{"id":"160a5ca8-9a41-465e-ae7f-bd7c89293917","arxiv_id":"2509.00097","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"PEGE replaces the straight-through estimator with a curriculum-driven blend of quantized and full-precision weights plus an additive discretization-error correction, reporting small accuracy gains on low-bit CNNs.","lead":"PEGE is a new backward-pass estimator for quantized neural networks that adds a scaled discretization-error term to the straight-through gradient and progressively mixes full-precision and quantized weights. The paper reports small accuracy gains over STE and EWGS on CIFAR-10 and ImageNet, but omits code, seeds, error bars, and the values of its central hyperparameters.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'surpasses full-precision' claim rests on a single unseeded 0.21pp ImageNet gap, with schedulers and hyperparameters selected on test data.","rationale":"The paper's central claim is that PEGE lets low-precision models match or exceed full-precision accuracy. The strongest experimental support is the ImageNet W4A4 ResNet-18 result (70.01% vs. 69.80%). But this comparison is a single run with no error bars or seeds, and the reported gain is only 0.21 pp. The ablations select the scheduler and µ scheme by comparing test accuracy across multiple options, which is test-set selection. Without reproducibility details or multiple seeds, the central empirical claim is not statistically established. I do not see an internal mathematical contradiction in Eq. (5): it can be viewed as a straight-through gradient of an augmented loss L + (µ/2)||x_c − x_q||², though the paper does not present it that way. The more load-bearing weakness is the fragility of the empirical evidence, not the absence of a convergence proof. The reader's verdict of CONDITIONAL is appropriate; my concern strengthens the need for reproducibility checks but does not move the verdict.","tokens_in":7400,"tokens_out":4325,"duration_ms":44664,"concrete_test":"Rerun the ImageNet W4A4 ResNet-18 comparison (PACT forward; STE, EWGS, PEGE backward; plus full-precision baseline) with at least 5 independent seeds under the same protocol. Report mean ± std top-1 accuracy and a paired test for PEGE vs. full-precision and vs. STE. If the PEGE−FP mean difference is below ~0.2 pp or its 95% CI includes zero, the claim that PEGE 'surpasses' full precision is not established, and the verdict should remain conditional pending further evidence.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In §4.1, Table 2, the central ImageNet claim is PEGE at 70.01% vs. the paper's full-precision 69.80%, a 0.21 pp advantage; Table 1's CIFAR gains over STE are at most 0.45 pp. No error bars, seeds, or repeated runs are reported anywhere, and Figure 2 shows single training curves. A 0.21 pp gap is within typical run-to-run variation for ResNet-18 on ImageNet. Moreover, the ablations in §4.2 (Tables 3–4, Figure 3) select the logarithmic replacement scheduler and the exponential µ scheduler based on test accuracy, with no validation split, so the reported configuration is chosen on the test set. The constants in Eq. (4) (B, k, b) and Eq. (5) (µ_max, k) are never specified, making independent reproduction impossible. The heuristic in Eq. (5) could be defended as a penalty-style gradient; the load-bearing problem is that the headline improvement is supported by a single unseeded run with test-set-selected hyperparameters.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes PEGE, a gradient-estimation method for quantization-aware training (QAT). PEGE replaces STE's gradient with dL/dx_c = dL/dx_q + mu*(x_c - x_q), and progressively mixes full-precision and quantized parameters through a Bernoulli mask whose probability follows a logarithmic schedule. The authors claim this improves low-bit accuracy across CIFAR-10 (ResNet-20, VGG-16, W2A2) and ImageNet (ResNet-18, W4A4), and that PEGE enables low-precision models to match or surpass their full-precision counterparts. The method is intended as a drop-in backward-pass replacement compatible with different forward quantizers such as EWGS and PACT.","tokens_in":7769,"tokens_out":2797,"duration_ms":32129,"significance":"If the central claims held, PEGE would be a practically valuable contribution: it is simple, adds negligible overhead, and is forward-method-agnostic, addressing a known weakness of STE at low bit-widths. The paper also usefully frames QAT as a constrained optimization problem and provides ablations of the replacement scheduler and mu scheduler. However, the current evidence is not sufficient to establish the claims: the key gradient formula is asserted heuristically, the schedule hyperparameters are not reported, the ablation selects schedulers on test accuracy, and all experiments appear to be single unseeded runs. These issues directly affect the reproducibility and statistical reliability of the headline 'surpasses full-precision' result.","major_comments":[{"comment":"The central gradient estimator dL/dx_c = dL/dx_q + mu*(x_c - x_q) is introduced without derivation or theoretical justification. It is not shown to be a descent direction for the quantized loss, nor is it derived from the optimization objective in Eq. (6). The text calls the formulation 'theoretically grounded,' but no argument is provided. At minimum, the authors should derive the estimator from a surrogate loss (e.g., a penalty or proximal term) and discuss when it reduces to STE. The values of mu_max and k for the exponential mu schedule are also never reported, making the method impossible to reproduce.","section":"Sec. 3.2, Eq. (5)"},{"comment":"The logarithmic replacement scheduler p_T = min(log(kT+b)/B, 1.0) is under-specified: the base B, coefficient k, offset b, and the exact base of the logarithm are not given. More importantly, the choice of scheduler and the choice of mu scheduler are both made by comparing test accuracy on CIFAR-10 (Tables 3-4 and Fig. 3) without a held-out validation split. Thus the reported gains partly reflect test-set selection of hyperparameters. The authors should specify all constants and select schedules on a validation split, then report the corresponding test results.","section":"Sec. 3.2, Eq. (4) and Sec. 4.2, Tables 3-4, Fig. 3"},{"comment":"All experimental results are reported as single runs with no seeds, error bars, or statistical significance. The main ImageNet claim—PEGE at 70.01% vs. full-precision 69.80%, a 0.21 pp advantage—is within typical run-to-run variance for ResNet-18 on ImageNet, and the CIFAR-10 gains over STE are at most 0.45 pp. The paper needs multiple seeds with mean/std or a significance test to support the 'surpasses full-precision' claim. Without this, the headline conclusion is not robust.","section":"Sec. 4.1, Tables 1-2 and Fig. 2"},{"comment":"The comparison with 'STE (Orig. with FP: 71)' is not apples-to-apples: the cited PACT baseline uses a full-precision model at 71.0% accuracy, while the paper's own full-precision baseline is 69.80%. The reported -1.80 pp drop is therefore conflated with the baseline difference. The fair comparison is with the paper's own STE run (69.56%) and EWGS run (69.53%), which show PEGE ahead by 0.45-0.48 pp; this is still a small margin from a single run. The authors should either re-run PACT's original protocol under the same baseline or remove the 'Orig.' row.","section":"Sec. 4.1, Table 2"}],"minor_comments":[{"comment":"Typos: 'N etwork' in the title, and 'backprogation' in Section 5.","section":"Title and Sec. 5"},{"comment":"Figure 2b's caption says 'ResNet-20' but the accompanying text and Table 2 concern 'ResNet-18' on ImageNet.","section":"Sec. 4.1, Fig. 2 caption"},{"comment":"The description of Eq. (6) as a 'theoretically grounded' formulation is overstated; it is a standard constrained optimization statement. Consider tempering the language or providing a formal analysis.","section":"Sec. 3.3"},{"comment":"The notation is imprecise: 'Consider backpropagation step T, we generate an independent Bernoulli random variables, denoted as r_T'—the variable is a single Bernoulli, not plural; also clarify whether T indexes steps or epochs.","section":"Sec. 3.2, Eq. (3)"},{"comment":"The implementation details omit the initial learning rate, weight decay, data augmentation protocol, and the exact number of training epochs for ImageNet. These are needed for reproduction.","section":"Implementation Details, Sec. 4"},{"comment":"The vertical axis is labeled 'u' instead of the Greek mu used in the text.","section":"Fig. 3b"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my read on arXiv:2509.00097. PEGE is a simple gradient-estimation heuristic for quantization-aware training: use STE's gradient plus a scaled discretization-error term, and progressively switch from full-precision to quantized weights/activations with a log-curriculum schedule. Nothing here is individually new—EWGS already scales gradients by discretization error and progressive quantization is an old trick—but the specific combination, framed as a co-optimization of task loss and discretization error, is a reasonable drop-in that claims consistent small gains over STE and EWGS across CIFAR-10 and ImageNet. If the gains hold, it's useful for edge deployment. The writing is clear and the overhead is negligible.\n\nThe soft spots are real and load-bearing for the headline claims. The entire evaluation reports single runs: no seeds, no error bars, and the ImageNet 'surpasses full-precision' claim rests on a 0.21 pp advantage over the paper's own FP baseline—well inside run-to-run noise for ResNet-18. The ablations in Section 4.2 pick the logarithmic replacement scheduler and the exponential µ scheduler by test accuracy, with no validation split. So part of the reported gain is a fitted choice, not an independent test. On top of that, the constants in Eqs. (4) and (5)—B, b, k, µ_max—are never reported, so nobody can reproduce the method as described. Equation (5) itself is asserted without derivation; it could be defended as a penalty-style gradient, but the paper doesn't make that case.\n\nI don't think these problems amount to a fatal flaw. The heuristic is plausible, the results are consistently in one direction, and the paper is honest about its experiments. But the central claims—especially 'enables low-precision models to match or outperform full precision'—are not supported at the current level of evidence. The reader's conditional verdict is about right. One small thing: the citation to the author's own pruning paper [14] is not an issue; it's a normal reference.\n\nWho should read it: people working on QAT who want a simple alternative to STE and are willing to test it themselves. The paper is not ready for acceptance, but it deserves a serious referee. I'd send it to review with instructions to require seeds, error bars, validation-based schedule selection, hyperparameter values, and ideally code. If the gains survive that scrutiny, it's a modest but legitimate contribution. Without that, it's a promising tech report.","headline":"A plausible drop-in gradient estimator for QAT with small consistent gains, but the evidence is a single unseeded run with test-set-tuned schedules.","tokens_in":8130,"tokens_out":2175,"would_cite":false,"duration_ms":23862,"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 simple change to the gradient rule used in quantization-aware training lets low-bit networks match or beat full-precision accuracy.","keywords":["quantization-aware training","straight-through estimator","gradient estimation","low-bit quantization","discretization error","curriculum learning","model compression","image classification"],"falsifier":"Train a 2-bit ResNet-20 on CIFAR-10 with PEGE but set μ_max = 0, so the only change from STE is the progressive replacement schedule. If the reported 91.62% accuracy persists, the paper's core gradient-correction mechanism is not doing the work; if the reported accuracy requires specific μ_max and k values that the paper does not disclose, the central claim is not reproducible from the text.","tokens_in":7319,"feed_emoji":"🧮","tokens_out":4739,"duration_ms":48659,"temperature":0.7,"pith_summary":"The paper proposes PEGE, a direct replacement for the Straight-Through Estimator in quantization-aware training. Instead of passing the gradient through quantization as if the round function were the identity, PEGE progressively mixes full-precision and quantized values during training and adds a scaled discretization-error term to the STE gradient. The central claim is that this simple backward rule consistently improves low-bit accuracy across CIFAR-10 and ImageNet, letting a 2-bit ResNet-20/VGG-16 match its full-precision model and a 4-bit ResNet-18 surpass its full-precision counterpart on ImageNet. If correct, QAT can gain accuracy without changing the forward quantizer or adding meaningful overhead, since PEGE only modifies backpropagation.","feed_headline":"Gradient fix lets 2-bit nets match full precision","feed_subtitle":"Replacing STE with PEGE lifts low-bit accuracy on CIFAR-10 and ImageNet.","key_machinery":"The load-bearing object is the modified gradient estimator in Eq. (5): the STE gradient plus an element-wise correction μ·(x_c − x_q). The correction is scaled by μ, which grows exponentially with training steps, and is paired with a Bernoulli replacement mechanism whose probability follows a logarithmic curriculum, so fully quantized parameters gradually take over from full-precision ones. The mechanism lets the optimizer see both the task gradient and a direct penalty for deviating from quantized values, effectively co-optimizing prediction loss and discretization error.","core_discovery":"The paper's central claim is that the accuracy loss in low-bit quantization-aware training comes mainly from the backward pass, not the forward quantizer, and can be corrected by a gradient rule of the form ∂L/∂x_c = ∂L/∂x_q + μ·(x_c − x_q). Here x_c is the clipped full-precision value, x_q its quantized counterpart, and μ a growing scale. The first term is the standard STE gradient; the second term explicitly feeds the discretization error back into the optimization. PEGE combines this with a stochastic replacement schedule: at each step a Bernoulli variable decides whether the network trains on quantized parameters, with the replacement rate following a logarithmic curriculum, and μ growin","pith_inferences":["If the μ·(x_c − x_q) term is what matters, PEGE is effectively a penalty or regularization view of QAT; an ablation with μ_max=0 would isolate whether the replacement schedule alone, rather than the correction term, drives the reported gains.","The same correction idea may transfer to other nondifferentiable discrete operations, such as weight pruning masks or binary/ternary networks, where an STE-like identity gradient also ignores discretization error.","Because the paper does not report the values of μ_max, k, and B used in Eqs. (4)–(5), the practical claim is sensitive to schedule hyperparameters; a sensitivity study would be needed before the method can be used as a drop-in."],"forward_implications":["PEGE is backward-only: any existing forward quantizer, including EWGS and PACT, can keep its forward pass and gain accuracy by swapping the backward rule.","Low-bit models become practical: 2-bit CIFAR-10 models come within roughly 0.1–0.3 points of full precision, and 4-bit ResNet-18 exceeds full precision on ImageNet, at a 16x/8x compression ratio.","Training dynamics improve: PEGE converges faster than STE and EWGS in the reported learning curves, and it continues to benefit from longer training, unlike EWGS.","The gains are reported consistently across ResNet-20, VGG-16, and ResNet-18 on both CIFAR-10 and ImageNet, suggesting the effect is not tied to one architecture or dataset."],"supporting_citations":[{"why":"Introduces the Straight-Through Estimator that PEGE replaces; supplies the baseline backward method in all comparisons.","marker":"[1]"},{"why":"EWGS is the prior element-wise gradient scaling method and the forward quantizer used in the CIFAR-10 experiments.","marker":"[8]"},{"why":"PACT is the forward method used in the ImageNet experiments and provides the original STE baseline accuracy.","marker":"[2]"},{"why":"MQbench motivates the paper's goal by showing no QAT method consistently wins across architectures.","marker":"[11]"},{"why":"Defines quantization-aware training as the setting the paper targets.","marker":"[7]"}],"fun_headline_variants":["New backprop rule closes low-bit accuracy gap","PEGE gradient fix lifts 2-bit nets to full precision","Backward pass correction boosts quantized training","Progressive gradient beats STE for low-bit accuracy"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"Equation (5) assumes that adding a scaled discretization-error term to the STE gradient steers the full-precision weights toward a better optimum of the quantized loss; the paper gives no derivation or convergence argument for this, and the values of μ_max and the schedule constants are not reported.","fun_headline_variants_meta":{"raw":{"variants":["New backprop rule closes low-bit accuracy gap","PEGE gradient fix lifts 2-bit nets to full precision","Backward pass correction boosts quantized training","Progressive gradient beats STE for low-bit accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00022,"raw_usage":{"total_tokens":1297,"prompt_tokens":772,"completion_tokens":525,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":516,"completion_tokens_details":{"reasoning_tokens":464}},"tokens_in":516,"tokens_out":525,"duration_ms":5691,"temperature":1.0,"reasoning_tokens":464,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T15:13:15.877680+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train a 2-bit ResNet-20 on CIFAR-10 with PEGE but set μ_max = 0, so the only change from STE is the progressive replacement schedule. If the reported 91.62% accuracy persists, the paper's core gradient-correction mechanism is not doing the work; if the reported accuracy requires specific μ_max and k values that the paper does not disclose, the central claim is not reproducible from the text.","supporting_citations":[{"cited_title":"Network quantization with element-wise gradient scaling","cited_arxiv_id":null,"evidence_quote":"EWGS is the prior element-wise gradient scaling method and the forward quantizer used in the CIFAR-10 experiments."},{"cited_title":"MQBench: Towards Reproducible and Deployable Model Quantization Benchmark","cited_arxiv_id":"2111.03759","evidence_quote":"MQbench motivates the paper's goal by showing no QAT method consistently wins across architectures."}],"review_version":1}