acceptodds
Under review as a conference paper at ICLR 2027

HeRo-Q: A General Framework for Stable Low Bit Quantization via Hessian Conditioning

Abstract

Post-Training Quantization (PTQ), a mainstream model compression technique, often leads to the paradoxical “low-error, high-loss” phenomenon because it focuses solely on minimizing quantization error. The root cause lies in the Hessian matrix of the LLM loss landscape: a few high-curvature directions are extremely sensitive to perturbations. To address this, we propose the Hessian Robust Quantization (HeRo-Q) algorithm, which applies a lightweight, learnable rotation-compression matrix to the weight space prior to quantization. This joint framework reshapes the loss landscape by reducing the largest Hessian eigenvalue, thereby significantly enhancing robustness to quantization noise. HeRo-Q requires no architectural modifications, incurs negligible computational overhead, and integrates seamlessly into existing PTQ pipelines. Experiments on Llama and Qwen models show that HeRo-Q consistently outperforms state-of-the-art methods—including GPTQ, AWQ, and SpinQuant—not only achieving superior performance under standard W4A8 settings, but also excelling in the highly challenging W3A16 ultra-low-bit regime, where it boosts GSM8K accuracy on Llama-3-8B to 70.15% and effectively avoids the logical collapse commonly seen in aggressive quantization.

Primary area:
foundation or frontier models, including LLMs
Keywords:
PTQ.+Hessian

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_KN6tOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

HeRo-Q is a post-training quantization framework that combines Hessian-proxy-based diagonal conditioning with orthogonal rotation optimization. The conditioning targets effective curvature, while rotation is updated by minimizing output reconstruction error induced by weight quantization. The paper provides a local second-order analysis and evaluates weight-only and weight-activation quantization across dense LLMs, MoE models, ViTs, and a VLM. Compatibility with existing quantization algorithms (GPTQ, AWQ, and RTN) is evaluated, along with composition with QuIP. The experiments broadly cover PPL, downstream accuracy, calibration cost, and decoding throughput.

Critical strengths
  1. Evaluation broadly covers dense LLMs, MoE models, ViTs, and a VLM under multiple quantization settings. The results generally show improvements over the reported baselines across these model families (Tables 1–3 and additional experiments in Appendix B).

  2. The framework is compatible with different base quantizers, including GPTQ, AWQ, and RTN (Table 7). Combining it with QuIP also improves over both standalone methods in the reported settings (Table 21).

  3. The paper includes component ablations and controlled comparisons rather than relying only on comparisons between complete methods. Tables 5 and 18 provide evidence for the conditioning’s contribution within the evaluated settings, although their scope remains limited.

  4. Calibration time, peak calibration memory, and measured decoding throughput are reported (Tables 4 and 16). The deployment description also distinguishes transformations absorbed offline from those remaining online (Table 15), making the implementation assumptions more explicit.

  5. The Hessian-based analysis provides a useful distinction between diagonal conditioning and orthogonal rotation. This offers a principled motivation for combining the two operations, with a spectral-compression result under the stated diagonal-proxy assumptions (Theorem 2).

Critical weaknesses
  1. Though the Hessian-based analysis and conditioning method are valuable, HeRo-Q fundamentally combines channel-wise scaling with learned rotations, leaving its methodological novelty limited. Furthermore, activation quantization is excluded during optimization (Appendix C.4), while the component ablation uses W4A16 (Table 5), so the conditioning's additional contribution under W4A4/W4A8 remains unclear.

  2. The comparisons in Table 1 do not cover recent transformation-based methods, such as OSTQuant [1], DartQuant [2], and OptRot [3], leaving competitiveness against these methods unclear.

  3. In Table 16(b), HeRo-Q's throughput is close to the transformation-free baseline, despite retaining online R4 at the SwiGLU–down interface (Table 15). Further clarification of the exact inference setup, including the W4A16 kernel used for each method and a timing ablation isolating the online transform, is needed to establish its overhead and the competitiveness of the inference baseline.

  4. Gains over QTIP in Table 20 are small (0.48 pp on MMLU and 0.31 pp on GSM8K). Without uncertainty estimates or repeated calibration and optimization runs, it is unclear whether these gains are reliable and consistent.

  5. Although GSM8K (Table 1), HumanEval, and MathQA (Table 9) results are reported, HeRo-Q's robustness on recent reasoning models and challenging long-generation tasks is not evaluated. This leaves its applicability to these increasingly important workloads unclear.

Questions for authors

I have a few questions for the authors, roughly aligned to the weaknesses:

  1. What does the proposed Hessian-based conditioning contribute beyond existing scaling–rotation approaches? Please extend the component ablation in Table 5 to W4A4/W4A8, isolating conditioning-only, rotation-only, and combined effects, and broaden the conditioning-rule comparison in Table 18 beyond Llama-3.1-8B W3A16.

  2. Can the authors compare against recent transformation-based methods, such as OSTQuant [1], DartQuant [2], and OptRot [3], using matched checkpoints, quantization settings, calibration budgets, and evaluation protocols?

  3. Which W4A16 kernels and library versions are used for each method in Table 16(b)? Please provide a matched comparison against an optimized transformation-free baseline and a timing ablation isolating the online R4 transform listed in Table 15.

  4. Are the small gains over QTIP in Table 20 consistent across calibration samples and optimization seeds? Please provide repeated-run results and appropriate uncertainty estimates for the downstream accuracy differences.

  5. Do the gains reported in Tables 1 and 9 hold on recent reasoning models and challenging long-generation tasks, such as harder math (AIME 2025), science (GPQA Diamond), and coding benchmarks (LiveCodeBench v5)?

  6. The code URL provided in Appendix C.1 is expired. Can the authors provide an accessible anonymized repository with the implementation and scripts needed to reproduce the main results?

7 comments
AuthorsOpenReview ↗

Rebuttal by Authors

Q1

Regarding methodological novelty, please refer to our responses to Reviewer CoaR's Weaknesses 2 and Question 1. Here, we focus on the requested component ablations and expanded conditioning comparisons.

We extend the component ablation in the original Table 5 to W4A4 and W4A8 settings across three models. The following table reports WikiText-2 perplexity.

ModelPrecisionGPTQ onlyConditioning onlyRotation onlySmoothQuant+SpinQuantHeRo-Q
Llama-3.1-8BW4A41470±22146.6±5.814.66±0.5612.48±0.7310.29±0.10
Llama-3.1-8BW4A88.81±0.178.06±0.147.73±0.117.43±0.096.96±0.03
Llama-3.2-1BW4A42060±364135.0±10.624.12±1.0520.63±1.0618.13±0.90
Llama-3.2-1BW4A816.18±0.3414.98±0.2014.35±0.2813.60±0.3012.67±0.04
Llama-3.2-3BW4A41640±34795.8±6218.37±0.8915.96±0.8513.18±0.54
Llama-3.2-3BW4A89.82±0.079.80±0.099.77±0.029.74±0.049.73±0.07

Under W4A4, HeRo-Q reduces perplexity by 24.8%–29.8% compared with rotation only and by 12.1%–17.5% compared with SmoothQuant+SpinQuant.

We also extend the conditioning-rule comparison in the original Table 18 to additional models and precisions. Each cell below reports WikiText-2 PPL↓ / C4 PPL↓ / MMLU↑ / GSM8K↑, in that order.

Model / PrecisionQuIP scaling + matched rotationHeRo-Q
Llama-3.1-8B / W3A168.40±0.19 / 14.33±0.22 / 63.02±0.34 / 68.29±0.638.02±0.12 / 13.98±0.08 / 63.80±0.28 / 70.15±0.42
Llama-3.1-8B / W4A410.75±0.20 / 15.38±0.46 / 44.65±0.27 / 41.17±0.5710.29±0.10 / 14.84±0.09 / 45.31±0.24 / 42.33±0.12
Llama-3.1-8B / W4A87.11±0.09 / 12.98±0.11 / 65.73±0.18 / 74.33±0.436.96±0.03 / 12.94±0.07 / 66.31±0.16 / 75.26±0.37
Llama-3.2-3B / W4A413.90±0.68 / 22.61±0.81 / 39.44±0.43 / 35.16±1.1713.18±0.54 / 20.14±0.15 / 41.10±0.18 / 37.15±0.47
Llama-3.2-1B / W3A1615.43±0.23 / 27.00±0.19 / 41.50±0.14 / 25.80±0.2714.05±0.09 / 25.80±0.16 / 42.50±0.16 / 29.10±0.23
Qwen2.5-7B / W4A86.81±0.05 / 13.47±0.06 / 70.61±0.11 / 85.31±0.186.78±0.03 / 13.43±0.06 / 71.02±0.18 / 86.12±0.29

Q2

We have added comparisons with OSTQuant, DartQuant, and OptRot. Each cell below reports WikiText-2 PPL↓ / C4 PPL↓ / MMLU↑ / GSM8K↑, in that order.

The W4A16 results are as follows:

MethodLlama-3.1-8BQwen2.5-7B
FlatQuant6.46±0.03 / 12.32±0.01 / 66.93±0.08 / 76.32±0.136.71±0.02 / 13.35±0.03 / 71.17±0.07 / 87.22±0.19
OSTQuant6.43±0.01 / 12.37±0.02 / 66.78±0.09 / 76.28±0.176.73±0.02 / 13.31±0.03 / 71.27±0.16 / 87.02±0.13
DartQuant6.55±0.04 / 12.41±0.03 / 66.60±0.17 / 75.71±0.196.76±0.04 / 13.43±0.05 / 70.97±0.27 / 86.77±0.30
OptRot6.52±0.02 / 12.41±0.02 / 66.50±0.11 / 76.04±0.156.76±0.05 / 13.38±0.03 / 71.06±0.22 / 87.09±0.35
HeRo-Q6.33±0.01 / 12.30±0.01 / 67.10±0.12 / 76.85±0.176.68±0.03 / 13.28±0.03 / 71.35±0.12 / 87.50±0.31

The W3A16 results are as follows:

MethodLlama-3.1-8BLlama-3.2-1B
FlatQuant8.77±0.10 / 14.55±0.21 / 62.56±0.16 / 68.38±0.1916.11±0.22 / 27.47±0.38 / 40.18±0.23 / 24.97±0.29
OSTQuant8.69±0.08 / 14.62±0.16 / 62.41±0.30 / 68.68±0.4616.26±0.19 / 27.24±0.36 / 40.39±0.18 / 24.68±0.37
DartQuant9.02±0.14 / 14.82±0.24 / 61.96±0.32 / 67.57±0.7716.58±0.24 / 27.99±0.23 / 39.70±0.22 / 24.14±0.28
OptRot8.93±0.21 / 14.87±0.15 / 62.13±0.28 / 67.89±0.2916.45±0.40 / 27.81±0.22 / 39.91±0.30 / 24.48±0.35
HeRo-Q8.02±0.12 / 13.98±0.08 / 63.80±0.28 / 70.15±0.4214.05±0.09 / 25.80±0.16 / 42.50±0.16 / 29.10±0.23

HeRo-Q achieves the best mean performance on all four metrics in the comparisons above. Under W3A16, compared with the best result among the listed baselines for each metric, HeRo-Q reduces WikiText-2 perplexity by 7.7% and 12.8% and improves GSM8K accuracy by 1.47 and 4.13 percentage points on Llama-3.1-8B and Llama-3.2-1B, respectively.

AuthorsOpenReview ↗

Rebuttal by Authors

Q3

All W4A8 experiments use the Marlin kernel in vLLM v0.17.1, with symmetric INT4 weight quantization at group size 128 and dynamic per-token symmetric INT8 activation quantization. All methods share the same GEMM kernel, and end-to-end latency includes activation quantization and the required online transformations. Timing excludes prefill and includes the remaining online transformations. After three warm-up runs, we synchronize the GPU before and after the timed region.

To isolate the cost of online rotation, we add a timing ablation at the SwiGLU–down interface, comparing “pre-rotated input + down GEMM” with “online R4 + down GEMM”:

ModelPre-rotated input + down GEMM (ms/step)Online R4 + down GEMM (ms/step)Increment (ms/step)Fraction of total decoding time
Llama-3.2-1B1.160±0.0271.420±0.0290.260±0.0071.97%
Llama-3.2-3B2.080±0.0462.520±0.0480.440±0.0162.31%
Llama-3.1-8B4.320±0.0944.890±0.1240.570±0.0321.61%

The pre-rotated-input path is used to isolate the timing cost of the online transformation; it does not imply that R4 can simply be removed from the full deployment. This rotation is still executed in actual deployment.

Under the tested configurations, the online R4 path adds 0.260–0.570 ms/step, accounting for 1.61%–2.31% of total decoding time. This is consistent with the small overall throughput differences reported in the paper. Our efficiency conclusion is that diagonal conditioning at compatible interfaces can be absorbed offline, avoiding a separate online scaling operator; the remaining online rotations still incur costs, which are included in the decoding measurements.

Q4

The small accuracy differences in the original Table 20 require support from repeated experiments. Under the Llama-3.2-1B W3A16 setting, the results of 15 paired evaluation runs are as follows:

TaskQTIPHeRo-QPaired difference (mean±SD, percentage points)Runs where HeRo-Q leads
MMLU42.02±0.1542.50±0.16+0.48±0.1315/15
GSM8K28.79±0.1429.10±0.23+0.31±0.1315/15

HeRo-Q achieves higher accuracy in all 15 paired evaluation runs on both tasks. Its average gains on MMLU and GSM8K are 0.48 and 0.31 percentage points, respectively.

Q5

We extend the evaluation to Qwen3-8B and Qwen3-14B and add AIME, GPQA Diamond, and LiveCodeBench v5. The compared methods include FP16, GPTQ, AWQ, FlatQuant, and HeRo-Q.

Each cell below reports scores on AIME / GPQA Diamond / LiveCodeBench v5, in that order.

MethodQwen3-8BQwen3-14B
FP1667.3 / 62.0 / 57.570.4 / 64.0 / 63.5
GPTQ64.5 / 59.9 / 53.569.1 / 63.5 / 61.2
AWQ63.1 / 59.0 / 55.068.1 / 62.1 / 62.1
FlatQuant65.5 / 60.7 / 54.469.5 / 63.1 / 61.7
HeRo-Q66.4 / 61.7 / 55.670.3 / 63.9 / 62.8

HeRo-Q achieves the highest scores among the listed quantization methods on all three tasks for both models. Compared with the strongest quantization baseline on each task, the gains are 0.9, 1.0, and 0.6 percentage points on Qwen3-8B and 0.8, 0.4, and 0.7 percentage points on Qwen3-14B, respectively.

Q6

We have updated Appendix C.1.

AuthorsOpenReview ↗

Follow up on rebuttal

Dear Reviewer KN6t,

We apologize for disturbing you. As the discussion deadline approaches, we are writing to follow up to ensure that our previous response and revisions have fully addressed your concerns.

We believe that these improvements have significantly strengthened the quality of the paper. We would be extremely grateful if you could take a moment to review our reply and re-evaluate our work. We are joyous to engage in further discussion if needed.

Best regards, The Authors

Reviewer_YeU5OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper addresses the reduction of storage requirements for neural networks by quantisation of the parameters. The problem is that direct quantisation is inefficient because some parameters are more important than others, and it is hard to deduce their relative importance, making it difficult to determine which parameters to quantise accurately. For that problem, this paper proposes an improvement to Hessian-aware methods, where the parameter space is transformed before quantisation. The rotation part of that transform is then optimised for best accuracy. They furthermore give a theoretical motivation for the efficiency of the method, which is also validated by experiments.

Critical strengths
  • The proposed method reduces the numerical range of parameters such that fixed-bitrate quantisation is more efficient.
  • Experiments clearly verify the advantages.
  • Extensive theoretical derivation provides a clear motivation for the benefits of the method.
Critical weaknesses

Major issues:

  • The paper is clearly aimed at advanced experts, making it a laborious read for others. This is not a language issue; the writing is very clear, but it just assumes a lot of prior knowledge.

Minor issues:

  • Figure 1 is not accessible to non-experts, as there is very little explanation of what the figure means. For 1c in particular, the caption and the figure don't connect, and 1b is also difficult to understand. I think I mostly understand 1a.
  • Figure 2 looks like a standard AI-generated visualisation. It's an effective visualisation, so I don't really mind it, but I've just seen the style so very many times before.
Questions for authors

None

4 comments
AuthorsOpenReview ↗

Rebuttal by Authors

Weakness 1

Thank you for pointing this out. We will revise the appendix to provide more background for readers who are new to post-training quantization. We will explain how low-bit rounding perturbs the weights and use Figure 1 to show why perturbations of the same magnitude can affect the loss differently. We will then explain how the Hessian captures this directional sensitivity before presenting the formal analysis.

Weakness 2

Figure 1(b) shows how weight perturbations of the same magnitude can cause different increases in loss depending on their direction. Both curves start at the same full-precision weight vector. The horizontal plane represents the direction and magnitude of the perturbation, and the vertical axis shows the increase in local loss relative to the starting point. Under the second-order approximation, a perturbation of magnitude rr along a direction with curvature λ\lambda increases the loss by approximately 12λr2\tfrac12\lambda r^2. For the same perturbation magnitude, the high-curvature direction (blue) therefore produces a larger loss increase than the low-curvature direction (red).

Weakness 3

Figure 1(c) shows how HeRo-Q changes the distribution of these curvatures. The horizontal axis shows the Hessian eigenvalues, each of which gives the local curvature along its corresponding eigendirection. The vertical axis shows their density. The red curve shows the distribution in the original coordinates, and the blue curve shows it after HeRo-Q’s diagonal rescaling. The shorter right tail of the blue curve illustrates the reduction in large eigenvalues. The largest eigenvalue gives the curvature along the sharpest direction in Figure 1(b). Reducing it lowers the worst-case increase in local loss for a rounding error of a given norm in the quantization coordinates.

AuthorsOpenReview ↗

Follow up on rebuttal

Dear Reviewer YeU5,

We apologize for disturbing you. As the discussion deadline approaches, we are writing to follow up to ensure that our previous response and revisions have fully addressed your concerns.

We believe that these improvements have significantly strengthened the quality of the paper. We would be extremely grateful if you could take a moment to review our reply and re-evaluate our work. We are joyous to engage in further discussion if needed.

Best regards, The Authors

AuthorsOpenReview ↗

Rebuttal by Authors

Dear Reviewer YeU5,

We apologize for disturbing you. As the discussion deadline approaches, we are writing to follow up to ensure that our previous response and revisions have fully addressed your concerns.

We believe that these improvements have significantly strengthened the quality of the paper. We would be extremely grateful if you could take a moment to review our reply and re-evaluate our work. We are joyous to engage in further discussion if needed.

Best regards, The Authors

Reviewer_YeU5OpenReview ↗

Concerns have been adressed

All my critical concerns have been adressed.

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

The authors propose HeRo-Q, a post-training quantization method that combines (i) Hessian-derived diagonal conditioning to reduce the dominant effective curvature and (ii) a optimized orthogonal rotation to change the finite-bit residual and its alignment in the conditioned coordinates. The authors present theoretical results like the transformed-coordinate loss bound and a strict spectral-compression result. Empirically, HeRo-Q works for a wide range of language, vision, multimodal, and MoE models.

Critical strengths
  1. Empirical improvement over baselines Table 5 provides a useful comparison where HeRo-Q beats SmoothQuant+SpinQuant.
Critical weaknesses
  1. Limited theoretical contribution Theorem 1 is a standard Rayleigh-quotient bound, and doesn't make it clear that decreasing λmax⁡(Hq)\lambda_{\max}(H_q) alone does not establish an improvement in quantization error: a simple rescaling T↦T/cT\mapsto T/c decreases λmax⁡(Hq)\lambda_{\max}(H_q) by c2c^2 while increasing ∥ϵ∥22\|\epsilon\|_2^2 by c2c^2, so nothing really changes. QuIP’s Theorems 1 and 7 provide finer-grained guarantees.
  2. Limited novelty QuIP already has diagonal rescaling, although a heuristic one instead of a grid search, and the comparison does not clearly separate the benefit of the proposed scaling rule from additional optimization/search effort. Moreover, SpinQuant already uses Cayley SGD. The SmoothQuant+SpinQuant baseline is not clearly specified.
Questions for authors
  1. What exactly is your SmoothQuant+SpinQuant baseline, and how does it differ from HeRo-Q's methodology?
  2. Can you do an ablation against the "diagonal rescaling" in QuIP Algorithm 1 line 4? I'm curious whether it's beneficial to replace the Dα=(diag⁡(H)/diag⁡(WTW))α4D_\alpha = \left(\operatorname{diag}(H)/\operatorname{diag}(W^T W)\right)^\frac{\alpha}{4} in QuIP with Dα=diag⁡(H)α2D_\alpha = \operatorname{diag}(H)^\frac{\alpha}{2} in your paper.
5 comments
AuthorsOpenReview ↗

Rebuttal by Authors

Weakness 1

A decrease in the transformed Hessian’s largest eigenvalue alone does not guarantee lower quantization error. Under uniform scaling with a correspondingly rescaled quantization range, the reduction in curvature is canceled by the increase in residual scale. HeRo-Q applies nonuniform, curvature-dependent conditioning that changes relative channel scales. We explicitly account for the transformed weight scale below.

Under the positive diagonal proxy assumption of Theorem 2, fix Hproxy=diag⁡(h1,…,hd)≻0H_{\mathrm{proxy}}=\operatorname{diag}(h_1,\ldots,h_d)\succ0, ww, and bb, with 0≤α≤10\le\alpha\le1. Let QbQ_b use 2b2^b equally spaced levels over the symmetric interval determined by the maximum absolute value. For

z=Tα−1w=R⊤Dαw,ϵ=Qb(z)−z,z=T_\alpha^{-1}w=R^\top D_\alpha w,\qquad \epsilon=Q_b(z)-z,

nearest-neighbor rounding gives

∥ϵ∥22≤d(2b−1)2∥z∥∞2≤d(2b−1)2∥z∥22.\Vert\epsilon\Vert_2^2\le\frac{d}{(2^b-1)^2}\Vert z\Vert_\infty^2\le\frac{d}{(2^b-1)^2}\Vert z\Vert_2^2.

Since RR is orthogonal and Dα=diag⁡(hiα/2)D_\alpha=\operatorname{diag}(h_i^{\alpha/2}),

∥z∥22=∥R⊤Dαw∥22=∑i=1dhiαwi2.\Vert z\Vert_2^2=\Vert R^\top D_\alpha w\Vert_2^2=\sum_{i=1}^{d}h_i^\alpha w_i^2.

Applying Theorem 1 to the proxy loss and using Eq. (23),

ΔLproxy≤12λmax⁡(Hq,proxy(α))∥ϵ∥22.\Delta L_{\mathrm{proxy}}\le\frac12\lambda_{\max}(H_{q,\mathrm{proxy}}(\alpha))\Vert\epsilon\Vert_2^2.

≤dhmax⁡1−α2(2b−1)2∑i=1dhiαwi2.\le\frac{d h_{\max}^{1-\alpha}}{2(2^b-1)^2}\sum_{i=1}^{d}h_i^\alpha w_i^2.

≤dhmax⁡2(2b−1)2∑i=1d(hi/hmax⁡)αwi2\le\frac{d h_{\max}}{2(2^b-1)^2}\sum_{i=1}^{d}(h_i/h_{\max})^\alpha w_i^2 ≤dhmax⁡2(2b−1)2∥w∥22\le\frac{d h_{\max}}{2(2^b-1)^2}\Vert w\Vert_2^2 Because 0<hi/hmax⁡≤10 < h_i/h_{\max}\le1, this joint upper bound is nonincreasing in α\alpha, and strictly decreasing if some ii satisfies hi<hmax⁡h_i < h_{\max} and wi≠0w_i\neq0. Thus, nonuniform conditioning can improve the bound even after accounting for the transformed weight scale, without assuming a fixed residual scale. Unlike uniform scaling, it leaves the channel-dependent factor (hi/hmax⁡)α(h_i/h_{\max})^\alpha.

However, this bound uses a norm relaxation; its monotonic decrease does not imply monotonically decreasing actual quantization loss. Although α=1\alpha=1 makes the effective Hessian isotropic under these assumptions, requantization also changes the residual’s magnitude and direction. We therefore search over α∈[0,1]\alpha\in[0,1] using post-quantization error rather than fixing α=1\alpha=1 or selecting solely by curvature reduction.

Weakness 2

Both methods use curvature information, but differ in their motivation, conditioning rules, and rotation selection. QuIP uses random orthogonal transformations to improve incoherence, distributing weight magnitudes and preventing Hessian eigenvectors from concentrating on a few coordinate axes. HeRo-Q focuses on how effective curvature amplifies actual quantization residuals. As illustrated in Figure 1(b), equal-norm perturbations can incur different local losses along different directions.

Let hih_i denote a Hessian-proxy diagonal entry and gig_i the corresponding channel’s weight energy:

(DQuIP)ii=(higi)1/4,(Dα)ii=hiα/2.(D_{\mathrm{QuIP}})_{ii}=\left(\frac{h_i}{g_i}\right)^{1/4},\qquad (D_\alpha)_{ii}=h_i^{\alpha/2}.

QuIP balances Hessian and weight scales through a closed-form rule. HeRo-Q constructs a tunable Hessian-power conditioning family, whose spectral effect is characterized in Theorem 2, and selects α\alpha through reconstruction error after requantization. Thus, weight scale and its induced quantization distortion still influence candidate selection even though weight energy does not explicitly appear in DαD_\alpha.

Hessian incoherence and residual–curvature alignment also describe different objects. The former concerns eigenvector concentration relative to coordinate axes; our alignment coefficient,

a(ϵ,Hq)=ϵ⊤Hqϵλmax⁡(Hq)∥ϵ∥22,a(\epsilon,H_q)=\frac{\epsilon^\top H_q\epsilon}{\lambda_{\max}(H_q)\Vert\epsilon\Vert_2^2},

measures the normalized curvature experienced by the actual residual. For the same HqH_q, equal-norm residuals along high- and low-curvature directions have identical Hessian incoherence but different curvature-weighted losses.

Accordingly, QuIP constructs random rotations with favorable incoherence properties, while HeRo-Q optimizes or selects admissible rotations through requantization and reconstruction feedback after conditioning. Alignment is an explanatory diagnostic; the optimization objective remains the full reconstruction error. We acknowledge that SpinQuant already uses Cayley SGD. HeRo-Q’s distinction lies in its conditioning construction and transformation selection through quantization-reconstruction feedback, rather than a new Cayley optimizer.

Table 19 in Appendix D.3 supports this analysis using the full Gram proxy on Llama-3.1-8B W4A16. Relative to SmoothQuant+SpinQuant, HeRo-Q reduces the relative residual norm from 0.88 to 0.81, relative curvature-weighted error from 0.46 to 0.29, and alignment score from 0.21 to 0.13. The improvements concern both residual magnitude and its curvature-weighted cost.

AuthorsOpenReview ↗

Rebuttal by Authors

Q1

Our SmoothQuant+SpinQuant baseline applies SmoothQuant’s channel-scaling rule under W4A16. It first computes scaling factors from calibration activation and weight magnitudes and performs equivalent reparameterization, then reoptimizes rotations on the scaled model, and finally quantizes weights to 4 bits using GPTQ while retaining FP16 activations. Thus, the baseline includes rotation readaptation after scaling.

In Table 5, its WikiText-2 perplexities on Llama-3.2-1B and Llama-3.1-8B are 12.62 and 6.75, compared with 11.74 and 6.33 for HeRo-Q.

Both methods combine scaling and rotation, but their design motivations differ. SmoothQuant+SpinQuant applies magnitude-driven scaling followed by rotation optimization. HeRo-Q jointly considers effective curvature, residual magnitude, and residual–curvature alignment. Conditioning adjusts effective curvature; rotation preserves the conditioned spectrum while changing the requantized residual’s magnitude and direction to reduce its curvature-weighted cost.

Q2

Table 18 in Appendix D.2 compares QuIP scaling with HeRo-Q conditioning and strength selection on Llama-3.1-8B W3A16, sharing the rotation-optimization procedure and GPTQ backend. HeRo-Q reduces WikiText-2/C4 perplexity from 8.40/14.33 to 8.02/13.98, while increasing MMLU/GSM8K accuracy from 63.02/68.29% to 63.80/70.15%. This supports the conditioning scheme, including strength selection, under the shared pipeline.

We further compare five configurations within QuIP: no diagonal scaling, original QuIP scaling, fixed-strength HeRo-Q scaling, and versions allowing each scaling family to search its strength. Here, hi=(Hproxy)iih_i=(H_{\mathrm{proxy}})_{ii} and gi=(W⊤W)iig_i=(W^\top W)_{ii}.

Entries are WikiText-2 PPL↓ / C4 PPL↓ / MMLU↑ / GSM8K↑, reported as mean±SD; accuracies are percentages.

Diagonal entry DiiD_{ii}Llama-3.2-1BLlama-3.1-8B
1117.04±0.42 / 28.38±0.36 / 39.94±0.41 / 23.62±0.699.10±0.12 / 14.97±0.29 / 62.03±0.20 / 67.22±0.26
(hi/gi)1/4(h_i/g_i)^{1/4}15.50±0.35 / 27.00±0.44 / 41.50±0.34 / 26.00±0.558.50±0.16 / 14.40±0.19 / 63.20±0.28 / 69.00±0.51
hi1/4h_i^{1/4}15.63±0.18 / 26.86±0.24 / 41.42±0.30 / 26.16±0.578.45±0.09 / 14.47±0.19 / 63.28±0.25 / 68.88±0.63
(hi/gi)α/2(h_i/g_i)^{\alpha/2}, searched α\alpha15.30±0.40 / 26.79±0.34 / 41.72±0.30 / 26.49±0.728.42±0.15 / 14.34±0.12 / 63.33±0.22 / 69.22±0.26
hiα/2h_i^{\alpha/2}, searched α\alpha14.88±0.22 / 26.48±0.17 / 42.10±0.20 / 27.18±0.378.24±0.12 / 14.20±0.09 / 63.63±0.28 / 69.73±0.44

Compared with searched QuIP-style scaling, HeRo-Q scaling reduces WikiText-2/C4 perplexity from 15.30/26.79 to 14.88/26.48 on Llama-3.2-1B, with MMLU/GSM8K gains of 0.38/0.69 percentage points. On Llama-3.1-8B, perplexities decrease from 8.42/14.34 to 8.24/14.20, with accuracy gains of 0.30/0.51 percentage points.

AuthorsOpenReview ↗

Follow up on rebuttal

Dear Reviewer CoaR,

We apologize for disturbing you. As the discussion deadline approaches, we are writing to follow up to ensure that our previous response and revisions have fully addressed your concerns.

We believe that these improvements have significantly strengthened the quality of the paper. We would be extremely grateful if you could take a moment to review our reply and re-evaluate our work. We are joyous to engage in further discussion if needed.

Best regards, The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.