acceptodds
Withdrawn from ICLR 2027

QuantLRM: Quantization of Large Reasoning Models via Fine-Tuning Signals

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Weight-only quantization is important for compressing Large Reasoning Models (LRMs). However, existing post-training quantization (PTQ) methods are suboptimal for LRMs that are typically fine-tuned from base models, because they leave their fine-tuning traces wasted during PTQ. Inspired by the spirit of classical magnitude pruning, we study whether the magnitude of weight updates during reasoning-incentivized fine-tuning can provide valuable signals for quantizing LRMs. We hypothesize that the smallest and largest weight updates are more important than those of intermediate magnitude, a phenomenon we term "protecting both ends". Upon hypothesis validation, we introduce QuantLRM, the first method for weight quantization of LRMs that explicitly leverages fine-tuning signals. We fit simple restricted quadratic functions on weight updates to protect both ends. By multiplying the average quadratic values with the count of zero weight updates of channels, we compute channel importance that is more effective than using activation or second-order information, which guides quantization by protecting important weights. To demonstrate the effectiveness and applicability of QuantLRM, we evaluate sub-4-bit quantization performance on various fine-tuning strategies (SFT, DPO, and RLVR) over different model families (Llama, Qwen, and Olmo) on a range of reasoning benchmarks (AIME-120, FOLIO, temporal sequences, and GPQA-Diamond). QuantLRM delivers consistent gains in LRMs quantization, improving average accuracy by up to 6.55% over the best baselines.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Quantization, Weight-Only Quantization, LLMs Compression, LRMs Compression

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_956kOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

QuantLRM uses weight differences between a reasoning model and an earlier checkpoint to guide weight-only quantization. It assigns higher importance to small and large updates, separately handles zero updates, and incorporates these scores into an AWQ-style scaling search. Experiments across SFT, DPO, and RLVR models report improved reasoning accuracy, particularly at 3-bit precision.

Critical strengths

S1: Consistent gains across models and training stages. The method improves average accuracy over the strongest included baselines across several model families, supporting the practical value of checkpoint differences.

S2: Simple integration with relevant ablations. QuantLRM fits into an existing quantization workflow. Experiments compare different protected regions, mapping functions, and zero-update treatments, including comparisons using AWQ scaling.

Critical weaknesses

W1. The comparison with other delta-aware approaches is limited. DAQ already uses base/post-trained weight differences to guide quantization. Although it is discussed in §2, the experiments do not establish the advantage of QuantLRM’s specific scoring rule over alternative delta-aware objectives. DAQ’s FP8 setting differs from the paper’s W3 focus, so a controlled comparison or a clearer discussion of this distinction would strengthen the contribution.

W2. The interpretation and robustness of zero-update weighting remain unclear. Exact-zero counts can depend on checkpoint precision and rounding, which complicates their interpretation as weight importance. Appendix B also introduces matrix slicing to avoid overflow when computing the scores, but does not clearly establish whether this preserves Equation 5. These details matter because zero-update counts directly multiply the channel scores.

W3. The reported confidence intervals do not reflect the original benchmark sizes. Appendix C uses 50,000 paired draws per benchmark in each bootstrap replicate, substantially exceeding the actual question counts. Without an appropriate adjustment, this produces overly narrow intervals for the reported evaluation sizes. The significance of the smaller gains, particularly the approximately 0.25-point DPO improvement, therefore needs reassessment.

Questions for authors

Q1: Can you compare alternative delta-aware scoring rules or objectives under the same W3 quantizer and search budget, to isolate the benefit of the proposed two-ended score?

Q2: Is the matrix-slicing procedure equivalent to Equation 5? How sensitive are channel rankings and accuracy to the slicing configuration and the precision used to identify zero updates?

Q3: Do the smaller improvements remain statistically supported when paired bootstrap samples match each benchmark’s actual question count? How is stochastic decoding accounted for?

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

QuantLRM uses weight differences between a reasoning model and its preceding checkpoint to guide weight-only post-training quantization. A piecewise quadratic mapping emphasizes small and large updates, while zero-update counts amplify channel importance. These scores guide an AWQ-style scaling search. Experiments cover five models, three model families, SFT/DPO/RLVR, and four reasoning benchmarks. QuantLRM achieves the highest reported average 3-bit accuracy among the evaluated baselines on each model, with gains up to 6.55 percentage points.

Critical strengths
  1. A useful and practical source of quantization information. Differences between checkpoints provide information beyond the final model’s weights or activations. The method integrates this signal into a simple scaling procedure and requires no additional training when the preceding checkpoint is available.

  2. Encouraging results across several settings. Table 2 reports the best average accuracy on all five models, including improvements of 3.03 percentage points on R1-Qwen-32B and 6.55 points on Olmo-RLVR over their strongest evaluated baselines. Coverage across model families and training pipelines supports the method’s potential usefulness.

  3. Relevant ablations support parts of the design. The experiments examine alternative mappings, zero handling, and protected regions. Particularly, Table 4 shows that adding the high-update branch while retaining the same zero handling improves average accuracy from 70.83 to 73.58. This provides concrete support beyond the initial mixed-precision experiments.

Critical weaknesses
  1. The reported statistical significance is not established by the bootstrap procedure. Section 5.1 and Appendix C draw 50,000 paired observations per benchmark in each replicate, despite the much smaller evaluation sets, and directly use the resulting distribution to estimate uncertainty. Without correcting for the enlarged resample size, this artificially narrows the confidence intervals. More bootstrap repetitions reduce simulation error; they do not justify treating resampled observations as additional independent data. The claimed significance, especially for the 0.251-point DPO improvement, therefore requires reassessment. Repeated decoding variability is also unreported despite sampling at temperature 0.6. This does not invalidate the larger point estimates, but weakens the claim of consistently reliable improvements.

  2. The ablation study is incomplete. The paper lacks a zero-count-only baseline and a high-update-only variant with the same zero handling as the full method. These comparisons are needed to clarify the individual contributions of the two-ended mapping and zero-count weighting.

  3. The exact-zero rule may depend on numerical representation. Equation (5) treats exact equality between checkpoints as an importance signal, but Appendix B uses 16-bit arithmetic and model-specific slicing and averaging of zero counts. No sensitivity analysis establishes whether channel rankings and accuracy remain stable across checkpoint precision or a small-update threshold. This leaves the robustness and reproducibility of a central component uncertain.

  4. The practical evidence does not cover the targeted 3-bit deployment. Appendix B reports pseudo-quantized accuracy, while Table 3 measures throughput only with a 4-bit kernel. Consequently, the results establish 3-bit accuracy gains and 4-bit runtime compatibility separately; actual 3-bit memory and latency benefits remain unverified.

Questions for authors
  1. What are the corrected paired confidence intervals when each benchmark is resampled at its actual evaluated size? Do the smaller gains persist across repeated decoding runs?

  2. Under the Table 4 setup, how do zero-count-only scores and high-update scores with identical zero handling perform? How sensitive are channel rankings and accuracy to checkpoint precision and the definition of zero updates?

  3. Which choices were selected using the stated validation split, particularly the mapping range and count-slicing settings?

Reviewer_MBpxOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes QuantLRM, a post-training quantization method for large reasoning models that leverages weight changes introduced during fine-tuning. The authors hypothesize that weights with either very small or very large updates are particularly important for preserving reasoning capabilities after quantization. Based on this hypothesis, QuantLRM uses piecewise quadratic functions to map weight-update magnitudes into importance scores, further incorporating the number of unchanged weights within each channel. These scores guide channel-wise scaling during quantization. Experiments across Llama, Qwen, and Olmo models, covering SFT, DPO, and RLVR fine-tuning, demonstrate improvements over existing quantization methods on several reasoning benchmarks, particularly under 3-bit quantization.

Critical strengths
  1. A useful perspective on quantizing fine-tuned reasoning models.

The paper explores fine-tuning weight updates as an additional source of information for quantization, rather than relying exclusively on the final model’s weights or activations. The results in Table 1 suggest that these signals can provide complementary information for identifying weights worth preserving, offering a potentially useful direction for reasoning-model compression.

  1. Practical improvements across multiple model families and fine-tuning strategies.

Table 2 demonstrates improvements over competitive PTQ baselines across five reasoning models, with particularly notable gains under 3-bit quantization. The method also introduces limited offline overhead while maintaining inference efficiency comparable to AWQ (Table 3), supporting its practical applicability.

Critical weaknesses
  1. Insufficient validation of the proposed weight-importance interpretation.

The central hypothesis is that both small and large fine-tuning updates indicate weights that are particularly important for quantization. Although Tables 1 and 4 show that protecting both ends outperforms alternative update-based mappings, these results primarily establish the effectiveness of the proposed heuristic rather than directly validating the claimed relationship between update magnitude and quantization sensitivity. In particular, the interpretation that minimally updated weights are already well-optimized remains insufficiently supported by direct evidence.

  1. The reliability and generality of the reported improvements require further clarification.

While Table 2 reports consistent average improvements, the gains vary substantially across models and benchmarks, with only a 0.22-point advantage over AWQ on Olmo-3-7B-Think-DPO. More importantly, the significance tests in Section 5.1 and Appendix C use 50,000 bootstrap draws per benchmark despite the substantially smaller evaluation sets. Such resampling does not increase the number of independent observations, and the reported confidence intervals may therefore overstate the precision of the performance estimates. This raises concerns about the statistical support for the smaller improvements.

Questions for authors
  1. Can the authors provide a more direct analysis of the relationship between fine-tuning update magnitudes and quantization sensitivity? In particular, what evidence supports the interpretation that minimally updated weights are especially important to preserve?
  2. Could the authors clarify the bootstrap procedure in Appendix C and report confidence intervals obtained by resampling the original evaluation instances? This would help assess whether the smaller improvements remain statistically reliable.

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

Related papers

Discussion (0)

Sign in to comment.