acceptodds
Under review as a conference paper at ICLR 2027

Q-MINO: A Minimal-Norm Method for Quantization-Aware Training

Abstract

The Straight-Through Estimator (STE) is a widely used heuristic for Quantization-Aware Training (QAT), but its surrogate gradients can exhibit substantial mismatch with the underlying quantized objective, leading to noisy updates and parameter oscillations, particularly in ultra-low-bit regimes. We propose the Quantization-Aware Minimal-Norm Optimizer (Q-MINO), a temporal bundle method that combines gradient consensus, state-drift regularization, and an alignment constraint to construct stabilized, minimum-norm update directions from recent optimization states. Q-MINO solves the resulting constrained subproblem using a warm-started Frank–Wolfe procedure with a feasible fallback initialization. Theoretically, via a stochastic quasi-Lyapunov Kurdyka–\L ojasiewicz (KL) framework, we show that Q-MINO achieves almost sure convergence to a stationary point. Moreover, we detail numerical experiments with Q-MINO at various quantizations.

Primary area:
optimization
Keywords:
Quantization-Aware Training, Straight-Through Estimator, Bundle Methods, Frank-Wolfe Algorithm

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_EnMjOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

Q-MINO maintains a temporal bundle of EMA-filtered surrogate gradients and quantized parameter states. At each iteration, it approximately minimizes the norm of a convex combination of the stored gradients, with a quantized-state drift penalty and a current-gradient alignment constraint. A fixed number of Frank–Wolfe steps produces the direction for a proximal parameter update. The paper claims neighborhood convergence for a smooth surrogate objective through a stochastic Lyapunov/KL analysis. Experiments cover quantized synthetic optimization, a CNN on CIFAR-10 at INT8/INT4/INT2, and synthetic hyperparameter sensitivity studies.

Critical strengths

A coherent algorithmic construction with a feasible fallback. Combining temporal gradient aggregation with a penalty on stale quantized states is a reasonable approach to QAT instability. The current-gradient vertex is feasible for the alignment constraint, including when historical weights are no longer feasible. This gives the subproblem a clear interpretation and a valid fallback, although it does not establish convergence or an empirical advantage.

Critical weaknesses
  1. The supplied proof does not establish the central convergence claim. Appendix A, p.14, asserts a contraction of the tracking quantity VkV_k with forcing only of order ηk2\eta_k^2. A fixed-decay EMA of nonvanishing stochastic noise retains nonzero variance even at a fixed iterate, so this recursion needs a noise term and a derivation for the changing bundle direction. Equation (22) also contains ηkδ2\eta_k\delta^2, which is nonsummable for fixed δ>0\delta > 0 under the theorem's stepsize assumptions, while p.16 invokes Robbins–Siegmund with ηk2δ2\eta_k^2\delta^2 instead. The same page assumes compact sublevel containment without a supporting boundedness assumption and concludes distance convergence from a liminf stationarity estimate and vanishing increments, which are insufficient. There is a further implementation/proof mismatch: p.12 uses the norm bound implied by exact subproblem optimality, but the fixed-length Frank–Wolfe solver need not satisfy that bound, as the explicit example below shows. These gaps affect the main result.

  2. The empirical scope and related-work comparisons are insufficient to establish practical significance. Experiments are limited to 10-dimensional Rosenbrock/Rastrigin problems, a CNN on CIFAR-10, and synthetic sensitivity studies. There is no large-scale evaluation on ImageNet or larger vision/language models, despite the introduction's motivation from large-model deployment. This matters particularly because Section 5.2 acknowledges storage of twice the bundle size in full parameter-length auxiliary vectors, plus a Frank–Wolfe solve at every update. Without measured peak training memory, throughput, or time to a matched accuracy as model size increases, scalability and the accuracy–cost tradeoff remain unestablished. The acknowledged compute limitations explain the restricted evaluation but do not supply evidence for larger-scale applicability.

    Even within the tested setting, Table 1 reports Q-MINO accuracies of 76.63%, 75.06%, and 74.55%, versus Adam+STE's 83.79%, 83.32%, and 76.09% at INT8, INT4, and INT2. A smaller cross-bit accuracy drop does not establish an advantage when absolute accuracy remains lower. Seed variability and a complete training recipe are missing, and the sensitivity studies do not isolate the bundle against EMA/momentum alone or remove the drift and alignment components.

    Important related works are also absent: ProxQuant (Bai et al., ICLR 2019), which studies proximal formulations and stability for quantized training; Learned Step Size Quantization (Esser et al., ICLR 2020), an established low-bit QAT approach evaluated on ImageNet; and Oscillation-free Quantization for Low-bit Vision Transformers (Liu et al., ICML 2023), which directly addresses quantization-induced oscillations. The minimum-norm construction also needs positioning relative to Burke, Lewis, and Overton's gradient-sampling method (2005). Its sampled gradients differ from Q-MINO's temporal surrogate-gradient bundle, so its guarantees do not transfer automatically. Nagel et al. (2022) is already cited, but its oscillation-control methods are not experimentally compared. The concern is both incomplete attribution and a lack of evidence that the proposed mechanism adds value beyond relevant existing approaches; these methods are not all interchangeable optimizer baselines.

Questions for authors
  1. Convergence clarification. Can you derive a valid conditional recursion for VkV_k that includes the stochastic-noise contribution, and supply the boundedness and full-sequence convergence argument? How is the ηkδ2\eta_k\delta^2 term in Equation (22) handled without replacing it by a summable term, and what precise stationary-neighborhood guarantee follows?

  2. Finite-solver clarification. How does the implementation ensure the norm bound used on p.12? Take current gradient (1,0)(1,0), historical gradient (0.1,10)(0.1,10), ρ=0.1\rho=0.1, λ=0\lambda=0, and initialize at the current-gradient vertex. Both vertices are feasible. With the stated Frank–Wolfe stepsizes and five iterations, the historical weight becomes 1/151/15, giving d=(47/50,2/3)d=(47/50,2/3) and ∥d∥2=29881/22500>1\lVert d\rVert^2=29881/22500 > 1, the squared norm of the current gradient. Is an objective safeguard or an inexact-solve term missing from the algorithm and analysis?

  3. Empirical scope and related work. What existing measurements support scalability beyond the CIFAR-10 CNN, including peak training memory, throughput, and time to a matched accuracy with bundle storage and the inner solver included? Please provide the training/quantization configuration, tuning protocol, seed variability, and any matched EMA/momentum or oscillation-control comparisons. How does Q-MINO differ from the omitted quantization and minimum-norm methods identified above, and which benefits are established despite its lower accuracy than Adam+STE? Larger-model benchmarks and component ablations would strengthen a revision; the immediate clarification is what the current evidence supports.

Reviewer_NM3ZOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes Q-MINO, an optimizer for training neural networks with low-bit quantized weights. It combines recent gradients to make updates more stable, penalizes differences between past and current quantized weights, and keeps the update aligned with the current filtered gradient. The paper claims convergence to a neighborhood of stationary points under its assumptions. It evaluates the method on two synthetic functions and a CNN on CIFAR-10 at several quantization levels.

Critical strengths

1) The paper addresses an important problem: improving training stability under low-bit quantization. The proposed combination of gradient history and quantized-state information is well motivated.

2) The experiments cover several quantization levels and include sensitivity studies for the main hyperparameters.

Critical weaknesses

1) On page 14, lines 712–713, the proof claims E[Vk+1∣Fk]≤(1−β)Vk+C1ηk2\mathbb E[V_{k+1}\mid\mathcal F_k] \le (1-\beta)V_k+C_1\eta_k^2 without justification.

Since ηk→0\eta_k \to 0, this recurrence keeps E[Vk]\mathbb E[V_k] bounded. Let A:=lim sup⁡k→∞E[Vk]A:=\limsup_{k\to\infty}\mathbb E[V_k]. Taking expectations and then the limsup, we obtain A≤(1−β)AA\le(1-\beta)A. Since 0<β<10<\beta<1 and Vk≥0V_k\ge0, we obtain A=0A=0, hence E[Vk]→0\mathbb E[V_k] \to 0.

However, each EMA update introduces fresh gradient noise. Even for a constant objective and bundle size m=1m=1, bounded, independent, zero-mean gradient noise ξk\xi_k with E∣∣ξk∣∣2=v>0\mathbb E||\xi_k||^2=v>0 gives dk=gˉkd_k=\bar g_k and Vk=∣∣gˉk∣∣2V_k=||\bar g_k||^2. The EMA update therefore gives E[Vk+1]=β2E[Vk]+(1−β)2v\mathbb E[V_{k+1}]=\beta^2\mathbb E[V_k]+(1-\beta)^2v, which implies E[Vk]⟶1−β1+βv>0.\mathbb E[V_k]\longrightarrow \frac{1-\beta}{1+\beta}v>0. Thus, the claimed bound is missing a noise term. Since the convergence proof relies on this bound, the subsequent argument needs to be revised.

2) On page 16, lines 835–837, the proof states that the iterates remain in a compact sublevel set. However, the stated assumptions do not guarantee this. A function can be proper, lower semicontinuous, and bounded below while having unbounded sublevel sets: the constant function Φ(x)=0\Phi(x)=0 is a simple example. The proof of Lemma 2 on page 11, line 579, assumes sup⁡k∣∣xk∣∣≤D\sup_k||x_k||\le D, but this assumption is not included in Section 3.1 or Theorem 5, and no proof of boundedness is provided. The authors should state this additional assumption explicitly or prove boundedness, and explain why the claimed compactness follows.

3) In Equation (22), the error term is ηkδ2\eta_k\delta^2, but on page 16, line 848, it becomes ηk2δ2\eta_k^2\delta^2 when applying the Robbins–Siegmund theorem. I couldn't find a justification for this change. For fixed δ>0\delta>0, the assumption ∑kηk=∞\sum_k\eta_k=\infty means that ηkδ2\eta_k\delta^2 is not summable. Therefore, the convergence of the Lyapunov sequence does not follow from the argument as written.

4) On page 16, lines 855–859, the proof combines vanishing successive changes with a liminf bound on expected stationarity to conclude that E[dist⁡(xk,Uδ)]→0\mathbb E[\operatorname{dist}(x_k,U_\delta)]\to0. I couldn't follow this implication. Even a nonnegative sequence with vanishing successive changes and liminf equal to zero need not converge to zero. For example, ak=1+sin⁡(log⁡(k+1))a_k=1+\sin(\log(k+1)) satisfies both properties but keeps oscillating. An additional argument is needed to establish convergence of the expected distance to UδU_\delta.

5) The experiments don't clearly show better training stability. In Table 1, Q-MINO is less accurate than Adam+STE at every precision (e.g., 74.55% versus 76.09% at INT2) and less accurate than SGD+STE at INT8 and INT4. Its accuracy decreases less from INT8 to INT2, but it is already much lower at INT8. The synthetic sensitivity study averages over 100 randomized initial trajectories, but the CIFAR-10 results do not report variability across runs. Measuring weight oscillations, ablating the drift penalty and the alignment constraint separately on CIFAR-10, and reporting CIFAR-10 results across several random seeds with means and standard deviations would help show whether the method improves stability.

Justification of the score I recommend rejection because the convergence guarantee, a central contribution of the paper, is not established by the current proof, and the experiments don't clearly demonstrate improved training stability. I believe the issues detailed above require substantial corrections to the analysis and stronger experimental support.

Questions for authors

1) Can you derive the tracking-error recurrence on page 14, lines 712–713, including the contribution of fresh gradient noise, and explain how the corrected bound affects Theorem 5?

2) What guarantees that the iterates remain bounded? Is an additional assumption needed for the compactness argument on page 16?

3) How can Robbins–Siegmund be applied when Equation (22) contains the nonsummable term ηkδ2\eta_k\delta^2 for fixed δ>0\delta>0?

4) Can you provide the missing argument connecting the liminf stationarity bound and vanishing successive changes to E[dist⁡(xk,Uδ)]→0\mathbb E[\operatorname{dist}(x_k,U_\delta)]\to0?

Reviewer_s8eyOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper addresses gradient mismatch caused by the straight-through estimator and parameter oscillations in low-bit quantization-aware training. The proposed method, Q-MINO, keeps a history of exponential moving average gradients and quantization states. It uses this history to formulate a minimum-norm subproblem with a state-drift penalty and a constraint that encourages alignment with the current gradient. The update direction is then computed using the Frank–Wolfe algorithm. The paper also presents a convergence analysis based on stochastic Lyapunov arguments and the Kurdyka–Łojasiewicz property, along with experiments on synthetic functions and CIFAR-10.

Critical strengths

The paper tackles a practical issue in low-bit training, where parameter oscillations and gradient mismatch can make optimization difficult. I think this is a worthwhile problem to study.

The motivation for combining historical gradients with quantization-state information is clear. Using the current gradient as a fallback keeps the subproblem feasible, and I appreciate that the authors discuss the memory cost of storing past states.

Critical weaknesses

The theory does not account for errors from finite Frank–Wolfe iterations. The link between the smooth surrogate and the actual quantized objective is also unclear.

Q-MINO underperforms Adam+STE at all tested bit widths on CIFAR-10. A smaller accuracy drop from a lower starting point is not convincing evidence of robustness.

The ablations do not isolate the effects of the EMA, drift penalty, and alignment constraint. Direct stability measurements are also missing.

More details on training, baseline tuning, and results across random seeds are needed, along with runtime and memory comparisons.

Questions for authors

Please address the concerns listed under Weaknesses.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.