acceptodds
Under review as a conference paper at ICLR 2027

MirrorQuant: Learnable Householder Rotations for Low-Bit LLM Quantization

Abstract

Orthogonal rotations have become an effective technique for low-bit LLM quantization by expressing hidden representations in more quantization-friendly bases. However, existing approaches predominantly parameterize learned rotations as dense matrices, incurring quadratic learned state and costly optimization. Their deployment efficiency largely depends on foldability: at non-foldable interfaces, dense rotations must remain on the inference path, with both quadratic storage and computation of the rotation dimension. To overcome these limitations, we propose MirrorQuant, a unified framework for efficient and learnable rotations across Transformer interfaces. MirrorQuant parameterizes learned rotations using short chains of Householder reflections, a construction that preserves exact orthogonality, reduces the learned state to linear in the rotation dimension for fixed chain depth, and enables learnable log-linear online execution without materializing dense rotation matrices. We further introduce projection-guided learning, which complements representation reconstruction by penalizing quantization distortion after the immediate downstream linear projection. Experiments across 23 matched settings spanning model families, scales, and low-bit configurations demonstrate accuracy and efficiency gains. MirrorQuant improves 0-shot accuracy over DartQuant in every setting, with gains up to 1.81%, while its online rotation reduces prefill latency by 29.6% relative to a dense implementation. Source code can be found in the supplementary materials.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Large Language Models, Post-Training Quantization, Householder Rotations, Efficient Inference

Reviews

Among:
Shape:
1234this paper, Top 17%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.75 over 4 reviews, above 83% of the others.
Reviewer_4Jc8OpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes MirrorQuant, a PTQ framework that parameterizes learnable rotations as a fixed orthogonal base followed by a short chain of Householder reflections. This reduces learnable state from quadratic to linear complexity while preserving exact orthogonality. For the non-foldable MLP rotation R 4 ​ , MirrorQuant combines an FHT with learned reflections to enable efficient online execution. The paper also introduces projection-guided learning and progressive calibration. Experiments report consistent improvements over DartQuant across multiple model families and low-bit settings.

Critical strengths
  1. The paper addresses an deployment bottleneck of dense learned rotations.
  2. The Householder parameterization is simple, theoretically well motivated, and exactly orthogonal.
  3. The distinction between foldable and online rotations is clearly presented.
  4. Projection-guided learning has a reasonable motivation and is supported by ablations.
Critical weaknesses
  1. Missing comparison with Kronecker-structured online transformations. A FlatQuant-style Kronecker transform is a direct alternative for R4 ​ . It may offer greater expressiveness through learnable affine transformations and can be implemented using small, Tensor-Core-friendly GEMMs fused with activation quantization.

  2. Outdated and insufficient evaluation for W4A4. The evaluation mainly uses perplexity and legacy short-form zero-shot benchmarks such as ARC, HellaSwag, PIQA, LAMBADA, and MMLU. These tasks do not measure error accumulation during long chain-of-thought generation. Recent studies [1][2] show that W4A4/W4A4KV4 may preserve conventional QA accuracy while substantially degrading mathematical, scientific, and coding reasoning. [1] Quantization Hurts Reasoning? [2] PTQ Baselines for Reasoning LLMs on Ascend

  3. Lack of reasoning-model evaluation. The paper evaluates base Llama and Qwen models but not modern reasoning models such as DeepSeek-R1-Distill or QwQ. Calibration on short WikiText-2 sequences may also fail to represent long-CoT activation distributions. Therefore, the current results do not demonstrate that MirrorQuant preserves reasoning ability under aggressive activation quantization.

Questions for authors
  1. How does MirrorQuant compare with a parameter-matched FlatQuant-style Kronecker transform when only R 4 ​ is replaced and all other settings are fixed?
  2. Are the Householder/FHT operations fused with activation quantization? How would their latency compare with a fused Kronecker GEMM kernel?
  3. Can the authors evaluate W4A4KV4 on reasoning models like DeepSeek-R1-Distill or QwQ using MATH-500, AIME, GPQA-Diamond, and LiveCodeBench?
7 comments
AuthorsOpenReview ↗

Response to Reviewer 4Jc8 (1)

Thank you for the careful review and for recognizing the deployment bottleneck at the online MLP interface, the exact orthogonality of the parameterization, and the motivation for projection guidance. Your suggestions on fused Kronecker alternatives and reasoning-model evaluation helped us extend the experiments in both directions.

W1. Missing comparison with Kronecker-structured online transformations. Q1. How does MirrorQuant compare with a parameter-matched FlatQuant-style Kronecker transform when only R4 is replaced and all other settings are fixed? Q2. Are the Householder/FHT operations fused with activation quantization? How would their latency compare with a fused Kronecker GEMM kernel?

Thank you for suggesting fusion with activation quantization. Following this suggestion, we implemented fused execution of MirrorQuant's FHT/Householder R4R_4 and per-token A4 quantization. We agree that FlatQuant's Kronecker-factorized affine transform is a relevant structured alternative and compared it using a fused Kronecker GEMM+A4 kernel.

Reference: FlatQuant, ICML 2025

For all three models, we hold R1/R2R_1/R_2 and the quantization settings fixed, replace only the online R4R_4 with a FlatQuant-style Kronecker transform matched to MirrorQuant’s R4R_4 in trainable parameter count, and retrain it during calibration on the corresponding model. Both alternatives use their fused implementations under W4A4KV4. The Δ\Delta column compares accuracy with the Kronecker baseline within each model. E2E denotes end-to-end latency, and Op. denotes the R4R_4+A4 operator path, excluding the down-projection GEMM. Bold values mark the better result within each model, including ties.

ModelOnline R4R_40-shot₉ (%) ↑Δ\Delta (pp) ↑E2E prefill (ms) ↓E2E decode (ms/token) ↓Op. prefill (ms) ↓Op. decode (ms) ↓
Llama-2-7BFlatQuant-style Kronecker58.40—605.42191.350.5870.100
MirrorQuant58.58+0.18606.76192.850.6680.100
Llama-2-13BFlatQuant-style Kronecker62.48—998.79248.110.6510.109
MirrorQuant62.79+0.311008.17249.820.6230.106
Llama-3-8BFlatQuant-style Kronecker62.47—549.06176.110.6500.092
MirrorQuant62.60+0.13543.66174.480.5590.087

MirrorQuant improves accuracy by 0.13–0.31 pp across all three models, with end-to-end prefill and decode latency differing by less than 1% from Kronecker in every comparison. Operator-level results depend on the model. Kronecker has lower prefill latency on Llama-2-7B, while MirrorQuant has lower prefill and decode latency on Llama-2-13B and Llama-3-8B. These fused comparisons show a small accuracy benefit for MirrorQuant at closely matched end-to-end latency.

We will report the fused Kronecker comparison and describe the fusion and timing protocol in Appendix.

AuthorsOpenReview ↗

Response to Reviewer 4Jc8 (2)

W2. Outdated and insufficient evaluation for W4A4 about mathematical, scientific, and coding reasoning tasks. W3. Lack of reasoning-model evaluation. Q3. Evaluation on reasoning models like DeepSeek-R1-Distill or QwQ using MATH-500, AIME, GPQA-Diamond, and LiveCodeBench.

Thank you for recommending these studies. We agree that the original evaluation on base models and short-form tasks left long-CoT reasoning insufficiently tested. Quantization Hurts Reasoning? shows that low-bit reasoning performance varies across models and tasks, and the Ascend study reports long-generation failures in several W4A4KV4 configurations. These findings motivated us to add direct generation-based evaluation on a reasoning model. We will cite both studies and clarify the scope of the original evaluation in the revision.

Reference: Quantization Hurts Reasoning?, COLM 2025 · PTQ Baselines for Reasoning LLMs on Ascend, arXiv

We retained the short WikiText-2 calibration and evaluated DeepSeek-R1-Distill-Qwen-7B on the four requested benchmarks. We use W4A4KV16 to isolate the effects of 4-bit weights and activations. MirrorQuant's rotation learning addresses a complementary part of the quantization pipeline to dedicated KV-cache methods such as PM-KVQ and TurboQuant. Keeping the cache at 16 bits provides a direct test of weight–activation quantization after short calibration.

Reference: PM-KVQ, ICLR 2026 · TurboQuant, ICLR 2026

We followed the evaluation setup of the Ascend study and ran BF16, FlatQuant, and MirrorQuant locally on Ascend 910C under the same protocol. Scores are in percent, with the better W4A4KV16 result bolded. Δ\Delta is the MirrorQuant score minus the FlatQuant score in percentage points. Evaluation protocol

MethodMATH-500 ↑AIME-120 ↑GPQA-Diamond ↑LiveCodeBench ↑
DeepSeek-R1-Distill-Qwen-7B93.2048.3348.4833.58
FlatQuant92.0037.5043.4328.73
MirrorQuant92.8038.3346.4629.48
Δ\Delta+0.80+0.83+3.03+0.75

Under W4A4KV16, MirrorQuant improves on FlatQuant across all four benchmarks, by 0.80 pp on MATH-500, 0.83 pp on AIME-120, 3.03 pp on GPQA-Diamond, and 0.75 pp on LiveCodeBench. After short-sequence calibration, MirrorQuant stays within 0.40 pp of BF16 on MATH-500. The gaps to BF16 are 2.02 pp on GPQA-Diamond and 4.10 pp on LiveCodeBench, with the largest gap of 10.00 pp on AIME-120.

We will discuss the two reasoning studies in Related Work and report the W4A4 reasoning experiments in Appendix.


We hope the above responses have adequately addressed your concerns. Please feel free to contact us if you have any further questions. If not, we would greatly appreciate your consideration for a higher score.

AuthorsOpenReview ↗

Thank You for Your Constructive Review — Discussion Open Until November 18

Dear Reviewer 4Jc8,

We sincerely appreciate your detailed review and the related studies you recommended. Your suggestions helped us identify where the evaluation could be strengthened. We have shared the additional results and explanations in our point-by-point responses and hope they address your concerns.

If any questions remain about our responses, we sincerely hope to have the opportunity to continue the discussion with you before it closes on November 18. We would be very happy to provide further analysis, run additional experiments, or explain in more detail any aspect you would like to discuss.

Thank you again for your thoughtful review and constructive feedback!

Best regards,

Authors of Submission 2369

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

MirrorQuant replaces dense learned rotations for low-bit LLM post-training quantization with a fixed orthogonal base followed by short Householder chains. This preserves exact orthogonality, reduces learned state to O(Kd) , and enables O(dlogd) online execution with an FHT base. The authors also introduce projection-guided objectives and progressive calibration. Experiments on Llama and Qwen models from 1.5B to 70B report consistent zero-shot accuracy gains over DartQuant—up to 1.81 percentage points—and lower online latency than a dense rotation.

Critical strengths
  1. Elegant and deployment-aware parameterization. Householder chains provide exact orthogonality with compact learned state and efficient online application. This directly addresses both calibration cost and the important non-foldable MLP interface.
  2. Broad and controlled evaluation. The study covers multiple model families, scales, and quantization settings. The ablations usefully separate the contributions of R1,R2,R4, projection guidance, and cache regeneration.
Critical weaknesses

1.Robustness of the accuracy gains is unclear. Results appear to use one calibration seed, without variance or significance analysis. Several improvements over DartQuant are only 0.03–0.3 points, and task-level outcomes are mixed. Thus, consistency across 23 settings does not by itself establish reliable improvement. 2. The practical advantage over the strongest alternatives is incompletely established. There is no joint accuracy–runtime comparison against a learned dense online R4, nor direct evaluation against RUQuant or HARP. Moreover, MirrorQuant has lower calibration throughput and higher memory than DartQuant, while the Llama-3-70B results rely on a DartQuant-derived base whose construction cost is excluded. These gaps weaken the efficiency and novelty claims.

Questions for authors
  1. Can you report mean and standard deviation over at least three calibration seeds for the main DartQuant comparisons and the R4 ablation?
  2. Under identical rotation sets and an end-to-end calibration budget, how do accuracy, total calibration time, and memory compare with DartQuant and a learned dense online R4, including base-construction costs?
  3. What controlled evidence demonstrates an advantage over RUQuant/HARP, and does regenerating the Stage-II cache with upstream weight quantization enabled materially change the results?
4 comments
AuthorsOpenReview ↗

Thank you for your positive review – Response to Reviewer BGwa (1)

Thank you for the careful review and for recognizing the practical value of compact, exactly orthogonal rotations at the non-foldable MLP interface. Your questions helped us distinguish the robustness of the accuracy gains, the one-time cost of calibration, and the recurring cost of applying the learned rotation.

W1. Robustness of the accuracy gains is unclear. Results appear to use one calibration seed, ... consistency across 23 settings does not by itself establish reliable improvement. Q1. Can you report mean and standard deviation over at least three calibration seeds for the main DartQuant comparisons and the R4 ablation?

Thank you for this question. To address it directly, we additionally conducted an ablation study on three different seeds 0, 42, and 123, as shown below.

Seed 0

MethodWikiText-2 PPL ↓PTB PPL ↓C4 PPL ↓0-shot₉ (%) ↑Δ\Delta (pp) ↑
QuaRot8.9115.2014.6156.35−5.25
DartQuant7.4613.0812.1061.60—
MirrorQuant fixed FHT R4R_47.4313.0312.0462.34+0.74
MirrorQuant7.4012.9811.9962.60+1.00

Seed 42

MethodWikiText-2 PPL ↓PTB PPL ↓C4 PPL ↓0-shot₉ (%) ↑Δ\Delta (pp) ↑
QuaRot9.0215.2614.5556.47−5.63
DartQuant7.4412.9412.0662.10—
MirrorQuant fixed FHT R4R_47.4512.9112.0362.49+0.39
MirrorQuant7.3712.9511.9762.93+0.83

Seed 123

MethodWikiText-2 PPL ↓PTB PPL ↓C4 PPL ↓0-shot₉ (%) ↑Δ\Delta (pp) ↑
QuaRot9.0415.6714.7855.83−5.21
DartQuant7.4513.1412.1261.04—
MirrorQuant fixed FHT R4R_47.4312.8412.0661.68+0.64
MirrorQuant7.4212.8812.0362.01+0.97

Average (mean ± sample standard deviation)

The Δ\Delta column compares 0-shot₉ accuracy with DartQuant at the same seed. The summary reports the mean and sample standard deviation of these paired differences.

MethodWikiText-2 PPL ↓PTB PPL ↓C4 PPL ↓0-shot₉ (%) ↑Δ\Delta (pp) ↑
QuaRot8.99 ± 0.0715.38 ± 0.2614.65 ± 0.1256.22 ± 0.34−5.36 ± 0.23
DartQuant7.45 ± 0.0113.05 ± 0.1012.09 ± 0.0361.58 ± 0.53—
MirrorQuant fixed FHT R4R_47.44 ± 0.0112.93 ± 0.1012.04 ± 0.0262.17 ± 0.43+0.59 ± 0.18
MirrorQuant7.40 ± 0.0312.94 ± 0.0512.00 ± 0.0362.51 ± 0.47+0.93 ± 0.09

MirrorQuant consistently improves both mean PPL and 0-shot₉ accuracy over DartQuant across all three seeds.

PPL. MirrorQuant improves mean PPL over DartQuant across all three seeds. The improvement varies across corpora, so absolute PPL differences should be interpreted in context. The evidence here is the consistent improvement over a strong baseline whose PPL is already closely matched.

Accuracy. The downstream benefit is more apparent: 0-shot₉ accuracy improves by +1.00, +0.83, and +0.97 pp, yielding a paired gain of +0.93 ± 0.09 pp across seeds. For context, matched-model accuracy gains under each paper’s own settings are +0.13 pp for DartQuant (NeurIPS25) over SpinQuant (W4A8KV16), +0.58 pp for InfoQuant (EMNLP26) over non-distilled OSTQuant (W4A4KV4), and +0.48 pp for TwinQuant (ICML26) over FlatQuant (W4A4). Our results therefore demonstrate consistent downstream gains.

Learned-R4R_4 ablation across seeds. With R1/R2R_1/R_2 held fixed within each seed, learning R4R_4 improves 0-shot₉ accuracy over fixed FHT by +0.26, +0.44, and +0.33 pp at seeds 0, 42, and 123, respectively. The paired gain is +0.34 ± 0.09 pp (mean ± sample standard deviation).

Reference: DartQuant, NeurIPS 2025 · InfoQuant, EMNLP 2026 · TwinQuant, ICML 2026

We will report the these results in the revised paper.

AuthorsOpenReview ↗

Response to Reviewer BGwa (2)

W2. Moreover, MirrorQuant has lower calibration throughput and higher memory than DartQuant, while the Llama-3-70B results rely on a DartQuant-derived base whose construction cost is excluded. These gaps weaken the efficiency and novelty claims. Q2. Under identical rotation sets and an end-to-end calibration budget, how do accuracy, total calibration time, and memory compare with DartQuant and a learned dense online R4, including base-construction costs?

Thank you for this question. Our additional experiments show that MirrorQuant achieves higher accuracy at comparable total calibration time and enables faster inference than learned dense online R4.

Learned R4: comparable accuracy, faster prefill. We fix R1R_1~R3R_3 and quantization settings, varying only the learned online R4R_4. Both variants fold R4⊤R_4^\top into the down-projection weights. MirrorQuant’s FHT–Householder path, fused with A4 quantization, reduces end-to-end prefill latency by 14.9–26.5%, with accuracy within 0.11 pp of learned dense R4.

ModelOnline R4R_4E2E prefill (ms) ↓Op. prefill (ms) ↓Op. decode (ms) ↓0-shot₉ (%) ↑Δ\Delta (pp) ↑
Llama-2-7BLearned dense712.724.2170.38458.62—
MirrorQuant606.760.6680.10058.58−0.04
Llama-2-13BLearned dense1237.086.4530.56462.68—
MirrorQuant1008.170.6230.10662.79+0.11
Llama-3-8BLearned dense739.816.8820.60362.53—
MirrorQuant543.660.5590.08762.60+0.07

Runs use W4A4KV4, GPTQ weights, BF16 compute, batch size 1, and one RTX A6000. Prefill uses 2,048 tokens. Operator timings cover layer-16 R4R_4+A4 before the down-projection GEMM.

Matched calibration: +0.51 pp at comparable total time. On Llama-3-8B (W4A4KV4, seed 0), both methods use the same calibration settings and learn rotations at the same sites within each pair. Total calibration time includes data preparation, rotation training, and cache generation where applicable.

MethodOnline R4R_4Stage I (min)Stage II (min)Total (min) ↓Peak VRAM (GiB) ↓Mean PPL ↓0-shot₉ (%) ↑Δ\Delta (pp) ↑
DartQuantFixed FHT14.3/6.9/10.8—32.016.39/—/16.3910.8561.75—
MirrorQuantFixed FHT14.3/7.3/11.5—33.126.68/—/26.6810.8362.34+0.59
DartQuantLearned dense14.3/6.9/10.89.2/10.151.316.39/20.53/20.5310.8162.09—
MirrorQuantLearned Householder14.3/7.3/11.510.6/7.551.226.68/17.28/26.6810.7962.60+0.51

Stage I lists data preparation / R1R_1 training / R2R_2 training. Stage II lists cache generation / R4R_4 training. Peak VRAM is reported for Stage I / Stage II / the full run. Δ\Delta is the accuracy difference from DartQuant within each pair.

Accuracy–cost trade-off versus DartQuant. With fixed FHT, MirrorQuant gains +0.59 pp in accuracy at 33.1 vs. 32.0 min. When both pipelines learn R4R_4, MirrorQuant gains +0.51 pp at nearly identical end-to-end calibration time (51.2 vs. 51.3 min). Its Stage II is faster (18.1 vs. 19.3 min) and uses less peak memory (17.28 vs. 20.53 GiB). Its full-run peak is 26.68 GiB during Stage I, compared with 20.53 GiB for DartQuant.

70B Model: shared base refinement. For Llama-3-70B, we reused a pre-specified DartQuant-derived R1 base to evaluate MirrorQuant within our compute budget. Table 11 shows the effect of initialization. MirrorQuant reaches 67.59% with a Hadamard base and 69.21% with the learned base. With the learned base shared, MirrorQuant improves over DartQuant by +0.86 pp (68.35% → 69.21%). Learning R4 contributes +0.26 pp with R1/R2 held fixed. Constructing the reused R1R_1 base adds approximately 5% to the end-to-end GPU-hours of MirrorQuant calibration with the base already available.

We will report the calibration and dense R4R_4 benchmarks in Appendix.

AuthorsOpenReview ↗

Response to Reviewer BGwa (3)

W2. The practical advantage over the strongest alternatives is incompletely established. There is no joint accuracy–runtime comparison against a learned dense online R4, nor direct evaluation against RUQuant or HARP. Q3. What controlled evidence demonstrates an advantage over RUQuant/HARP, and does regenerating the Stage-II cache with upstream weight quantization enabled materially change the results?

We evaluated RUQuant and ButterflyQuant on the same Llama-3-8B checkpoint under W4A4KV4, using the same PPL and nine-task zero-shot metrics as for MirrorQuant. The Δ\Delta columns use RUQuant as the W4A4KV4 reference.

MethodMean PPL ↓Δ\Delta PPL ↓0-shot₉ (%) ↑Δ\Delta (pp) ↑
Full precision8.92—66.04—
ButterflyQuant11.81+0.8558.82−0.81
RUQuant10.96—59.63—
MirrorQuant10.79−0.1762.60+2.97

MirrorQuant gains 2.97 pp over RUQuant and 3.78 pp over ButterflyQuant, while lowering mean PPL by 0.17 and 1.02, respectively. It scores higher than both methods on each of the nine zero-shot tasks. HARP's published evaluation uses weight-only quantization. A comparison under the same W4A4KV4 setting would require evaluating HARP with 4-bit activations and KV cache as well.

We will report the structured-baseline comparisons and discuss HARP's evaluation setting in Appendix.

Reference: RUQuant KDD 2026, ButterflyQuant arXiv, HARP EMNLP 2026

W2 / Q3. Stage-II cache regeneration with upstream weight quantization.

Thank you for asking whether upstream weight quantization changes the outcome. The table shows the original full-precision cache (W16A16, Table 9B), the default regenerated cache (W16A4, Table 9D), and a new cache regenerated with GPTQ W4 and the same A4. All rows use Llama-3-8B, seed 0, the same projection-guided R4R_4 objective with truncated GℓG_\ell, and final W4A4KV4 evaluation. Both regenerated caches use the same calibration samples and frozen R1/R2R_1/R_2 to collect MLP down-projection inputs (Appendix B.1). Deltas are relative to W16A4.

Precision during cache generationWikiText-2 ↓PTB ↓C4 ↓Mean PPL ↓Δ\Delta PPL ↓0-shot₉ (%) ↑Δ\Delta (pp) ↑
W16A16 (original cache)7.4413.1312.0610.88+0.0962.43−0.17
W16A4 (default)7.4012.9811.9910.79—62.60—
W4A4 (GPTQ)7.3812.9612.0010.78−0.0162.65+0.05

Refreshing the original cache with the fixed upstream rotations and A4 lowers mean PPL by 0.09 and improves accuracy by 0.17 pp. Comparing the two regenerated caches isolates the effect of upstream W4, which lowers mean PPL by a further 0.01 and improves accuracy by 0.05 pp. The W4A4 variant incurs extra cost during Stage-II data construction because we use GPTQ to quantize the upstream weights for cache generation. Given the small additional gain in this setting, we retain W16A4 as the default to reduce cache-construction cost.

We will add the upstream-W4 cache result to Table 9 and discuss its construction cost in Appendix.


We hope the above responses have adequately addressed your concerns. Please feel free to contact us if you have any further questions. If not, we would greatly appreciate your consideration for a higher score.

AuthorsOpenReview ↗

Thank You for Your Positive Review — Discussion Open Until November 18

Dear Reviewer BGwa,

We would like to express our sincere gratitude for your careful review and your encouraging assessment of our work. We appreciate the thoughtful questions you raised and have included additional results and explanations in our point-by-point responses. We hope these additions address the points you raised.

As the discussion period remains open until November 18, we would warmly welcome any further thoughts or suggestions you may have. We would be happy to clarify any points you would like to discuss, provide further analysis, or run additional experiments if helpful.

Thank you again for taking the time to consider our work!

Warm regards,

Authors of Submission 2369

Reviewer_TdWnOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper found that existing learned-rotation methods for low-bit LLM quantization make the rotations impractical at non-foldable interfaces. To address it, this paper proposes MirrorQuant, a framework that parameterizes each learned rotation as a fixed orthogonal base followed by a short chain of K Householder reflections. It also introduces projection-guided learning, which penalizes the quantization error after the subsequent linear projection. Experiments on three models from 1.5B to 70B parameters show that this method improves zero-shot accuracy over DartQuant in all 23 matched settings.

Critical strengths
  1. This paper identifies a concrete bottleneck in learned-rotation quantization: learned rotations are effectively restricted to foldable interfaces since existing methods parameterize rotations as dense matrices. MirrorQuant makes model-specific rotations affordable at non-foldable interfaces, and cheaper to learn at foldable ones, which is directly relevant to the practical deployment of low-bit LLM.

  2. Each learned rotation consists of only a fixed base and a few reflections, so it has few parameters to learn. And calibration does not modify the model weights or quantization settings, so the method is easy to integrate into existing quantization pipelines.

  3. The experiments cover two model families, nine model sizes from 1.5B to 70B, and three low-bit configurations. And the paper compares against seven common quantization methods under a unified setting.

Critical weaknesses
  1. Small improvements over DartQuant. MirrorQuant shares the same rotation placement and quantization pipeline as DartQuant, but improves zero-shot accuracy by only 0.61 points on average on Llama models, and by only 0.2–0.4 points in many settings.

  2. Small benefits of the learned online rotation. The online R4 is the main motivation of this paper, but its contribution is limited. In Table 7, most of the gain over DartQuant comes from the R₁ and R₂ while the online R₄ adds only 0.26 points.

  3. There are tons of learnable rotation quantizaton method, like butterflyquant, are not studied in the experiment.

Questions for authors
  1. Given that the improvements are small, could you report results over multiple random seeds to show that the gains are not due to randomness?

  2. How were the model-specific hyperparameters selected?

  3. Could you report the gain and latency of R4 over the fixed Hadamard transform on other models and bit settings?

6 comments
AuthorsOpenReview ↗

Response to Reviewer TdWn (1)

Thank you for your careful review and for recognizing the deployment bottleneck at non-foldable interfaces and the role of the compact rotation parameterization. Your comments helped us distinguish the overall accuracy gain from the incremental benefit of the online rotation, and prompted us to test both more directly.

W1. Small improvements over DartQuant. MirrorQuant shares the same rotation placement and quantization pipeline as DartQuant, but improves zero-shot accuracy by only 0.61 points on average on Llama models, and by only 0.2–0.4 points in many settings. Q1. Given that the improvements are small, could you report results over multiple random seeds to show that the gains are not due to randomness

Thank you for this question. To check whether the advantage persists across calibration runs, we additionally conducted an ablation study on three different seeds 0, 42, and 123, as shown below.

Seed 0

MethodWikiText-2 PPL ↓PTB PPL ↓C4 PPL ↓0-shot₉ (%) ↑Δ\Delta (pp) ↑
QuaRot8.9115.2014.6156.35−5.25
DartQuant7.4613.0812.1061.60—
MirrorQuant7.4012.9811.9962.60+1.00

Seed 42

MethodWikiText-2 PPL ↓PTB PPL ↓C4 PPL ↓0-shot₉ (%) ↑Δ\Delta (pp) ↑
QuaRot9.0215.2614.5556.47−5.63
DartQuant7.4412.9412.0662.10—
MirrorQuant7.3712.9511.9762.93+0.83

Seed 123

MethodWikiText-2 PPL ↓PTB PPL ↓C4 PPL ↓0-shot₉ (%) ↑Δ\Delta (pp) ↑
QuaRot9.0415.6714.7855.83−5.21
DartQuant7.4513.1412.1261.04—
MirrorQuant7.4212.8812.0362.01+0.97

Across seeds (mean ± sample standard deviation)

The Δ\Delta column summarizes the paired accuracy differences across seeds.

MethodWikiText-2 PPL ↓PTB PPL ↓C4 PPL ↓0-shot₉ (%) ↑Δ\Delta (pp) ↑
QuaRot8.99 ± 0.0715.38 ± 0.2614.65 ± 0.1256.22 ± 0.34−5.36 ± 0.23
DartQuant7.45 ± 0.0113.05 ± 0.1012.09 ± 0.0361.58 ± 0.53—
MirrorQuant7.40 ± 0.0312.94 ± 0.0512.00 ± 0.0362.51 ± 0.47+0.93 ± 0.09

Accuracy. On Llama-3-8B (W4A4KV4), the paired 0-shot₉ gains over DartQuant are +1.00, +0.83, and +0.97 pp at seeds 0, 42, and 123, respectively. The paired gain is +0.93 ± 0.09 pp (mean ± sample standard deviation).

PPL. Averaged over WikiText-2, PTB, and C4, MirrorQuant’s PPL is lower than DartQuant’s by approximately 0.09, 0.05, and 0.13 at seeds 0, 42, and 123, respectively. MirrorQuant has lower PPL in eight of nine corpus–seed comparisons.

For context, matched-model accuracy gains under each paper’s own settings are +0.13 pp for DartQuant (NeurIPS25) over SpinQuant (W4A8KV16), +0.58 pp for InfoQuant (EMNLP26) over non-distilled OSTQuant (W4A4KV4), and +0.48 pp for TwinQuant (ICML26) over FlatQuant (W4A4).

Reference: DartQuant, NeurIPS 2025 · InfoQuant, EMNLP 2026 · TwinQuant, ICML 2026

We will report the three-seed results in Appendix of the revised paper.

AuthorsOpenReview ↗

Response to Reviewer TdWn (2)

Q2. How were the model-specific hyperparameters selected?

Thank you for asking about the selection protocol. We use a common calibration schedule and a default reflection depth of four across models. The Llama-3-8B ablations support this depth for both R2 accuracy and the R4 accuracy–latency trade-off (Tables 4 and 10).

For the model-specific loss weights λv\lambda_v and λact\lambda_{\mathrm{act}}, we evaluate a small set of candidate values and select them based on reconstruction error on held-out samples. The R2 output-projection weight is shared across models, and head balancing is used in the grouped-query variant for Llama-3 and Qwen2.5. Appendix B.2 lists the final settings.

W2. Small benefits of the learned online rotation. The online R4 is the main motivation of this paper, but its contribution is limited. In Table 7, most of the gain over DartQuant comes from the R₁ and R₂ while the online R₄ adds only 0.26 points. Q3. Could you report the gain and latency of R4 over the fixed Hadamard transform on other models and bit settings?

Thank you for this question. Controlled comparison with fixed Hadamard. For each model and bit setting, we hold R1/R2R_1/R_2 and the quantization settings fixed and replace the fixed FHT with MirrorQuant’s learned R4R_4. In the A4 runs, MirrorQuant fuses the online FHT/Householder rotation with per-token A4 quantization. Operator latency includes R4R_4 and activation quantization at the bit width shown. Δ\Delta is the accuracy difference from fixed FHT. Bold marks the better result within each pair, including ties.

ModelBitsOnline R4R_4E2E prefill (ms) ↓Op. prefill (ms) ↓Op. decode (ms) ↓0-shot₉ (%) ↑Δ\Delta (pp) ↑
Llama-2-7BW4A4KV4Fixed FHT602.620.6370.10058.17—
MirrorQuant606.760.6680.10058.58+0.41
Llama-2-13BW4A4KV4Fixed FHT987.290.6320.10462.24—
MirrorQuant1008.170.6230.10662.79+0.55
Llama-3-8BW4A4KV4Fixed FHT539.020.3760.09562.34—
MirrorQuant543.660.5590.08762.60+0.26
Llama-3-8BW4A4KV16Fixed FHT499.100.3800.10162.31—
MirrorQuant504.720.6340.10262.78+0.47
Llama-3-8BW4A8KV16Fixed FHT500.870.3800.09965.04—
MirrorQuant507.510.6180.10165.22+0.18

Accuracy gains at comparable latency. We agree that the 0.26 pp gain in Table 7 on Llama-3-8B (W4A4KV4) is modest. At the same precision, learned R4R_4 improves 0-shot₉ accuracy over fixed FHT by 0.41 pp on Llama-2-7B and 0.55 pp on Llama-2-13B. On Llama-3-8B with a 16-bit KV cache, it improves accuracy by 0.47 pp under W4A4KV16 and 0.18 pp under W4A8KV16. End-to-end prefill and operator-level decode latencies are comparable to fixed FHT across all five settings.

We will report the fixed-Hadamard accuracy and latency comparisons across models and bit widths in Appendix.

W3. There are tons of learnable rotation quantizaton method, like butterflyquant, are not studied in the experiment.

Thank you for suggesting a comparison with other structured rotations. We will firstly discuss these methods in the related work. Additionally, we evaluated ButterflyQuant and RUQuant on Llama-3-8B under W4A4KV4 using the same three-corpus mean PPL and nine-task zero-shot metrics as for MirrorQuant. The Δ\Delta columns use ButterflyQuant as the reference within W4A4KV4. ButterflyQuant, arXiv · RUQuant, KDD 2026

MethodMean PPL ↓Δ\Delta PPL ↓0-shot₉ (%) ↑Δ\Delta (pp) ↑
Full precision8.92—66.04—
ButterflyQuant11.81—58.82—
RUQuant10.96−0.8559.63+0.81
MirrorQuant10.79−1.0262.60+3.78

MirrorQuant improves on RUQuant by 2.97 pp and lowers mean PPL by 0.17. It scores higher than both baselines on each of the nine zero-shot tasks. This comparison evaluates the full methods, while the fixed-FHT comparison isolates the contribution of the online rotation.

We will discuss structured rotations and report the ButterflyQuant and RUQuant comparisons in the revised version.


We hope the above responses have adequately addressed your concerns. Please feel free to contact us if you have any further questions. If not, we would greatly appreciate your consideration for a higher score.

AuthorsOpenReview ↗

Thank You for Your Review. We Welcome Further Discussion Before November 18

Dear Reviewer TdWn,

Thank you very much for the time and effort you devoted to reviewing our work and for your constructive suggestions. Your feedback gave us the opportunity to explain our experimental choices more clearly and include additional comparisons in our point-by-point responses. We hope these additions address your questions.

If you have any further questions or suggestions, we would welcome the opportunity to discuss them with you before the discussion period closes on November 18. We would be glad to explain any part of our responses in more detail, share further analysis, or carry out additional experiments.

Thank you again for your careful review and valuable suggestions!

Best regards,

Authors of Submission 2369

Reviewer_A97ZOpenReview ↗
Rating
3(+1)
Confidence
2(−1)
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes MirrorQuant. The main contribution of this paper is twofold.

  • First, it leverages Householder Rotations for Low-Bit LLM Quantization to eliminate the overhead of the unfoldable interface of rotation-based quantization methods. This technique replaces the remaining dense online interface (which is mainly R4R_4 described in the paper) with a lightweight online transformation, whose computation complexity is O(dlog⁡d)O(d \log d), which is lower than the complexity of dense O(d2)O(d^2)

  • Second, it proposes projection-guided progressive learning to mitigate the propagation of quantization error. It divides the learning into two stages. In the first stage, MirrorQuant learns the Foldable Rotations. In the second stage, MirrorQuant learns the Online MLP Rotation with foldable quantization enabled and frozen.

Critical strengths
  • The paper identifies a real deployment asymmetry. R1R_1 and R2R_2 can be folded into adjacent weights, whereas R4R_4 follows the gated MLP’s elementwise product and must remain on the inference path.

  • The paper proposes to replace the original dense R4R_4 with a fixed Hadamard base followed by K Householder reflections, which preserves orthogonality, reduces the learned state to O(Kd)O(Kd), and gives O(dlog⁡d)O(d \log d) online computation at fixed depth.

  • Projection-guided learning is well motivated, and the two-stage schedule calibrates R4R_4 on activations produced by the fixed upstream rotation and quantization configuration.

Critical weaknesses
  • The overall accuracy gain seems small. Relative to the primary matched baseline, DartQuant, the average 0-shot₉ improvement is 0.61 points, and the gains under W4A8KV16 are often around 0.1 points.

  • Limited consideration of error propagation. Projection-guided learning only considers the immediate downstream linear projection, without explicitly accounting for error propagation through subsequent Transformer layers. It remains unclear whether local projection error adequately reflects the impact on final model outputs.

  • Insufficient ablation and efficiency analysis. The contributions of Householder parameterization and projection-guided learning are not fully disentangled. Moreover, MirrorQuant exhibits lower training throughput and higher peak GPU memory than DartQuant in Table 3, despite its claimed efficiency advantages.

  • Limited standalone scalability evidence. The Llama-3-70B results rely on a DartQuant-derived rotation base. With the default Hadamard base, MirrorQuant performs worse than DartQuant, raising concerns about its independent scalability.

Questions for authors
  1. What fundamentally distinguishes projection-guided learning from existing output-reconstruction objectives? Could the authors provide ablations separating the effects of the Householder parameterization and the learning objective?

  2. Why is immediate projection error considered a sufficient proxy for downstream sensitivity? Would block-level or multi-layer output reconstruction provide better guidance?

  3. Could the authors report end-to-end calibration time, total GPU-hours, and peak memory under comparable settings, particularly against DartQuant?

  4. Can MirrorQuant achieve competitive performance at the 70B scale without relying on a previously learned DartQuant rotation? How would including the base construction cost affect the efficiency comparison?

7 comments
AuthorsOpenReview ↗

Response to Reviewer A97Z (1)

Thank you for recognizing the deployment cost of the non-foldable R4R_4, the compact orthogonal Householder construction, and the motivation for projection-guided calibration.

W1. The overall accuracy gain seems small. Relative to the primary matched baseline, DartQuant, the average 0-shot9 improvement is 0.61 points, and the gains under W4A8KV16 are often around 0.1 points.

Thank you for raising this point. Gains are smaller under W4A8KV16, where higher activation precision leaves less quantization error to recover. To assess the consistency of our improvement under more aggressive quantization, we evaluate Llama-3-8B at W4A4KV4 across seeds 0, 42, and 123.

Seed 0

MethodWikiText-2 PPL ↓PTB PPL ↓C4 PPL ↓0-shot₉ (%) ↑Δ\Delta (pp) ↑
QuaRot8.9115.2014.6156.35−5.25
DartQuant7.4613.0812.1061.60—
MirrorQuant7.4012.9811.9962.60+1.00

Seed 42

MethodWikiText-2 PPL ↓PTB PPL ↓C4 PPL ↓0-shot₉ (%) ↑Δ\Delta (pp) ↑
QuaRot9.0215.2614.5556.47−5.63
DartQuant7.4412.9412.0662.10—
MirrorQuant7.3712.9511.9762.93+0.83

Seed 123

MethodWikiText-2 PPL ↓PTB PPL ↓C4 PPL ↓0-shot₉ (%) ↑Δ\Delta (pp) ↑
QuaRot9.0415.6714.7855.83−5.21
DartQuant7.4513.1412.1261.04—
MirrorQuant7.4212.8812.0362.01+0.97

Across seeds (mean ± sample standard deviation)

The Δ\Delta column summarizes the paired accuracy differences across seeds.

MethodWikiText-2 PPL ↓PTB PPL ↓C4 PPL ↓0-shot₉ (%) ↑Δ\Delta (pp) ↑
QuaRot8.99 ± 0.0715.38 ± 0.2614.65 ± 0.1256.22 ± 0.34−5.36 ± 0.23
DartQuant7.45 ± 0.0113.05 ± 0.1012.09 ± 0.0361.58 ± 0.53—
MirrorQuant7.40 ± 0.0312.94 ± 0.0512.00 ± 0.0362.51 ± 0.47+0.93 ± 0.09

On Llama-3-8B under W4A4KV4, MirrorQuant improves both mean PPL and 0-shot₉ accuracy over DartQuant at all three seeds.

PPL. MirrorQuant achieves lower seed-averaged PPL on all three datasets.

Accuracy. MirrorQuant improves 0-shot₉ accuracy by +1.00, +0.83, and +0.97 pp at seeds 0, 42, and 123, respectively, for a paired gain of +0.93 ± 0.09 pp (mean ± sample standard deviation). Across matched models under each published study's own settings, the average gains are +0.13 pp for DartQuant over SpinQuant (W4A8KV16, Table 2), +0.58 pp for InfoQuant over non-distilled OSTQuant (W4A4KV4, all available pairs in Table 1), and +0.48 pp for TwinQuant over FlatQuant (W4A4, Table 1).

Reference: DartQuant, NeurIPS 2025 · InfoQuant, EMNLP 2026 · TwinQuant, ICML 2026

We will report the three-seed results in Appendix of the revised paper.

AuthorsOpenReview ↗

Response to Reviewer A97Z (2)

W2. Limited consideration of error propagation. Projection-guided learning only considers the immediate downstream linear projection, without explicitly accounting for error propagation through subsequent Transformer layers. It remains unclear whether local projection error adequately reflects the impact on final model outputs. Q2. Why is immediate projection error considered a sufficient proxy for downstream sensitivity? Would block-level or multi-layer output reconstruction provide better guidance?

Thank you for your question.

Relation to block-level reconstruction. When only R4 varies, the residual branch of the current Llama block is fixed. Because the block output adds this residual to the down-projection output, the full down-projection output difference equals the change in the current block’s output. Our rank-256 projection term approximates this local output error.

Relation to final-logit error. On Llama-3-8B, we enable A4 at one down-projection input at a time, with weights and the KV cache kept at 16 bits. We use the same 16 held-out sequences at each prespecified layer, totaling 2,032 token positions per layer. For each sequence, we compute Spearman’s ρ\rho between each local error measure and the relative squared error of the final logits across token positions. The table reports the median across sequences.

Layer (zero-indexed)Activation quantization error (ρ\rho)Immediate projection error (ρ\rho)
160.6590.669
200.6480.685
280.7040.748

Immediate projection error has a higher median correlation with final-logit error at all three tested layers, including layer 16, where the perturbation passes through 15 subsequent Transformer blocks. The median correlations of 0.669–0.748 support its use as a proxy for ranking tokens by downstream sensitivity under this single-site intervention.

Calibration with deeper reconstruction targets. We replace the immediate down-projection target with the output after four or eight blocks while keeping the R4 parameterization, activation reconstruction term, and Stage-II training steps fixed. All runs use Llama-3-8B at W4A4KV4 with seed 0. Deltas are relative to the immediate target.

Additional reconstruction targetMean PPL ↓Δ\Delta PPL ↓0-shot₉ (%) ↑Δ\Delta (pp) ↑
Immediate down-projection10.79—62.60—
Output after 4 blocks10.86+0.0762.37−0.23
Output after 8 blocks10.82+0.0362.22−0.38

Under the same number of optimization steps, the immediate target yields lower mean PPL and higher zero-shot accuracy than either tested multi-block target. The correlation and calibration results support using immediate projection error as a practical proxy for downstream sensitivity in this Stage-II setting.

We will add the final-logit error analysis and multi-block reconstruction experiments to Appendix.

AuthorsOpenReview ↗

Response to Reviewer A97Z (3)

W3. Insufficient ablation and efficiency analysis. The contributions of Householder parameterization and projection-guided learning are not fully disentangled. Q1. What fundamentally distinguishes projection-guided learning from existing output-reconstruction objectives? Could the authors provide ablations separating the effects of the Householder parameterization and the learning objective?

Thank you for asking us to clarify the learning objective, separate the contributions of the two design choices, and explain the efficiency trade-off with DartQuant.

Projection guidance at the immediate down-projection. As a representative block-level output-reconstruction method, FlatQuant jointly fits transformations, scaling, and clipping by minimizing the reconstruction error at the output of an entire Transformer block (Eq. 4). MirrorQuant freezes the upstream configuration and learns only R4 in Stage II. Its objective combines activation reconstruction with a rank-256 approximation to the error propagated through the immediate down-projection (Eq. 18). This gives greater weight to activation errors that the projection amplifies. The full and rank-256 metrics yield the same mean PPL (10.79) and differ by 0.05 pp in accuracy (Table 9).

Separating the objective and parameterization. We toggle the projection term for Householder and Givens on Llama-3-8B (W4A4KV4, seed 0). All four runs retain activation reconstruction and share the same trainable parameter count, FHT base, frozen R1/R2, Stage-II cache, and training budget.

R4R_4 parameterizationProjection termMean PPL ↓0-shot₉ (%) ↑Δ\Delta (pp) ↑
GivensOff10.8262.22—
GivensOn10.8162.41+0.19
HouseholderOff10.7962.22—
HouseholderOn10.7962.60+0.38

Projection guidance improves accuracy with both parameterizations. Householder and Givens tie in accuracy with activation reconstruction alone. With projection guidance enabled, Householder achieves 0.19 pp higher accuracy and 0.02 lower mean PPL at the same parameter count. This supports the choice of Householder for MirrorQuant’s projection-guided R4R_4.

We will clarify the distinction from block-level reconstruction and add the controlled Householder–Givens ablation to the appendix.

W3. Moreover, MirrorQuant exhibits lower training throughput and higher peak GPU memory than DartQuant in Table 3, despite its claimed efficiency advantages. Q3. Could the authors report end-to-end calibration time, total GPU-hours, and peak memory under comparable settings, particularly against DartQuant?

Thank you for asking for a comparison of total calibration time, GPU-hours, and peak memory. To quantify the full rotation-calibration cost, we benchmark both methods on Llama-3-8B at W4A4KV4 using one RTX A6000 per run. Both methods use the same calibration settings within each pair. The totals include data preparation, rotation learning, and Stage-II cache generation where applicable. We extend DartQuant with a learned dense R4 for the learned-R4 comparison.

MethodOnline R4R_4Stage I (min)Stage II (min)Total (min) ↓Peak VRAM (GiB) ↓Mean PPL ↓0-shot₉ (%) ↑Δ\Delta (pp) ↑
DartQuantFixed FHT14.3/6.9/10.8—32.016.39/—/16.3910.8561.75—
MirrorQuantFixed FHT14.3/7.3/11.5—33.126.68/—/26.6810.8362.34+0.59
DartQuantLearned dense14.3/6.9/10.89.2/10.151.316.39/20.53/20.5310.8162.09—
MirrorQuantLearned Householder14.3/7.3/11.510.6/7.551.226.68/17.28/26.6810.7962.60+0.51

Stage I lists data preparation / R1R_1 training / R2R_2 training. Stage II lists cache generation / R4R_4 training. Peak VRAM is reported for Stage I / Stage II / the full run. Δ\Delta is the accuracy difference from DartQuant within each pair.

Calibration time, memory, and accuracy. With fixed FHT, MirrorQuant gains 0.59 pp at 33.1 vs. 32.0 min (0.552 vs. 0.533 GPU-hours), with peak VRAM of 26.68 vs. 16.39 GiB. When both pipelines learn R4R_4, MirrorQuant gains 0.51 pp at 51.2 vs. 51.3 min (0.853 vs. 0.855 GPU-hours). Its Stage-II peak is lower (17.28 vs. 20.53 GiB), and its overall peak occurs in Stage I (26.68 vs. 20.53 GiB overall).

We will report the end-to-end calibration time, GPU-hours, and peak memory in Appendix.

Comments on the paper

1 comment
AuthorsOpenReview ↗

A Summary of Reviews

Thank you to all four reviewers for taking the time to review our paper and for providing valuable, constructive feedback! We are grateful that the reviewers recognized the following strengths of MirrorQuant.

  1. The paper identifies a real deployment asymmetry between foldable rotations and the online MLP rotation R4R_4. MirrorQuant makes model-specific rotations affordable at non-foldable interfaces (A97Z, TdWn, BGwa, 4Jc8).
  2. The deployment-aware Householder parameterization preserves exact orthogonality with compact learned state. For online R4R_4, an FHT base and a short Householder chain replace the quadratic O(d2)O(d^2) dense matrix. The chain’s learned state and computation are linear O(d)O(d) at fixed depth (A97Z, BGwa, 4Jc8).
  3. Projection-guided learning is well motivated and supported by ablations (A97Z, 4Jc8).
  4. The paper provides a broad and controlled evaluation across model families, scales, and quantization settings, with baselines compared under a unified setting (TdWn, BGwa).

We thank all reviewers for their invaluable feedback and have supplemented the paper with additional experiments and works, including:

  1. Conducted calibration with three additional seeds to assess the gains over DartQuant and the contribution of learned R4R_4.
  2. Separated the effects of Householder parameterization and projection guidance, tested how immediate projection error relates to final-logit error, and compared immediate projection with multi-block reconstruction.
  3. Benchmarked learned online R4R_4 against fixed Hadamard, learned dense, and fused Kronecker rotations. We also compared with ButterflyQuant and RUQuant.
  4. Measured end-to-end calibration time and peak memory, tested Stage-II cache regeneration with upstream weight quantization, and clarified hyperparameter selection.
  5. Examined the effect of the R1R_1 base at 70B and clarified how base construction contributes to first-run calibration cost.
  6. Extended the evaluation to long-form reasoning across benchmarks in mathematics, science, and code generation.

You can find our individual responses below your review comments. If you have any more concerns or questions, we are entirely open to continuing the discussion with you!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.