acceptodds
Under review as a conference paper at ICLR 2027

Fine-Tuning Leaves Clues: Calibration-Free Quantization via Task-Aware Basis

Abstract

Fine-tuning foundation models has become the standard way to create domain-specific expert models. While calibration-free quantization offers an efficient deployment solution without requiring downstream data, existing methods often suffer significant performance drop at low bit-widths. These methods typically neglect the task-specific information encoded in the task vector, which is defined as the parameter difference between the fine-tuned and pre-trained weights. In this work, we propose **T**ask-aware bas**I**s to **G**uid**E** calibration-f**R**ee **Q**uantization (TIGER-Q). TIGER-Q extracts a task-aware basis from the task vector to serve as a data-free proxy for the Hessian matrix. This basis establishes a task-sensitive subspace that guides the quantization process to minimize error along dimensions critical to the target task. To efficiently solve this objective, we introduce a decoupled continuous-discrete refinement pipeline. It computes closed-form analytical solutions for continuous quantization scales, while employing a local search for discrete weights, coupled with a row-wise acceptance mechanism to maintain monotonic error reduction. Extensive experiments show that TIGER-Q achieves superior performance over state-of-the-art calibration-free methods on both language and vision models.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
calibration-free quantization, PTQ

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_piyoOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper uses the task vectors (the difference between pre-trained and fine-tuned weights) to improve post-training data-free quantization to 3 and 4 bits.

Critical strengths
  1. The method seems effective, cheap, and novel.
  2. The paper is well-written and well-motivated.
Critical weaknesses
  1. It is not clear whether we pay some hidden cost by using the task vectors in the loss. Specifically, how does performance change on other fine-tuning tasks (for which we don't have task vectors), or on the pre-training loss? i.e., does the task vector cause the model to over-align to the specific fine-tuning tasks it was trained on, so we forget the pre-trained model more or degrade out-of-distribution? Do we need to tune alpha to make sure this doesn't happen too much?

  2. It would be good to explain more why this is a practically interesting setting (the setting in which we have access to the task vectors, but not the fine-tuning data).

Questions for authors

I would appreciate some experimental results to clarify the above weakness.

Minor comments: • From figure 2a, why should we choose \alpha>0 at all?

• In figure 3b, I wonder what happens for larger \beta ? Does the red line (L/14) start decreasing, or not? If not, then for L/14, we should just use the initial scale and not change it? For B/32 we \beta=0 seems optimal. Any idea what causes this difference?

• Why use SVD (eq. 9) and not the original objective (eq. 9)? Is it for mathematical convenience only, or does it help accuracy? show ablation if there is some impact.

• Table 7: which gray background? I don't see any gray background.

• Figure 1 was not very clear to me when I first read it. e.g., why is the quantization error on the diagonal of x_i and x_j? And why does minimizing this error fail to reach the green region? How does the rotation make the region reachable? Maybe better to put this figure after the method is defined mathematically.

• In table 3, the difference from Hadamard+RTN is negligible in 3-bit L/14 and 4-bit B/32. Also, in Table 10 4bit and table 11, comparison with other methods doesn't show a clear win. Any idea why, in these specific cases, the gain is so small or nonexistent?

5 comments
AuthorsOpenReview ↗

Response to Reviewer piyo (Part 1/2)

Thank you for your time with our work. All your concerns are carefully addressed as follows. We sincerely hope the revised paper by taken into these valuable comments would meet the expectations of the reviewer, and we are glad to provide further responses if needed.

Weakness 1: Quantization performance beyond the fine-tuning task

As suggested, we evaluate WizardMath-7B at 3 bits with groups of 128 weights, and report GSM8K accuracy for the target task and perplexity (PPL) on WikiText-2 and C4 for language modeling tasks (out-of-distribution and non-target tasks) in the table below.

TIGER-Q improves GSM8K accuracy significantly over SINQ. On both general text corpora, its PPL remains close to SINQ and lower than RTN. This shows that TIGER-Q retains the general text performance of a strong baseline while improving accuracy on the target task. At fixed β=0.02\beta=0.02, changing α\alpha has little effect on PPL on either corpus. At fixed α=1/(MN)\alpha=1/(MN), increasing β\beta from 0.02 to 0.2 lowers PPL on both corpora, while GSM8K accuracy also rises from 38.59% to 39.42%. We have added this analysis in our revised version (Appendix B.5)

MethodGSM8K Accuracy$\uparrow$ (%)WikiText-2 PPL ↓\downarrowC4 PPL ↓\downarrow
FP55.346.688.75
RTN32.528.5611.30
SINQ36.397.8310.51
Ours (α=0\alpha=0, β=0.02\beta=0.02)37.537.9510.75
Ours (α=1/MN\alpha=1/{MN}, β=0.02\beta=0.02)38.597.9510.75
Ours (α=10/MN\alpha=10/{MN}, β=0.02\beta=0.02)39.047.9610.75
Ours (α=1/MN\alpha=1/{MN}, β=0.2\beta=0.2)39.427.8110.54

Weakness 2: Explanation of the calibration-free setting

We would like to clarify that our design can address a practically relevant deployment scenario. With the growing push toward open-source foundation models, organizations often fine-tune public foundation models to their business tasks rather than train models from scratch. For example, organization A may fine-tune a public base model on proprietary data and release the resulting model weights while keeping its data confidential. Organization B or an individual deploying this model can obtain both the base and fine-tuned weights, but cannot access organization A’s private fine-tuning data. In this scenario, the deployer can compute the task vector directly from the available weights and use it to guide quantization without accessing the confidential dataset. Therefore, we believe it offers the open-source community a novel insight for calibration-free quantization. We have added this explanation in the revised version (Line53-69).

AuthorsOpenReview ↗

Response to Reviewer piyo (Part 2/2)

Question 1: Explanation and analysis of α\alpha and β\beta

  • About α\alpha: It is the most basic optimization objective at the beginning of our exploration, which is to minimize the weight quantization error. We retain this term as we step-by-step develop our method. Based on empirical experience, we set α>0\alpha > 0.
  • About β\beta: We further evaluate two larger β\beta (100 and 1000) on 3-bit ViT-L/14 in the table below. The results show ViT-L/14 accuracy stays nearly flat at the two larger β\beta values. Nevertheless, we should still adjust the scales for ViT-L/14, since table 4 in our submission (i.e., the result of β=∞\beta=\infty in the table below) shows that fixing the scales lowers accuracy from 90.94% to 90.41%. ViT-B/32 performs better with a smaller β\beta. These results suggest that the best value of β\beta can vary by model. For any finite β\beta, we still refine the scales, and β\beta controls how far they move from their initial values. Scale refinement helps both models, but the amount of adjustment that works best may differ.
ViT-L/14Average ↑\uparrow (%)
β=0\beta=086.16
β=1\beta=190.94
β=100\beta=10091.14
β=1000\beta=100091.14
β=∞\beta=\infty (Using initial scales)90.41

Question 2: Explanation of SVD and the original objective

Eq. (9) gives the idea behind our method. It uses the Gram matrix of the task vector to give more weight to quantization errors along directions that matter for the task. We use the SVD form in Eq. (11) because it shows each direction and its weight separately, making the objective function more intuitive and easier to understand. When all singular components are kept, the Gram form and the SVD form are mathematically equivalent. With the same coefficients and regularizers, both give the same solution in theory.

Question 3: The gray background

The gray background refers to the results for "baseline w. TIGER-Q". We have explicitly clarified this in the revised version (caption of Table 7).

Question 4: Clarity of Figure 1

Figure 1 is a simple sketch of the idea. We draw the quantization error on a diagonal to show an example with a component along both axes. The green region represents small errors along directions that matter for the task. Among the available quantized weights, an update can reduce the total error while leaving a large error along one of these directions. The mentioned diagonal path illustrates one such case. The task basis lets us give different weights to the error components. This changes which updates the objective favors and can guide the search toward the green region.

Question 5: Clarification about some modest cases

At these higher bit width settings, quantized models often approach full precision performance. This leaves less room for refinement and narrows the gaps between methods. It is worth noting that our TIGER-Q can also improve accuracy by refining the weights produced by Hadamard+RTN. For example, on ViT-L/14 at 3 bits, this raises average accuracy from 90.92% to 92.41%, a gain of 1.49 percentage points (Table 5 of our submission).

Dear Reviewer piyo,

Hope this message finds you well.

We truly appreciate the time and effort you’ve already dedicated to reviewing our submission. As the discussion period is nearing its end, we would like to check if our rebuttal and additional experiments have satisfactorily addressed your concerns. If you have the opportunity to take a look at our response, we would be grateful for any further thoughts or clarifications you might wish to share.

Thank you again for your contribution to the review process.

Best regards,

The Authors

Reviewer_Q6CxOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors propose TIGER-Q, a novel Post-Training Quantization method that uses task vectors to construct a task-aware basis to minimize task-aware error. They extract a task-aware basis from the task vector as a data-free proxy for the Hessian and design a decoupled continuous-discrete weight refinement pipeline. They empirically evaluate their method on vision and language models and show that TIGER-Q performs better than the baselines in preserving task-specific performance under aggressive quantization settings.

Critical strengths
  1. Using the parameter difference between the fine-tuned and pre-trained model weights as a data-free proxy for the Hessian matrix is novel.
  2. The paper is well written and is easy to follow
  3. Their method achieves significant gains (upwards of 11.62% improvements) at extremely low bit-widths and can be used as a plug-in for other calibration-free methods.
Critical weaknesses
  1. Their method shows minimal or marginal improvements at the 4-bit level, which is a standard quantization level used in the industry
  2. Their method assumes the availability of a base model and its fine-tuned version. Sometimes both are not available. Also, if someone can fine-tune a model, why not use the data used for fine-tuning to compute the Hessian?
  3. The core justification that connects the task vector's Gram matrix to the input covariance relies on SGD updates. How does this vary with optimizers like Adam?
  4. Have you tried 2 bits or ternary quantization?
  5. Can you compare your method with rotation based methods such as Quip or Spin quant?
Questions for authors

See weaknesses.

6 comments
AuthorsOpenReview ↗

Response to Reviewer Q6Cx (Part 1/2)

Thank you for your time with our work. All your concerns are carefully addressed as follows. We sincerely hope the revised paper by taken into these valuable comments would meet the expectations of the reviewer, and we are glad to provide further responses if needed.

Weakness 1: Marginal improvements at 4-bit

We would like to clarify this concern by detailing the performance trend across different bit-widths: 1) Performance saturation at 4-bit: Since the larger representational capacity at 4-bit allows most calibration-free methods to reach performance saturation, the performance gaps among different methods naturally appear narrower; 2) Substantial improvements at lower bit-widths: Although performance saturates at 4-bit, existing baselines suffer from a more severe drop in performance at lower precision (e.g., 3-bit and 2-bit), rendering them less practical at lower bit-widths. In contrast, TIGER-Q achieves significant gains in these settings.

By significantly mitigating the performance degradation at low bit-widths, we believe TIGER-Q takes a positive step toward making calibration-free quantization more effective and reliable.

Weakness 2: Motivation of calibration-free scenarios

We would like to clarify that our design can address a practically relevant deployment scenario. With the growing push toward open-source foundation models, organizations often fine-tune public foundation models to their business tasks rather than train models from scratch. For example, organization A may fine-tune a public base model on proprietary data and release the resulting model weights while keeping its data confidential. Organization B or an individual deploying this model can obtain both the base and fine-tuned weights, but cannot access organization A’s private fine-tuning data. In this scenario, the deployer can compute the task vector directly from the available weights and use it to guide quantization without accessing the confidential dataset. Therefore, we believe it offers the open-source community a novel insight for calibration-free quantization. We have added this clarification in the revised version (Line53-69).

AuthorsOpenReview ↗

Response to Reviewer Q6Cx (Part 2/2)

Weakness 3: Applicability of the theoretical justification beyond SGD

The RoBERTa models in Table 2 are all fine-tuned with AdamW. At 3 bits, TIGER-Q achieves an average score of 70.12 across eight GLUE tasks, 11.62 points above SINQ (58.50). This shows that task vectors can still guide quantization effectively when the models are fine-tuned with AdamW. We have clarified it in the revised version (Line315). Our SGD analysis motivates the use of task vectors. Prior studies [1*, 2*] likewise use plain gradient updates in their analyses and report gains on models fine-tuned with adaptive optimizers like Adam or AdamW.

Weakness 4: 2-bit and ternary evaluation

As requested, we add 2-bit results for WizardMath, WizardCoder, and ViT-B/32 in the table below. These results show that TIGER-Q can effectively mitigate the performance drop at lower bit-widths where baselines suffer from significant performance drops. Ternary quantization uses a different format from those TIGER-Q is designed for. We would like to leave it for exploration in future work.

MethodWizardMath-13B (GSM8K/MATH)WizardCoder-13B (HumanEval/MBPP)Average
SINQ25.09/2.3440.24/35.8025.87
TIGER-Q (Ours)37.38/5.5242.68/44.0032.40
MethodCarsDTDEuroSATGTSRBMNISTRESISC45SUN397SVHNAverage
SINQ0.725.6931.1112.6686.458.270.7832.8222.31
TIGER-Q (Ours)1.0320.2180.5659.7797.6038.174.4180.9747.84

Weakness 5: Comparisons with Rotation-based baselines

As requested, we add the 3-bit results using QuIP on WizardMath-7B (for GSM8K/MATH) and WizardCoder-7B (for HumanEval/MBPP). It should be noted that mentioned QuIP and SpinQuant belong to a different class of quantization methods from our TIGER-Q and are therefore not our main baselines. Both of QuIP and SpinQuant use calibration data, and SpinQuant further requires training to learn its rotations. TIGER-Q is a calibration-free and training-free method without calibration data. Therefore, we use other calibration-free methods as our main baselines.

MethodGSM8KMATHHumanEvalMBPP
QuIP40.257.0443.9043.40
SINQ36.395.5439.6339.20
TIGER-Q (Ours)38.596.7042.6841.80

Reference

[1*] Cheng R et al., Whoever Started the interference Should End It: Guiding Data-Free Model Merging via Task Vectors, ICML 2025.

[2*] Wang S et al., LoRA-GA: Low-Rank Adaptation with Gradient Approximation, NeurIPS 2024.

Dear Reviewer Q6Cx,

Hope this message finds you well.

We truly appreciate the time and effort you’ve already dedicated to reviewing our submission. As the discussion period is nearing its end, we would like to check if our rebuttal and additional experiments have satisfactorily addressed your concerns. If you have the opportunity to take a look at our response, we would be grateful for any further thoughts or clarifications you might wish to share.

Thank you again for your contribution to the review process.

Best regards,

The Authors

Reviewer_C6d4OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper targets at calibration-free quantization. It mainly establishes a task-aware basis from the task vector to form a task sensitive subspace and then guide the quantization process. Extensive experiments on computer vision and language tasks demonstrate their paper's effectiveness.

Critical strengths
  1. For experimental results: it brings large gains at very low bit-widths on small models. Examples: 2-bit Flan-T5 (79.76 vs 67.35), 3-bit RoBERTa (70.12 vs 58.50).

  2. The proxy is evaluated by direct comparison, but not only through accuracy. Table 6 / Table 9 compare τᵀτ against the measured E[xxᵀ] using cosine similarity and eigenspace overlap, alongside identity and random controls.

  3. Ideas: It would be useful as the task vector is free and always present when you deploy a fine-tune of a known base model, and calibration-free methods ignore it.

Critical weaknesses
  1. LLM results are weak sometimes. For example, HumanEval-7B at 4-bit: 52.44 vs 53.66 (Hadamard+RTN). HumanEval-13B at 4-bit: 61.59 vs 64.63 (HQQ).

  2. Missing baselines and comparisons. for example,

  • GPTQ with τᵀτ as its Hessian. This is the most natural comparison and ablation. Then it would isolate the contribution of the proxy from that of the custom solver. Without it, the case for the new solver is unclear.
  • GPTQ with real data as a calibration-based reference, alongside AWQ.
  • The data used for AWQ is not stated.
  1. Some results not consistent. for example, the main GLUE table shows 2-bit for Flan-T5 but 3-bit for RoBERTa, and the less favourable settings are moved to the appendix.
Questions for authors

Please response to the weakness i mentioned.

Also, i doubt the motivation here. As most time, we have the calibration dataset, so scenarios on no calibration dataset is rare. Also, usually recalibration does not take longer time.

Also, the authors' comparion latest work is only one from 2025. Can also compare latest work that has calibration so that can prove it works well to replace the relibration perfectly.

6 comments
AuthorsOpenReview ↗

Response to Reviewer C6d4 (Part 1/2)

Thank you for your time with our work. All your concerns are carefully addressed as follows. We sincerely hope the revised paper by taken into these valuable comments would meet the expectations of the reviewer, and we are glad to provide further responses if needed.

Weakness 1: Some weak LLM cases

The mentioned cases are among the few exceptions in our broader LLM evaluation. In Table 1, our TIGER-Q achieves the best calibration-free score in 13 of 16 LLM settings and attains the highest average performance among calibration-free methods at both 3 and 4 bits. Therefore, we believe it supports the advantage of our TIGER-Q across the evaluated models, tasks, and bit-widths.

Weakness 2-1: GPTQ with τ⊤τ\tau^\top\tau as its Hessian

In the table below, we provide the suggested comparisons by substituting τ⊤τ\tau^\top\tau for GPTQ’s Hessian estimate at 3-bit settings. These results indicate that GPTQ is not well suited to solving the quantization objective defined by our proposed proxy, highlighting the importance of compatibility between the proxy and the solver.

Our custom solver successfully exploits the same proxy, supporting its effectiveness in utilizing the proposed task-aware information to achieve calibration-free quantization improvements.

MethodSUN397CarsRESISC45EuroSATSVHNGTSRBMNISTDTDAverage
GPTQ57.6148.3691.2499.8197.1298.3499.6263.9482.01
GPTQ with τ⊤τ\tau^\top\tau8.267.0845.6896.6395.4296.0799.4434.8960.43
TIGER-Q (Ours)54.5140.7488.3598.1195.4997.1199.5363.0379.61

Weakness 2-2: GPTQ with real data as a calibration-based reference

As requested, we add the results using GPTQ on ViT-B/32, WizardMath (for GSM8K/MATH), and WizardCoder (for HumanEval/MBPP) at 3-bit settings. It should be noted that GPTQ belong to the calibration-based quantization methods and are therefore not our main baselines. TIGER-Q is a calibration-free method without calibration data. We therefore use GPTQ and AWQ as reference and use other calibration-free methods as our main baselines. We have added these GPTQ results in Table 1 and Table 3 of our revised version.

MethodSUN397CarsRESISC45EuroSATSVHNGTSRBMNISTDTDAverage
GPTQ57.6148.3691.2499.8197.1298.3499.6263.9482.01
SINQ41.3223.6983.7097.6394.9495.3599.5255.8574.00
TIGER-Q (Ours)54.5140.7488.3598.1195.4997.1199.5363.0379.61
MethodGSM8K (7B/13B)MATH (7B/13B)HumanEval (7B/13B)MBPP (7B/13B)
GPTQ40.49/55.806.74/10.7843.29/58.5444.20/52.60
SINQ36.39/51.185.54/8.5039.63/57.3239.20/51.80
TIGER-Q (Ours)38.59/54.736.70/9.4842.68/59.7541.80/52.20

Weakness 2-3: The data used for AWQ

For WizardMath models, we use calibration data from the GSM8K and MATH training splits for their respective evaluations. For WizardCoder models, we use calibration data from the MBPP training split. For ViT-B/32 and ViT-L/14 models, we use training images from the corresponding dataset for each benchmark. We have clarified these calibration sources in the revised version (Line344-348).

AuthorsOpenReview ↗

Response to Reviewer C6d4 (Part 2/2)

Weakness 3: Choice of quantization settings

Models differ in their sensitivity to quantization, so we evaluate each model under two different regimes, i.e., a lower bit-width and a higher bit-width. At higher bit-widths, quantized performance often approaches full precision, narrowing the gaps between methods. At lower bit-widths, larger quantization errors make it easier to see how well each method preserves task performance. To assess TIGER-Q in both regimes, we evaluate Flan-T5 at 2 and 3 bits and RoBERTa, WizardMath, WizardCoder, ViT-B/32, and ViT-L/14 at 3 and 4 bits. The main tables report both bit-widths for WizardMath, WizardCoder, and ViT-B/32. Due to the page limit, the higher-bit results for other models are provided in Appendix B.4. Across these comparisons, TIGER-Q achieves the highest average performance among calibration-free methods at both evaluated bit-widths.

Question 1: Motivation of calibration-free scenarios

We would like to clarify that our design can address a practically relevant deployment scenario. With the growing push toward open-source foundation models, organizations often fine-tune public foundation models to their business tasks rather than train models from scratch. For example, organization A may fine-tune a public base model on proprietary data and release the resulting model weights while keeping its data confidential. Organization B or an individual deploying this model can obtain both the base and fine-tuned weights, but cannot access organization A’s private fine-tuning data. In this scenario, the deployer can compute the task vector directly from the available weights and use it to guide quantization without accessing the confidential dataset. Therefore, we believe it offers the open-source community a novel insight for calibration-free quantization. We have added this clarification in the revised version (Line53-69).

Question 2: Baselines from 2026 and calibration-based baselines

We would like to clarify two points here:

  1. We do include a latest baseline from 2026, i.e., SINQ (ICML 2026), which is a recent state-of-the-art calibration-free method.

  2. We respectfully clarify that we never claim our method can replace calibration-based methods, and demonstrating such a replacement is beyond the scope of this work. Our primary goal is to advance calibration-free quantization by introducing a novel perspective to use task-specific information. Therefore, calibration-free methods are the most directly relevant comparisons. Calibration-based methods just provide calibration-based performance references.

Dear Reviewer C6d4,

Hope this message finds you well.

We truly appreciate the time and effort you’ve already dedicated to reviewing our submission. As the discussion period is nearing its end, we would like to check if our rebuttal and additional experiments have satisfactorily addressed your concerns. If you have the opportunity to take a look at our response, we would be grateful for any further thoughts or clarifications you might wish to share.

Thank you again for your contribution to the review process.

Best regards,

The Authors

Comments on the paper

2 comments

Dear Reviewers,

Thank you for the review. The authors have responded to the points you raised, and the response period closes on 18 November (AoE). I would be grateful if you could post a brief reply on whether your concerns are resolved, partly resolved, or still open, so that the authors have a fair chance to address them. If anything still needs clarifying, please ask now while they can still respond.

Regards, AC.

We appreciate the time and effort of the AC and all reviewers in reviewing our submission and engaging in the discussion. As the discussion period comes to a close, we would like to briefly summarize the reviewers’ feedback and our responses.

Reviewer piyo (initial rating: 3) recognizes that "the method is effective, cheap, and novel", and that "the paper is well-written and well-motivated". After discussion, the reviewer keeps the positive score and recommends acceptance.

Reviewer C6d4 (initial rating: 2) recognizes the "large gains at very low bit-widths on small models", "the evaluation of our proxy", and "usefull ideas". Following the reviewer’s suggestions, we report experiments with calibrated GPTQ and GPTQ using our proposed proxy. We also clarify the AWQ calibration data and explain the experimental settings. In the discussion, the reviewer’s remaining concerns seem to arise because the reviewer miss our answers to the last two questions in our initial response, “Response to Reviewer C6d4 (Part 2/2)”. We point the reviewer to these answers and present it again. The reviewer has raised no further concerns.

Reviewer Q6Cx (initial rating: 2) recognizes the "novelty of the data-free Hessian proxy", "significant gains at low bit-widths, and "clear writing". Following the reviewer’s suggestions, we add 2-bit experiments and comparisons with the calibration-based QuIP. The reviewer subsequently acknowledges that the 2-bit results “show promise.” In further discussion, the reviewer’s remaining concerns mainly focus on access to calibration data, comparisons with calibration-based methods, and theory and practical deployment. We explain why third-party model publishers may lack access to user traffic and emphasize that TIGER-Q offers a safer alternative to traffic-based calibration by eliminating the need to collect or share user traffic. Regarding the theoretical analysis, we explain its motivating role and highlight experiments showing that TIGER-Q also works on the evaluated models fine-tuned by AdamW. We would like to clarify that our goal is not to replace calibration-based methods, but to advance the frontier of calibration-free quantization through a novel insight that makes these methods more practical. We therefore believe TIGER-Q makes a meaningful contribution and opens a promising direction for further research. The reviewer has raised no further concerns.

We thank the AC and all reviewers again for their time and consideration, and we hope this summary is helpful in reaching the final decision.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.