, coupling weight, quantizer, and normalization updates into one optimization. This paper argues that the standard pipeline overlooks a decisive lever: the per-channel scale and shift of normalization layers. The proposed method, NAP, freezes all convolutional and linear weights and tunes only these affine parameters, on the same fake-quantized graph used at inference. On ImageNet MobileNetV2, the collapsed 4-bit baseline jumps from 0.33% to 66.11% top-1 accuracy (beating the 60.71% of a reconstruction-based quantizer) while updating about 1.4% of the trainable parameters. If the claim is right, a large share of quantization robustness lives in a subspace the field has been treating as an afterthought, and it can be bought almost for free.","feed_headline":"Normalization-only tuning lifts 4-bit MobileNetV2 from 0.3% to 66%","feed_subtitle":"With all other weights frozen, updating only normalization scale and shift beats full-model quantization tuning.","key_machinery":"The central object is the normalization affine subspace Φ = {γ_l, β_l}: each channel of a normalized activation is rescaled and shifted by a_{l,c} = γ_{l,c}·h_{l,c} + β_{l,c}, so one or two scalars broadcast control over a whole channel across all spatial positions or tokens. Around it sits a three-way decomposition — backbone weights W define features, quantization parameters Ω define the discretization grid, and Φ regulates the channel-wise responses the quantizer sees — with NAP freezing W and updating only Φ, always under the same fake-quantized graph used for evaluation. The compensating identity is Eq. (13), floor(a_c) = α_c·a_c + δ_c + e_c: given per-channel affine error (α_c, δ_c), c","core_discovery":"The paper's central claim: normalization affine parameters — γ and β in batch norm, γ alone in RMSNorm — form a low-dimensional response-control subspace with outsized leverage over quantization robustness. On a saturated QAT checkpoint, tuning only these 34.2K parameters lifts CIFAR-100 MobileNetV2 W4A4 from 68.21% to 69.83% while extra full QAT barely moves it. On ImageNet, RTN collapses MobileNetV2 to 0.33% at W4A4; NAP alone reaches 66.11%, beating QDrop's 60.71% without updating any conv or linear weight. The paper's local model: if a quantized pre-BN channel equals its affine map plus a residual, Eqs. (14)-(15) cancel the affine part exactly, leaving only nonlinear rounding and clippin","pith_inferences":["The per-channel cancellation theorem is proved under a per-channel error model, but the ImageNet activation quantizer is per-tensor; the theorem is thus an existence proof for channel-wise controllers, and its fit to the per-tensor experiments is an empirical bridge the paper does not fully justify. Switching activation quantization to per-channel would test whether the affine mechanism actually c","The sharp data-dependence cliff (3.88% at 10K images) suggests NAP's bottleneck is distribution coverage, not parameter count; synthetic or distilled calibration data mimicking the full training set is a direct test of whether the subspace claim survives without the full dataset.","The OOQ-Freeze failure suggests a cheap selection rule for post-QAT NAP: measure the alignment between the channel-wise affine component of the quantization error and the normalization Jacobian's column space, and skip the affine stage when alignment is low.","The broadcast-control argument should extend to any per-channel affine structure inside a quantized graph — LayerNorm weights, or an inserted per-channel scaling before the quantizer; NAP's distinctive move is reusing parameters already present rather than adding new ones."],"forward_implications":["Reconstruction-based PTQ gains a near-free preconditioner: on ImageNet W3A4, NAP+QDrop reaches 53.45% vs 51.75% for QDrop alone; on Cityscapes W3A4, 61.36% vs 58.21% mIoU.","Saturated QAT checkpoints are not locally optimal along the affine direction — NAP improves five of the tested W4A4 checkpoints (e.g., OOQ-Freeze 70.02→70.68) — so standard full-parameter training under-optimizes normalization responses.","Response adaptation and quantization-grid adaptation are complementary subspaces: alternating NAP-QAT on Qwen2.5-3B at W4A4 gives 166.65 perplexity on OSCAR vs 483.57 for AWQ-style scaling and 233.27 for grid-only tuning.","NAP adds no inference branch, keeps batch-norm folding intact, and tunes only ~1.4% of the parameters, at measured cost 1.0× vs 1.83× for full tuning.","The paper's own boundaries: compensation is backend-specific (mismatched graph costs 17.39 accuracy points), needs broad tuning-data coverage (10K ImageNet images give 3.88% vs 66.11% for the full set), and can conflict with specialized QAT checkpoints (W2A4 OOQ-Freeze drops 67.42→62.81)."],"supporting_citations":[{"why":"Defines batch normalization and its learnable per-channel affine parameters γ, β, the parameters NAP tunes.","marker":"(Ioffe and Szegedy 2015)"},{"why":"Defines RMSNorm with a learnable scale only, the analogous subspace NAP optimizes for LLMs.","marker":"(Zhang and Sennrich 2019)"},{"why":"QDrop, the reconstruction-based PTQ backend that NAP beats at W4A4 and combines with at low bit-widths.","marker":"(Wei et al. 2022)"},{"why":"LSQ, the learned-step-size QAT formulation whose saturated checkpoint dynamics motivate the post-QAT use of NAP.","marker":"(Esser et al. 2020)"},{"why":"LSQ+, the saturated QAT checkpoint used for the diagnostic study and post-QAT adaptation.","marker":"(Bhalgat et al. 2020)"},{"why":"OOQ oscillation-dampening and iterative-freezing QAT variants whose checkpoints NAP post-adapts.","marker":"(Nagel et al. 2022)"},{"why":"AWQ, the activation-aware scaling baseline and the basis of the AWQ-style equalization in the LLM experiments.","marker":"(Lin et al. 2024)"},{"why":"Norm Tweaking, the closest prior normalization-only LLM PTQ that NAP extends and explicitly differentiates from.","marker":"(Li et al. 2024)"},{"why":"Supplies the knowledge-distillation objective that is the main learning signal of NAP.","marker":"(Hinton, Vinyals, and Dean 2015)"}],"fun_headline_variants":["Freeze all but batch norm: 4-bit MobileNetV2 jumps to 66%","Scale-shift-only tuning: from 0.3% to 66% on 4-bit","Subspace trick: tune 34K params, beat full QAT on 4-bit","Normalization affine tuning rescues 4-bit models from collapse","Tune 34K params: 4-bit MobileNetV2 climbs from 0.3% to 66%"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The load-bearing premise is that each channel's quantization error decomposes into a channel-wise scale-and-offset term that batch-norm γ and β can cancel, but the implemented activation quantizer is per-tensor (one scale for all channels), so the theorem is a conditional existence result whose experimental setting only approximately matches it — and the headline ImageNet gain additionally assumes 20 training epochs over the full training set.","fun_headline_variants_meta":{"raw":{"variants":["Freeze all but batch norm: 4-bit MobileNetV2 jumps to 66%","Scale-shift-only tuning: from 0.3% to 66% on 4-bit","Subspace trick: tune 34K params, beat full QAT on 4-bit","Normalization affine tuning rescues 4-bit models from collapse","Tune 34K params: 4-bit MobileNetV2 climbs from 0.3% to 66%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000697,"raw_usage":{"total_tokens":3047,"prompt_tokens":864,"completion_tokens":2183,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":608,"completion_tokens_details":{"reasoning_tokens":2063}},"tokens_in":608,"tokens_out":2183,"duration_ms":14568,"temperature":1.0,"reasoning_tokens":2063,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T05:44:17.595649+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Disable the affine parameters entirely (fix γ=1, β=0 in every batch-norm layer) and rerun the NAP pipeline at W4A4 on ImageNet MobileNetV2; if accuracy still approaches 66%, the affine response subspace is not the causal mechanism. Complementarily, switch activation quantization from per-tensor to per-channel: the theorem predicts NAP's gain should grow if channel-wise affine error dominates, while flat or lower accuracy would refute the cancellation story.","supporting_citations":[],"review_version":1}