acceptodds
Under review as a conference paper at ICLR 2027

IDScore: Task-Conditioned Likelihood Correction for Low-Bit QAT

Abstract

Quantization-aware training (QAT) is a key approach for optimizing quantized neural networks. However, deterministic quantization introduces piecewise-constant mappings that make direct gradient-based optimization difficult, so QAT relies on surrogate gradients whose quality is critical to effective training. We propose Identification-Score (IDScore), a correction that improves surrogate gradient estimation in QAT with little additional training cost. This improvement comes from a new perspective: motivated by quantized system identification, IDScore exploits additional structure information from the quantizer itself, beyond the external signals used by existing methods. Concretely, IDScore constructs a local likelihood score from a task-selected neighboring target, the latent parameter, and quantization thresholds, and uses it to form a bounded residual correction. Our theory explains why IDScore provides a meaningful correction direction and how its correction strength should be allocated across parameters. Formally, we prove calibrated target-direction consistency and constrained-optimal likelihood allocation, and connect the resulting correction to threshold progress and a local sufficient condition for task-loss descent. Experiments on vision and language tasks show that IDScore consistently improves the tested LSQ and QuEST baselines at low additional cost. Evaluations span W1A1-W4A4 precision and models with up to 430M non-embedding parameters. It also yields further gains when combined with FOGZO or CAGE. Controlled ablations support the likelihood-based allocation mechanism characterized by our theory. Together, these results provide both an efficient approach with practical potential for low-bit training and a broader perspective on how quantization structure can inform backward optimization.

Primary area:
optimization
Keywords:
Quantization-Aware Training, Low-Bit Neural Networks, Efficient Training, Gradient Estimation, Likelihood-Based Learning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_ZsZ1OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This article proposes IDScore, a task conditional likelihood correction method for low bit quantization perception training. This method utilizes task gradients to select adjacent quantization targets, and combines latent parameters with quantization thresholds to construct ordinal likelihood scores, which are added as bounded residuals to existing backpropagation rules. In theory, this article analyzes the consistency of target direction, residual allocation under limited budget, and local threshold approximation properties. The experiment covered visual and language tasks, multiple bit widths, and up to 430 million parameter scales on LSQ and QuEST, and further evaluated the combination effect with FOGZO and CAGE, optimizer adaptability, computational cost, and ablation results.

Critical strengths
  1. It can improve the tested QAT baseline in both visual and linguistic scenarios, as well as in various quantization accuracies. In some high demand low bit settings, the improvement is particularly significant. At the same time, the measurement cost introduced by this method in the QuEST benchmark test report is relatively small, which makes it practically applicable.

  2. This article not only reports the performance of the final task, but also investigates different amplitude allocation, residual support, optimizer paths, and combinations with FOGZO and CAGE. These experiments help to separate the several components of the proposed correction method and demonstrate that IDScore can serve as a supplement to existing gradient estimation methods without completely replacing the underlying QAT method.

Critical weaknesses
  1. The biggest problem is that the current evidence is not sufficient to clearly prove that a complete ordinal likelihood construction is necessary for the reported performance improvement. The core motivation of IDScore is to utilize ordered threshold structures through target conditional ordinal likelihood. However, Boundary-NLL control has improved the performance of QuEST from 37.262 PPL to 36.611 PPL, while the full IDScore has reached 36.502. Therefore, in this case, a simpler single bound likelihood seems to be able to restore most of the performance improvement. This makes it difficult for me to determine whether aggregating all quantization threshold information is crucial for this method, especially when ordinal scores already exist.

  2. This article proposes IDScore to improve proxy gradient estimation, but the main evaluation focuses on downstream accuracy or perplexity. Is there direct evidence to suggest that IDScore can generate more reliable gradient estimates, such as higher consistency with finite difference or zero order reference values in easily manageable scenarios?

  3. The PPL of Boundary-NLL is 36.611, while the complete IDScore is 36.502 and QuEST is 37.262. Can the author provide a comparison of the matching between Boundary-NLL and the complete ordinal IDScore in stages W1A1, W2A2, W3A3, and W4A4?

Questions for authors

Please see the weaknesses.

2 comments
AuthorsOpenReview ↗

Response to Weaknesses 1 and 3: Added value of all-threshold aggregation

Thank you for pointing out that Boundary-NLL already captures most of the improvement. We agree with this observation: in the original W2A2 experiment, Boundary-NLL reduces PPL from 37.262 to 36.611, while ordinal IDScore further reduces it to 36.502. Our main contribution is a task-conditioned likelihood correction for QAT, with the all-threshold ordinal likelihood as our default construction. The gains from Boundary-NLL support the usefulness of this shared likelihood-based correction principle. Our revised comparison therefore examines the additional value of all-threshold aggregation.

For a comparison with matched residual support, we added Boundary-NLL-matched, which uses the same residual support and post-processing rules as Ordinal. The original Boundary-NLL scores only coordinates with feasible neighboring transitions. The matched variant additionally handles coordinates whose clipped target remains at the current level, using the nearest finite boundary of the retained cell. Its W2A2 PPL is 36.673. We retain the original 36.611 result in Table 14 and distinguish these two implementations in Appendix D.5.

The W2A2–W4A4 comparisons use the same code snapshot and matched training configurations within each bit width. All runs use QuEST-30M on C4 with a sequence length of 512 and AdamW with a cosine learning-rate schedule. W1A1 uses 11,444 updates with a global batch of 256 sequences (about 1.5B tokens), W2A2 uses 91,552 updates with a global batch of 64 (about 3B tokens), and W3A3 and W4A4 use 91,552 updates with a global batch of 32 (about 1.5B tokens each). For the W2A2–W4A4 reruns, Boundary-NLL-matched and Ordinal use the same training configuration within each bit width.

Across three seeds at each of W2A2, W3A3, and W4A4, Ordinal achieves lower PPL in all nine paired comparisons, with mean paired reductions of 0.172, 0.197, and 0.141, respectively (as shown in the table below). At W1A1, there is only one threshold, so the two score constructions coincide under the same adapter and post-processing rules. We therefore did not separately train this identical control. The reported W1A1 setting uses 11,444 updates with global batch size 256, corresponding to approximately 1.5B input tokens.

Table 1. Final validation PPL on QuEST-30M (lower is better).

PrecisionSeedQuESTBoundary-NLLOrdinalPPL gain
W1A1054.759--49.841--
154.349--49.816--
253.660--49.598--
Mean54.256±0.45354.256\pm0.453--49.752±0.10949.752\pm0.109--
W2A2037.23336.69036.467+0.223
137.37436.73636.562+0.174
237.17836.59436.476+0.118
Mean37.262±0.08337.262\pm0.08336.673±0.05936.673\pm0.05936.502±0.04336.502\pm0.043+0.172±0.043+0.172\pm0.043
W3A3033.62533.41233.226+0.186
133.79333.50133.309+0.192
233.81633.55933.346+0.213
Mean33.745±0.08533.745\pm0.08533.491±0.06033.491\pm0.06033.294±0.05033.294\pm0.050+0.197±0.012+0.197\pm0.012
W4A4030.87730.77630.660+0.116
131.05630.95330.797+0.156
230.99230.88830.737+0.151
Mean30.975±0.07430.975\pm0.07430.872±0.07330.872\pm0.07330.731±0.05630.731\pm0.056+0.141±0.018+0.141\pm0.018

*At W1A1, there is only one threshold, so the two score constructions coincide under the same adapter and post-processing rules. We therefore did not separately train this identical control. Positive paired gains indicate lower PPL for Ordinal.

The distinction between the constructions is that a single-boundary score encodes the selected threshold event, whereas the ordinal score encodes the target level's relationships to all thresholds. These constructions produce different coordinate-wise correction magnitudes. The matched comparisons support the additional empirical value of its ordinal construction in the tested multi-bit settings.

This additional benefit comes with 6.34% and 2.03% more step time at W3A3 and W4A4, respectively, relative to Boundary-NLL-matched, with nearly unchanged peak allocated memory. W2A2 timings vary more across trials, so we do not interpret the negative mean time difference as a speedup. The complete timing measurements and protocol are reported in Appendix D.5, Table 17.

Together, these results support our choice of the ordinal construction as the default implementation. The revised text following Eq. (5) clarifies the role of all-threshold aggregation, Section 4.4 discusses its added empirical value, and Appendix D.5 reports the cross-bit-width comparisons and efficiency measurements. All corresponding revisions are highlighted in blue.

AuthorsOpenReview ↗

Response to Weakness 2: Direct comparison with a zeroth-order reference.

Thank you for your insightful comment. We added a controlled local diagnostic comparing LSQ-STE and LSQ+IDScore at the same frozen ResNet-18/ImageNet-100 checkpoints while retaining the same forward quantizer. For each sampled weight pair, we evaluate the task loss at all 16 joint quantized states and construct a local Gaussian-smoothed loss from these evaluations.

Let (ui,uj)(u_i,u_j) be the normalized coordinates of a sampled weight pair. Let Lkℓ\mathcal L_{k\ell} be the task loss when the pair is set to quantization levels indexed by (k,ℓ)(k,\ell), with all other weights, quantization scales, and validation inputs held fixed. Define pik(ui;σ)=Pr⁡(ui+σZi∈Ck)p_{ik}(u_i;\sigma)=\Pr(u_i+\sigma Z_i\in C_k), where CkC_k is the quantization cell for level kk and Zi,ZjZ_i,Z_j are independent standard Gaussian variables. The smoothed loss and its reference gradient are

= \sum_{k=0}^{3}\sum_{\ell=0}^{3} p_{ik}(u_i;\sigma)\\,p_{j\ell}(u_j;\sigma)\\, \mathcal{L}_{k\ell},$$ $$\mathbf{h}_{\sigma} = \nabla_{(u_i,u_j)} \widetilde{\mathcal{L}}_{\sigma}(u_i,u_j).$$ We compute cosine similarity by pooling the sampled coordinates within each seed, with both backward signals expressed in the same normalized coordinates. The reference is the gradient of this local smoothed loss, rather than of the hard-quantized objective. The paired cosine gain is positive for each of the three seeds at both smoothing scales. The mean gain is $9.33\times10^{-5}$ at $\sigma=1$ and $7.27\times10^{-5}$ at $\sigma=0.5$. These results indicate a small but consistent improvement in local directional alignment across the tested seeds. | Smoothing scale | Seed | LSQ-STE | LSQ+IDScore | Paired gain ($10^{-5}$) | |---|---|---:|---:|---:| | $\sigma=1$ (primary) | 0 | 0.92027323 | 0.92033914 | +6.591 | | | 3 | 0.93780127 | 0.93789688 | +9.561 | | | 4 | 0.90085123 | 0.90096947 | +11.824 | | | Mean | 0.91964191 | 0.91973516 | +9.33 | | $\sigma=0.5$ | 0 | 0.95049302 | 0.95053788 | +4.486 | | | 3 | 0.95666936 | 0.95675166 | +8.229 | | | 4 | 0.93895217 | 0.93904298 | +9.081 | | | Mean | 0.94870485 | 0.94877751 | +7.27 | We use LSQ for this comparison because IDScore preserves the support and signs of the weight gradient in the tested setting, allowing us to examine the effect of reweighting relative gradient magnitudes. In QuEST, the correction can additionally supply nonzero signals outside the trust mask, so alignment changes would reflect both support changes and magnitude reweighting. We have added this comparison to Section 4.4, "Alignment with a zeroth-order reference." This diagnostic complements the downstream performance results by directly examining alignment with a reference constructed from task-loss evaluations.
Reviewer_apgmOpenReview ↗
Rating
3
Confidence
4(+2)
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This work proposes IDScore, a method that aims to improve the quality of the estimated gradient (of task loss) in Quantization-Aware Training, by adding a residual correction term to the estimated gradient (of task loss). Here is how i understand the proposed method to be:

After backprop, each parameter in the neural network gets a (possibly estimated) partial derivative from the task loss. The sign of this partial derivative identifies an immediate neighbouring quantization level (either the left immediate neighbour or the right immediate neighbour). The parameter takes a few steps towards the neighbour (as guided by the task loss partial derivative), but IDScore adds a few additional steps in the same direction (towards the identified immediate neighbor). The number of additional steps correlates with the distance between the parameter’s current location and the neighbour’s location. The number of additional steps also correlates with the magnitude of the parameter’s partial derivative relative to the magnitudes of partial derivatives of other parameters in the neural network. The number of additional steps is upper bounded by a hyperparameter c_max. While this identified immediate neighbor, in practice, for a fixed training iteration, can only be one of two immediate neighbors depending on the parameter’s current location, the authors (conceptually) model the identified neighbor as a sample from a normal distribution with some underlying mean (which is called the “identification” and I understand it to be an temporarily optimal value of the parameter) and some underlying standard deviation (which the authors try to guess in practice by setting a hyperparameter tau). Therefore, instead of only pushing towards the identified immediate neighbouring level, IDScore actually pushes the parameter towards all quantization levels jointly, although the push force towards farther away levels is smaller. A third hyperparameter lambda is introduced to control the strength of IDScore.

The paper provides theoretical analysis that reveals various insights about IDScore, experimental evidence that shows IDScore performs well in practice, and various ablation studies empirically justifying the design choices of IDScore.

Critical strengths

The method is novel and makes meaningful progress in the field of QAT gradient estimation. The empirical results are thorough and convincing, spanning multiple architectures, datasets, and model sizes.

Critical weaknesses

There are no critical weaknesses.

Questions for authors

The following are merely suggestions. Addressing them is optional and does not affect my decision to recommend acceptance.

  1. Consider adding a short description (e.g. a single sentence) explaining the connection between each theorem and the actual method used in practice. Such short descriptions help readers understand why they are reading the theorems. Section 3.1, consider adding a single sentence: “we hope/have reason to believe that in practice, s from Equation (4) resembles the behaviour of S from Theorem 3.1, because …”. Section 3.2, consider adding a single sentence: “we hope that D_S being small is a good thing because …”.
  2. Consider expanding the x-axis of Figure 2 to show the existence of a knee. Hyperparameters (HPs) typically have a “knee”, or an optimal value. Increasing the HP before the HP reaches the knee generally improves model accuracy, whereas increasing the HP after reaching the knee generally decreases model accuracy. If the knee does not exist, setting an HP should be trivial. Just set the HP to extreme values like 0 or infinity. I am guessing that tau, lambda, and c_max do have such knees. If such an HP sweep is too expensive, consider verbally explaining what happens when these HPs are set to extreme values like (almost) 0 or (almost) infinity.
  3. Equation 5: consider adding a single sentence explaining what happens when the loss is minimized, or when its gradient is zeroed, preferably immediately after the equation is introduced.
  4. Consider adding a short description (e.g. a single sentence/paragraph) discussing why the authors think IDScore performs better (i.e. in terms of global QAT convergence) than baselines like the STE in practice. The empirical results already support this claim (that IDScore performs better than baseline methods in terms of global QAT convergence), but I am interested in the “why”, not just the “whether”. This short description does not need to be rigorously proved. Even a hypothesis or a potential explanation can significantly help/inform the community and potentially inspire future research.
  5. Consider making the abstract more accessible to first-time readers. For example, the word “identification” first shows up in the abstract around line 017, but it means very little to a first-time reader until they reach theorem 3.1 which lives on page 4. I understand that these are complex concepts, and that there is limited space in the abstract, and I also have no concrete suggestions on how to improve readability. However, I do believe that with sufficient effort, complicated concepts can be explained in a more accessible manner, even when there is limited space.
5 comments
AuthorsOpenReview ↗

Response to Questions 1 and 2

Question 1: How does each theorem relate to the practical method?

Thank you for the helpful suggestion. We have added brief practical interpretations in Sections 3.1–3.3.

In Section 3.1, we clarify that the task-selected code ss serves as a deterministic pseudo-observation of the same threshold events as SS. We do not assume that ss follows the working model’s sampling distribution of SS. The calibrated target-direction result establishes a direct link between likelihood modeling and task-aware QAT: task-selected pseudo-targets align likelihood descent with the direction favored by the task gradient, providing a principled basis for the backward correction.

In Section 3.2, we explain that a small DSD_{\mathcal S} indicates better task-weighted agreement with the neighboring targets preferred by the current linearized loss. This clarifies the practical meaning of the mismatch measure used in the allocation analysis.

In Section 3.3, we explain how the correction promotes progress toward the selected thresholds under the stated direct sign-preserving update, and identify a sufficient condition for the resulting quantized transitions to decrease the task loss.

Together, these additions clarify how the theory explains IDScore’s correction direction, magnitude allocation, and resulting quantized transitions. All corresponding revisions are highlighted in blue.

Question 2: Do the hyperparameters have optimal values? What happens at extreme settings?

Thank you for your insightful suggestion. We have extended the one-parameter-at-a-time sweeps for τ\tau, λ\lambda, and cmax⁡c_{\max} on QuEST-30M W2A2. The reported results combine the original runs with new range extensions, all using seed 0 and final validation after 91,552 updates. The other parameters remain at their default values during each sweep.

The extended results show lower PPL at intermediate values of τ\tau and λ\lambda than at the tested extremes. For τ\tau, PPL is 35.822 at 0.25, compared with 37.937 at 0.0625 and 37.359 at 2. For λ\lambda, PPL is 36.349 at 0.01, compared with 36.802 at 0.002 and 37.317 at 0.05. These results show that making the temperature smaller or the correction stronger does not continually improve performance.

For cmax⁡c_{\max}, tight clipping reduces the gain: PPL is 37.152 at cmax⁡=0.5c_{\max}=0.5, compared with 36.467 at the default cmax⁡=3c_{\max}=3. Removing the upper cap gives a similar observed PPL of 36.564 in this seed-0 comparison. These results do not identify a sharply defined optimum for cmax⁡c_{\max}.

The main experiments retain the reported default configuration; the extended sweeps characterize sensitivity. Figure 2 has been updated, and the complete results are reported in Appendix D.7, Table 19.

Original and extended sensitivity results on QuEST-30M W2A2. All entries are final validation PPL for seed 0 after 91,552 updates. One parameter is varied at a time; the others retain their default values (τ,λ,cmax⁡)=(1,0.005,3)(\tau,\lambda,c_{\max})=(1,0.005,3). Bold marks the lowest PPL within each sweep.

QuEST reference (seed 0): 37.233.

(a) Likelihood temperature

τ\tauλ\lambdacmax⁡c_{\max}Final PPL ↓\downarrow
0.06250.005337.937
0.1250.005337.337
0.250.005335.822
0.50.005336.149
0.750.005336.338
10.005336.467
20.005337.359

(b) Residual coefficient

τ\tauλ\lambdacmax⁡c_{\max}Final PPL ↓\downarrow
10.002336.802
10.005336.467
10.01336.349
10.05337.317

(c) Score-magnitude upper cap

τ\tauλ\lambdacmax⁡c_{\max}Final PPL ↓\downarrow
10.0050.537.152
10.005136.795
10.005336.467
10.005∞\infty36.564

*Repeated default entries refer to the same historical run. cmax⁡=∞c_{\max}=\infty removes the upper cap. These single-seed sweeps characterize sensitivity rather than establish statistically significant differences between nearby settings.

AuthorsOpenReview ↗

Response to Questions 3-5

Question 3: What does it mean when Eq. (5) is minimized or its gradient vanishes?

Thank you for this insightful suggestion. We have added an explanation immediately after Eq. (5). For a fixed internal target under calibration, the likelihood loss reaches its minimum inside the target quantization cell. At this point, the likelihood gradient vanishes, so the corresponding IDScore residual becomes zero for that fixed target. The added explanation is highlighted in blue.

Question 4: Why does IDScore perform better than baselines such as STE in practical training?

Thank you for this meaningful suggestion. We have added an intuitive explanation in Section 4.4. IDScore adapts the correction to both task sensitivity and threshold-relative position. Coordinates with similar task gradients may require different movements to reach their selected neighboring levels. The likelihood score captures this quantization geometry, while the task-gradient magnitude weights the correction by task sensitivity. Normalization and clipping keep the additional correction bounded. Repeated over training, this state-dependent allocation can help translate task preferences into useful quantized transitions. This interpretation is consistent with our norm-matched ablations, which support the importance of coordinate-wise allocation beyond the total residual norm. The added discussion is highlighted in blue.

Question 5: How can the abstract be made more accessible to first-time readers?

Thank you for the helpful suggestion. We have clarified “quantized system identification” at its first occurrence in the abstract as “inferring continuous parameters from quantized observations.” We also split the original sentence into two shorter sentences for readability, introducing this idea before describing the IDScore construction. This helps readers understand the motivation without first consulting the theoretical analysis. The revised text is highlighted in blue.

Thank you for addressing my concerns. I have no further questions.

Reviewer_eEi5OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes IDScore, a bounded residual correction to existing quantization-aware training gradients. The upstream task gradient selects an adjacent quantization target, and a Gaussian ordinal likelihood determines the correction magnitude from the latent coordinate and quantization thresholds. The correction preserves the forward quantizer and requires no additional model evaluation or optimizer state. The theory establishes calibrated direction consistency, constrained allocation optimality for a frozen likelihood potential, and conditional task-loss descent for direct sign-preserving updates. Experiments cover LSQ and QuEST on vision and language tasks, W1A1 through W4A4, models up to 430M non-embedding parameters, and combinations with FOGZO and CAGE.

Critical strengths
  1. Useful empirical gains at low measured cost. On QuEST-30M W2A2, validation perplexity improves from 37.262 to 36.502, with approximately 2% additional step time and 0.4% additional peak allocated memory in the synchronized benchmark (Table 5). Improvements over QuEST appear across the tested scales and bit widths. Adding IDScore also improves CAGE, even though CAGE is stronger on its own in those comparisons (Tables 2–3). The combination of accuracy, measured cost, and compatibility supports the method's practical value.

  2. Informative controls for the correction mechanism. The norm-matched controls preserve the pseudo-target rule, support, correction sign, task-gradient magnitude factor, and per-tensor residual norm while changing the coordinate allocation. IDScore outperforms all three controls on each of the three matched seeds (Table 12). The norm is matched to the counterfactual IDScore correction at each control's own training state. Together with the calibration and adapter audits, these results provide evidence that coordinate allocation matters beyond the total correction norm.

Critical weaknesses
  1. The theoretical explanation is limited for the evaluated optimizer dynamics. Theorem 3.2 establishes first-order optimality for the method's frozen likelihood potential, with the norm budget set by the IDScore residual itself. It does not establish optimal allocation for task-loss reduction. Theorem 3.3 assumes a direct sign-preserving update and a condition controlling the finite-transition remainder; it does not cover the AdamW or Muon dynamics evaluated in Table 4. The paper explicitly acknowledges these distinctions. My concern is therefore the limited explanatory reach of the theory, rather than an incorrect claim of global convergence.

  2. The supplied PDF leaves important experimental details and efficiency coverage unresolved. Section 4 and Appendices B–C provide adapters, correction hyperparameters, and seed statistics, but do not enumerate complete training schedules and baseline tuning configurations for the main comparisons. The DeiT-specific adapter is also not explicitly identified in Appendix B.1. Table 1 mixes fully matched local comparisons with clearly disclosed, partially unpaired author-log comparisons, and the 430M result uses one matched seed. The overhead measurement is restricted to 30M W2A2, whereas the ordinal calculation scales with the number of thresholds (Appendix B.4). These limitations reduce confidence in reproducibility and in extending the measured efficiency to all tested settings; they do not invalidate the matched smaller-scale gains.

Questions for authors
  1. Experimental clarification. Can you provide the complete training configurations and baseline tuning protocol for Tables 1–5, identify the DeiT adapter and the numerical coordinate/threshold scaling used by QuEST, and specify which training settings were matched in the author-log comparisons? Are timing and memory measurements available at another bit width or scale?

  2. Mechanism clarification. Is there trajectory-level evidence connecting the correction to task-beneficial quantized transitions under AdamW or Muon? Table 10 shows that restricting the residual to the QuEST trust mask worsens perplexity from 36.502 to 39.965. The approximately 4.35% outside-mask residual energy in Table 8 is measured only at final checkpoints, so it cannot by itself explain the training effect. Any existing transition or optimizer-update diagnostics would help assess this connection.

4 comments
AuthorsOpenReview ↗

Response to Question 1: Experimental Configuration, Fairness of Comparisons, and Efficiency Measurements

Thank you for pointing out the need for these clarifications. We have expanded the adapter descriptions in Appendix B, added the protocols for Tables 1–5 in Appendix C, replaced the previously author-log-based comparisons with matched local runs, and extended the efficiency measurements in Appendix D.2.

1. Training configurations and baseline tuning. Appendix C.1–C.5 specifies the data processing, batch sizes, optimizers, learning-rate schedules, training budgets, seeds, and evaluation protocols. Baseline hyperparameters were taken from the cited reference recipes, with the explicitly documented local adaptations. The 30M–200M QuEST settings follow the FOGZO recipes, while the 430M setting adopts the CAGE training recipe with our local QuEST quantizers. IDScore parameter choices are explicitly stated.

2. DeiT adapter. Appendix B.1 now specifies that the likelihood score is evaluated at the clipped, normalized LSQ coordinates, full-tensor score normalization, and clipping-based residual support. We also clarify that the corrected round backward can affect the learned-scale gradient, while retaining the scale parameterization, gradient-rescaling factor, and optimizer rule.

3. QuEST coordinate and threshold scaling. Appendix B.2 specifies the Hadamard transformation, row-wise RMS scaling, bit-dependent quantization step Δ\Delta, score coordinate ui=wh,i/max⁡(Δ,10−8)−1/2u_i=w_{h,i}/\max(\Delta,10^{-8})-1/2, and normalized thresholds. It also documents the temperature units, normalization and support choices, numerical stabilization, and mapping back to the weight coordinates.

4. Matched local comparisons. We completed the missing local baseline runs and replaced the author-log aggregates. The revised comparisons match training configurations, initialization seeds, data-order settings, and evaluation protocols. The 430M comparison now also includes three matched seeds. We have also clarified the language-training budget labels in Table 1 and explicitly identified its shared legacy PPL evaluator, defined in Appendix C.1, which differs from token-weighted corpus perplexity.

5. Additional efficiency measurements. Appendix D.2 and Table 13 report five-repeat benchmarks on 30M W4A4 and 100M W2A2 using two A800 GPUs, global batch size 64, and sequence length 512. Relative to QuEST, the paired step-time overheads are 3.955±0.521%3.955\pm0.521\% and 3.449±0.261%3.449\pm0.261\%, respectively, with peak allocated-memory overheads of 0.409% and 0.845%. Timing uncertainties are sample standard deviations across five trials. Each trial uses 300 warmup and 1,000 timed updates, and all trials are retained. These additional benchmarks use cached CPU inputs, while CPU pinning and host-to-device transfers remain timed. They are reported separately from Table 5, which includes data loading. For workload consistency, the W4A4 timing benchmark uses batch size 64, compared with batch size 32 in its training runs.

AuthorsOpenReview ↗

Response to Question 2: What trajectory-level evidence supports IDScore's mechanism?

Thank you for your insightful suggestion. We report the original training curves to show how performance differences develop over time. We have also added an instrumented AdamW case study that records transitions after actual optimizer updates and evaluates the local task-loss effects of the observed level changes. Section 4.4 summarizes these findings, and Appendix D.8 provides the protocols and results.

Performance along the original training trajectories. Figure 3 presents the original periodic validation-loss curves over three seeds, showing how performance differences develop during training, rather than only comparing final outcomes. Under Muon-AdamW, IDScore performs better at most evaluation points for each of the three seeds, reducing mean final PPL from 39.830 to 39.508. We also present the original validation curves for IDScore and IDScore-trust, whose final PPLs are 36.502±0.04336.502\pm0.043 and 39.965±0.07739.965\pm0.077, respectively.

Quantized transitions after actual AdamW updates. We conducted paired QuEST-30M W2A2 runs with seed 0, global batch size 64, and 91,552 updates. The runs share initialization and data order while retaining their own optimizer histories. Recording level changes after every update, we find that IDScore reduces the cumulative transitions by 23.62%, increases the target-hit rate among feasible transitions from 64.91% to 66.14%, and lowers final full-validation PPL from 37.436 to 36.499 (see the table below). Feasible transitions are actual level changes whose pre-update task gradient selects a neighboring target different from the current level; a target hit reaches that target. Absolute target-hit counts also decreased, so the observed change concerns the transition composition, rather than more target hits.

Instrumented AdamW trajectories on QuEST-30M W2A2, seed 0, global batch size 64.

MetricQuESTIDScore
Actual parameter-update norm (RMS)0.1150570.115399
Cumulative transitions (10910^9)17.999313.7481
Cumulative target hits (10910^9)10.23237.8769
Target-hit rate among feasible transitions64.91%66.14%
Final full-validation PPL ↓\downarrow37.43636.499

*The update norm is the temporal RMS of the ℓ2\ell_2 norm over quantized weights, measured after the AdamW step. Counts aggregate rank-coordinate observations; ranks are not independent seeds. These instrumented runs exhibit numerical drift relative to native execution and are reported separately from the original experiments.

Local task benefits of the observed level changes. At three prespecified updates (9,000, 45,000, and 82,000), we insert the observed post-update weight quantization levels, while holding each method's other pre-update parameters, quantization scales, and validation data fixed and using the same validation inputs. IDScore's joint level changes reduced loss by 0.001678, 0.004959, and 0.005746, respectively. These within-state interventions directly test the local task-loss effects of the observed changes at each method's own training state. Appendix D.8 also reports an exploratory checkpoint-replay analysis separating target-hitting transitions from the remaining transitions and evaluating their group-level task-loss effects (Table 21).

Outside-mask effects over training. At the same three points, we replace only the levels outside the pre-update QuEST trust mask with their observed post-update levels, retaining all other levels and settings at their pre-update values. Relative to retaining all old levels, this outside-only intervention yields loss reductions of −0.000679-0.000679, +0.001511+0.001511, and +0.005704+0.005704: it increases loss at the early point but reduces it at the middle and late points. In addition, the outside-mask residual squared-energy fractions over updates 1–30,000, 30,001–60,000, and 60,001–91,552 are 3.533%, 2.745%, and 3.098%, measured before distributed gradient aggregation and clipping. These measurements complement the approximately 4.35% final-checkpoint statistic with stage-wise correction-signal measurements and local loss effects of outside-mask level changes.

Together, these results provide trajectory-level empirical evidence connecting observed quantization-level changes under AdamW with local task-loss benefits. In this case study, lower final PPL accompanies fewer transitions, a higher conditional target-hit rate, and beneficial joint level changes at all three prespecified diagnostic points.

Thank you for addressing both of my questions. Replacing the author-log comparisons with matched local runs, including three matched seeds at 430M, resolves my main concern about comparability. The added details on the DeiT and QuEST adapters clarify how to reproduce the correction, and the timing and memory measurements give a clearer picture of its cost at a second bit width and a larger model size.

I particularly appreciate the AdamW experiment. In the instrumented run, applying only the observed changes outside the trust mask increased validation loss at the early checkpoint and reduced it at the two later checkpoints. This helps explain how those coordinates can matter even though they account for a small fraction of the residual energy. The single-seed experiment leaves open how much these local effects contribute to the final perplexity gap, and how broadly the mechanism applies under Muon. My rating remains 3 (weak accept).

AuthorsOpenReview ↗

Follow-up: Additional multi-seed diagnostics under AdamW and Muon+AdamW

Thank you for your thoughtful follow-up. To further address the remaining questions, we added paired 20-update diagnostics under both AdamW and Muon+AdamW, covering three seeds and checkpoints from both QuEST and IDScore trajectories.

At each checkpoint, the correction-off and IDScore branches start from identical model parameters, optimizer states, RNG states, and subsequent training batches, and then evolve independently for 20 updates. We use a global batch size of 64 on two GPUs, fixed learning rates, and four common validation batches. These are local diagnostics near the original training endpoints, not full-training reruns.

We report the changes introduced by IDScore relative to the correction-off branch. Transition counts are reported as relative changes, while the other quantities are reported as percentage-point differences.

OptimizerCheckpoint sourceTransition count change (%)Feasible-coordinate transition rate (pp)Conditional target-hit rate (pp)Feasible-coordinate target-hit rate (pp)
AdamWQuEST−0.180−0.000788+0.018128−0.000413
AdamWIDScore−0.416−0.001113+0.012603−0.000672
Muon+AdamWQuEST+3.015+0.014008+0.212757+0.010571
Muon+AdamWIDScore−1.357−0.004767−0.016073−0.003393

These results suggest that IDScore changes the composition of actual quantized transitions, rather than simply increasing their frequency; no single transition statistic alone explains all observed task-loss benefits.

We further evaluate the loss after applying the actual optimizer updates, using the same fixed validation batches. Positive improvement denotes lower loss with IDScore.

OptimizerCheckpoint sourceSeedCorrection-off lossIDScore lossLoss improvement
AdamWQuEST03.6654773.663707+0.001770
13.6697353.664483+0.005252
23.6656823.661581+0.004101
Mean ± SD3.666965 ± 0.0019613.663257 ± 0.001227+0.003708 ± 0.001448
AdamWIDScore03.6434483.638211+0.005237
13.6493253.647139+0.002186
23.6441983.643098+0.001100
Mean ± SD3.645657 ± 0.0026123.642816 ± 0.003650+0.002841 ± 0.001751
Muon+AdamWQuEST03.7378073.733509+0.004299
13.7336703.731996+0.001674
23.7304873.730443+0.000044
Mean ± SD3.733988 ± 0.0029973.731983 ± 0.001252+0.002006 ± 0.001753
Muon+AdamWIDScore03.7258723.723033+0.002839
13.7298893.727347+0.002542
23.7277163.724019+0.003697
Mean ± SD3.727826 ± 0.0016423.724800 ± 0.001846+0.003026 ± 0.000490

IDScore yields positive local loss improvements in all 12 paired seed comparisons across the four optimizer–checkpoint combinations. These results extend the local task-loss evidence beyond the original single-seed AdamW case study and show that the correction remains locally beneficial under both AdamW and Muon+AdamW.

Together, these diagnostics provide additional evidence connecting actual optimizer updates, quantized transition patterns, and local task-loss improvements. They establish reproducible local effects across the tested seeds and optimizer routes, while not claiming to quantitatively decompose the final perplexity gap or establish general optimizer convergence guarantees.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.