acceptodds
Under review as a conference paper at ICLR 2027

How Much do Language Models Regurgitate under Post-Training Quantization?

Abstract

Quantization has become a standard practice for efficient deployment of today's large language models. However, there is still a lack of understanding of its impact on verbatim regurgitation—where models may regurgitate training data samples, which can be a privacy vulnerability. In this work, we present the first comprehensive study of how post-training quantization affects verbatim regurgitation in LLMs, uncovering multiple key insights. First, we show that reducing precision consistently lowers regurgitation rate across all model sizes, but at the cost of decreased downstream performance. Second, we find that for the same set of sequences, larger quantized models regurgitate more than smaller models, consistent across varying prefix & suffix lengths. We also propose a scaling law to model verbatim regurgitation under quantization. Lastly, fine-grained layer-wise and component-wise analyses reveal a consistent trend: quantizing attention modules leads to a higher decrease in regurgitation compared to feed-forward networks, and quantizing later layers has a more pronounced impact than early-middle layers, further supported by flip rate analyses. Our empirical study covers three modern post-training quantization methods and six models from the Pythia suite, providing novel insights into the impact of quantization on verbatim regurgitation.

Primary area:
interpretability and explainable AI
Keywords:
Verbatim Regurgitation, Quantization, Language Models

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.80 over 5 reviews, above 27% of the others.
Reviewer_sjGkOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

The authors study verbatim regurgitation as a function of PTQ bit widths. The findings are that larger models have more regurgitation, and that aggressive quantization reduces regurgitation. The experiments are done on the outdated pythia suite, why not OLMo2 or 3? The experiments do not control for general degradation: if low-bit PTQ degrades general capabilities of a model, it does not mean anything that verbatim regurgitation is 0. Overall, I do not see any verbatim regurgitation specifics.

Critical strengths
  • The paper is easy to follow along, and the flip-rate analysis is interesting.
Critical weaknesses
  • Study limited to the Pythia family, which is outdated for all of the models of the family are trained on the exact training corpus of ~300B tokens. This means that all of the models are in a heavily under-trained regime relative to modern LLMs. Moreover, if all of the model sizes are trained on the same number of tokens, the relative under-training confounds the model size experiments. Why not OLMo2 or 3?
  • There is no control for regurgitation-specific behavior in the main body. I would expect very similar analysis to hold for general PTQ-induced degradation. The same confound appears in the comparison between attn vs MLP, moreover, these do not have the same number of parameters.
  • I wonder if the analysis that the later layers are more sensitive to PTQ is an artifact of Pythia, or whether it holds as a general statement. Is this true for other models?
  • The setting under study is particularly narrow, and it is unclear why the relationship between PTQ and verbatim regurgitation is of particular interest.
  • An idea to isolate regurgitation specific degradation versus general degradation, with robustness on a per-layer basis would be to use a mixed-precision allocator à la dynamic HIGGS [malinovskii] twice under the same bit budget, once with per-layer coefficients estimated on the NLL of memorized sequences and once on general language modelling NLL. Analysing whether and where bit allocations differ would resolve to confound for me.

[malinovskii]: Pushing the Limits of Large Language Model Quantization via the Linearity Theorem, Malinovskii et al.; https://arxiv.org/abs/2411.17525

Questions for authors

Above

Reviewer_u3PCOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper investigates how post-training quantization (PTQ) affects the verbatim reproduction of training data by language models. The authors conduct experiments using six Pythia models of varying sizes, multiple quantization precision levels, three quantization methods, and a range of prefix lengths.

The results indicate that lower quantization precision generally reduces verbatim regurgitation but also degrades downstream task performance. Larger quantized models tend to reproduce more training sequences verbatim than smaller models. The authors also introduce Flip Rate, a metric that quantifies changes in token predictions after quantization, and propose an empirical scaling law for predicting regurgitation rates based on model size, precision, and prefix length.

Finally, the authors examine the effects of quantizing individual model components and layers. They find that quantizing attention modules has a greater effect on regurgitation than quantizing feed-forward networks, while quantizing later layers has a greater effect than quantizing earlier layers.

Critical strengths

1. Comprehensive empirical investigation

The paper provides a systematic study of how PTQ affects verbatim regurgitation. The experiments cover multiple model sizes, precision levels, quantization methods, and prefix lengths. The results provide useful evidence about how quantization changes memorization behavior and the trade-off between reducing regurgitation and preserving model performance.

2. Useful analyses of model components and scaling behavior

The layer-wise and component-wise experiments provide interesting observations, particularly that quantizing attention modules and later layers tends to reduce regurgitation more strongly. The proposed scaling law also achieves good prediction accuracy on held-out model sizes within the Pythia family. These findings may help researchers better understand the relationship between quantization and memorization.

Critical weaknesses

1. Limited technical novelty and theoretical contribution

Although the empirical study is comprehensive, the technical contributions are relatively limited. The paper mainly characterizes the effects of existing quantization methods rather than providing a deeper explanation of why these effects occur. In particular, the proposed Flip Rate is essentially a measure of how often quantization changes the most likely predicted token. Its relationship to exact-match regurgitation under greedy decoding is fairly straightforward. Moreover, the theoretical analysis shows that an upper bound on Flip Rate increases as precision decreases, but this does not establish that the actual Flip Rate must increase. As a result, the theoretical contribution is relatively modest, and the paper's main value comes from its empirical observations.

2. Limited generalizability and insufficiently controlled experiments

All experiments use models from the Pythia family, making it unclear whether the findings generalize to other model architectures or training datasets. Although the scaling law performs well when predicting held-out model sizes, Appendix G.3 shows that prediction errors increase substantially when INT3 or INT4 is excluded from fitting. This raises concerns about its reliability across precision levels. The component-wise analysis also has limitations. The experiments compare quantizing attention modules with quantizing feed-forward networks, but they do not adequately control for differences in the number of quantized parameters or the resulting quantization errors. Therefore, the stronger effect of attention quantization does not necessarily mean that memorization depends more strongly on attention modules.

These limitations weaken the generality of the conclusions and make the component-wise findings difficult to interpret.

Questions for authors

1. Interpretation of component-wise results

Could the authors provide additional experiments that better control for differences between attention and feed-forward modules, such as the number of quantized parameters or the magnitude of quantization errors? This would help determine whether attention modules are particularly important for regurgitation or are simply more sensitive to quantization.

2. Practical significance of the theoretical analysis

it is quite important issue. The theoretical results establish an upper bound on Flip Rate rather than its actual behavior. Could the authors evaluate how closely this bound matches the observed Flip Rate across different precision levels and quantization methods? This would help clarify how much the theoretical analysis explains the empirical findings.

Reviewer_uf8rOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates how post-training quantization affects a language model's verbatim reproduction of training data. Across six scale levels of Pythia models and three quantization methods, it is found that reducing precision generally diminishes verbatim reproduction, with larger models retaining more reproducing capability. The paper further proposes an empirical formula to predict reproduction rates and analyzes the quantization sensitivity of attention modules and varying depth layers.

Critical strengths
  1. The research questions have practical significance, covering different model sizes, quantization precisions, and quantization methods, with relatively systematic experiments.
  2. Modular and hierarchical analysis is more insightful than overall repetition rate changes; the proposed morpheme reversal rate also facilitates low-cost analysis of verbatim repetition behavior.
Critical weaknesses
  1. The representativeness of the evaluation samples is limited. The primary experiments only used samples that all original models could reproduce, thus measuring the retention of existing reproduction capabilities rather than reflecting newly extractable samples or overall leakage risks after quantization. The strict exact match metric also cannot rule out the possibility of partial content leakage.
  2. Insufficient support for some core conclusions. When the prefix length changes, the suffix simultaneously shortens, making it impossible to distinguish between the impact of increased context and the shortening of the generation target. Additionally, when INT3 or INT4 is reserved in the appendix, the prediction error of the empirical formula reaches approximately 20–25 percentage points, limiting its predictive value.
Questions for authors
  1. Does the conclusion still hold on random training samples or on replicable samples of each model itself? Will quantification generate new extractable samples?
  2. After fixing the suffix length, is the effect of prefix length still significant? How to explain the large prediction error of empirical formulas in key low bit intervals?
Reviewer_tBStOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies how post-training quantization (PTQ) affects verbatim regurgitation (VR) in LLMs. By evaluating on six Pythia models with three different PTQ methods, the authors find: (1) reducing bit-width lowers the VR rate; (2) larger models have a higher VR rate than smaller models; (3) quantizing attention modules and quantizing later layers lowers the VR rate more.

Critical strengths
  1. The paper is clearly written and easy to follow.

  2. As a primarily empirical study, the paper presents several meaningful results: (a) It defines a sequence-wise flip rate to measure whether the quantized model’s greedy token predictions deviate from those of the full-precision model, and provides theoretical support for its underlying intuition. (b) It provides a simple scaling law relating VR to model size and bit-width.

Critical weaknesses
  1. Limited validation. It is unclear whether the findings that VR is more sensitive to quantizing attention modules and later layers depend on the model architecture or the PTQ algorithm. The evidence for these findings is currently restricted to Pythia and HQQ. Results from additional model families and PTQ algorithms would strengthen these conclusions.

  2. More importantly, the paper’s contributions and novelty are not sufficiently clarified through its discussion of prior work. This concern has two aspects:
    (a) Although the authors emphasize that the effect of quantization on VR has not previously been studied, it remains unclear whether the degradation of VR is specific to quantization or reflects a general property of model compression (e.g., pruning). Moreover, how does the effect of model compression on VR differ from the effect on memorization? The similarities and differences to previous work should be discussed [1][2].
    (b) It remains unclear whether the stronger sensitivity of VR to quantizing attention and later layers reflects a VR-specific effect or general numerical sensitivity. The relationship to prior work on fine-grained quantization sensitivity should be discussed more clearly.

[1] Zhang et al., Catastrophic Failure of LLM Unlearning via Quantization, ICLR 2025

[2] Ni et al., Controllable Memorization in LLMs via Weight Pruning, EMNLP 2025

Questions for authors

The weaknesses above constitute my main concerns. In addition, I have the following questions and suggestions:

  1. The evaluation only considers sequences that the FP32 models can regurgitate exactly. Are there cases where a quantized model can reproduce a sequence that the FP32 model cannot?

  2. The exact-match definition follows prior work, but does not capture partial reproduction of sensitive information. Do the findings still hold if evaluating the regurgitation of specific privacy-sensitive or key-information patterns rather than the full sequence? Can quantized models that fail to reproduce an entire sequence still reproduce such information?

  3. Please add citations when the model family and the PTQ methods are first introduced by name in the Introduction.

Reviewer_763DOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper studies how post-training quantization affects verbatim regurgitation in six Pythia models. Using 338,440 sequences that all six models regurgitate at FP32, it reports that lower precision reduces regurgitation, and that larger quantized models regurgitate more than smaller ones on the same sequences. It also proposes a Gompertz-style scaling law over precision, model size, and prefix length. Quantizing attention modules or later layers is reported to reduce regurgitation more than quantizing FFN modules or early/middle layers. The paper introduces “Flip Rate” as a cheap teacher-forcing proxy for exact greedy regurgitation, with a theoretical upper bound on expected flip rate under quantization.

Critical strengths

New and relevant question: the effect of PTQ on memorization/privacy has not been studied, and the setup is clean (shared sequence set, three seeds, Pythia with released extractable sequences). Flip Rate is a practical idea. The equivalence to exact greedy regurgitation (Proposition 1) is correct and allows cheap, parallel evaluation. The KL-divergence counterexample is a useful motivation. Appendix is thorough, including scaling-law form comparisons, leave-one-out validation, and a frank discussion that some gaps are small.

Critical weaknesses

The main finding is largely expected: lower precision degrades the model, so exact-match regurgitation falls along with accuracy. The paper frames this as a privacy benefit but never measures it. There is no analysis of approximate leakage (edit distance, partial suffix match, sampling instead of greedy), so the privacy conclusion may be overstated. The “larger models regurgitate more” claim is confounded. The sequence set is conditioned on all six models regurgitating at FP32, and larger models likely just tolerate quantization noise better (more redundancy, larger logit margins). This is robustness to quantization, not more memorization, and the paper does not separate the two. Theorem 1 seems rely on an unestimated Lipschitz constant and a symmetric uniform quantizer, which does not describe HQQ or GPTQ. Corollary 1 is trivial, and the bound is never compared to measured flip rates. The component and layer findings are weakly supported. The paper admits the gaps are “not significant for all instances,” and there are no confidence intervals or tests. Attention vs. FFN is not parameter-matched, and 2.8B and 6.9B are excluded without explanation. The scaling law has limited credibility. It has six free parameters fit on a small grid, with one parameter at a boundary valueand an ad hoc form. Leave-one-out on INT3/INT4 gives 20–25% error, and since most points are saturated at 0 or 100%, a high R^2 is not informative.

Questions for authors

Does quantization reduce approximate memorization too (e.g., partial suffix overlap, or sampling with temperature), or only exact greedy match? Can you report this? Can you separate the model-size effect from robustness to quantization noise, for example by matching relative quantization error or logit margin across sizes? How was the Lipschitz constant handled, and how does the Theorem 1 bound compare quantitatively to observed flip rates? Does the symmetric uniform assumption hold for HQQ/GPTQ? Why are 2.8B and 6.9B excluded from the component- and layer-wise analyses? Can you add statistical tests and parameter-matched comparisons (attention vs. FFN)? How sensitive are the results to group size, calibration data, and the quantization method at INT3/INT2? Do the trends hold for GPTQ and BitsAndBytes beyond INT4/INT8? How robust is the scaling law to alternative parameterizations, and what is the evidence that c2 = 10 is not an artifact of constraining the optimizer? Do findings transfer to other families or to a model trained with different data, given that Pythia is the only family tested?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.