acceptodds
Under review as a conference paper at ICLR 2027

A Solvable Theory of Pre-training Data Poisoning: Regime-Dependent Scaling Exponents

Abstract

Pre-training data poisoning of large language models is usually studied using targeted backdoors and their survival through safety post-training, which leaves open a more basic question: how does a model's clean data performance degrade as the poison rate grows? Motivated by controlled pre-training runs of OLMo-style models, in which the relative clean data validation perplexity increase between poisoned and clean models matched in architecture, token budget, and optimization schedule is well fit by a power law with a non-integer exponent, we ask what such a law requires theoretically. We first prove an analyticity barrier: whenever the contaminated objective depends analytically on around a nondegenerate clean data optimum, is generically quadratic in , so a generic non-integer exponent is a signature of genuinely singular structure. We then supply that structure in solvable truncated ridge regression with heavy-tailed covariates, controlled by , and a label-shift poisoning. Our central result is that the excess risk scaling exponent depending on the order of limits: in the higher dimensional proportional regime it is , whereas taking the ample-data limit first gives , and the limits do not commute. We confirm this prediction through several numerical simulations. Finally, we argue that finite training time plays the role of truncation on the curvature spectrum in local LLM pre-training, deriving the observed scaling law under heavy tailed inverse curvature spectrum as a modeling hypothesis.

Primary area:
learning theory
Keywords:
LLM pre-training, Neural scaling laws, Data poisoning

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_SR9qOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates how pre-training data poisoning affects language models' performance on clean data, with a particular focus on the scaling relationship between the poisoning fraction (\epsilon) and relative perplexity degradation. Motivated by empirical observations of non-integer scaling exponents in OLMo-style models, the authors establish an analyticity barrier showing that smooth contamination of a nondegenerate optimum generically produces quadratic excess-risk scaling. They then construct a solvable heavy-tailed truncated regression model and derive two distinct exponents, (q_\star/(q_\star+2)) and (2-2/q_\star), depending on the order of the high-dimensional and small-poisoning limits. Numerical simulations support these theoretical predictions. Finally, the paper proposes a connection to LLM pre-training through finite-time spectral filtering and heavy-tailed inverse-curvature distributions, supported by numerical curvature measurements.

Critical strengths
  1. The paper studies aggregate clean-performance degradation rather than targeted attack success, offering a complementary perspective on pre-training data poisoning (Sec. 1). The relative perplexity metric provides a clear way to quantify degradation (Eq. (1)), and the empirical observations suggest that poisoning sensitivity depends on model size and training progress (Fig. 1). Understanding such relationships could provide useful insights into how data contamination affects large-scale pre-training.

  2. The analyticity barrier provides a principled explanation of quadratic scaling under regularity assumptions (Result 1), while the adaptive-truncation framework introduces an interpretable bias–poison trade-off that produces fractional exponents (Result 2). The random-coordinate model and Poisson-process formulation enable explicit analysis of the contamination mechanism (Sec. 4.1–4.2). Particularly interesting is the non-commutativity of the high-dimensional and small-poisoning limits, which reveals how different statistical regimes can produce distinct scaling exponents (Result 9).

  3. The numerical simulations support the predicted exponents across multiple heavy-tail parameters and aspect ratios (Figs. 2–3). The LLM experiments cover multiple model sizes, training checkpoints, and two poisoning constructions (Figs. 1, 7–10). Additionally, the connection to finite-time spectral filtering is an interesting attempt to explain poisoning sensitivity through optimization dynamics (Sec. 5). The detailed Gauss–Newton spectral analysis and numerical reliability discussions are also appreciated (Appendix E).

Critical weaknesses
  1. The experiments insert poisoned data into the original clean corpus while maintaining matched total training-token budgets. Consequently, increasing the poisoning fraction also changes the effective amount of clean data processed during training. Since reducing clean-data exposure can itself degrade language-model performance, it is unclear how much of the observed scaling behavior is attributable to poisoning-specific effects. No clean-data dilution or benign data-mixing controls are provided. Moreover, the fitted fractional exponents are not compared against simpler analytic alternatives, such as linear-plus-quadratic models (Figs. 1, 7–8).

  2. The theoretical exponents are derived by optimizing poisoning-dependent truncation thresholds, while the empirical LLM exponents are fitted at fixed training checkpoints (Secs. 4.3–5). The proposed LLM mechanism similarly relies on a poisoning-dependent crossover horizon, but this relationship is not directly validated. Furthermore, the solvable model uses specialized assumptions, including coordinate-sparse features and poison labels that depend on the truncation threshold (Appendix A.2). The key theoretical quantities (\beta) and (\gamma) are not independently measured to verify whether they predict the observed LLM exponents. No direct evidence found in the manuscript.

  3. The experiments consider only gibberish and language-switching poisoning, both constructed using similar trigger-based formats (Appendix D). These settings do not adequately distinguish random noise, distribution shift, and targeted behavioral contamination. Interestingly, the two poisoning types exhibit different fitted exponents under comparable configurations (Figs. 7–8), but these differences are not systematically analyzed. It remains unclear whether the proposed scaling mechanism is general across contamination types or primarily characterizes the specific data constructions studied. Additional controlled comparisons would substantially strengthen the empirical interpretation.

  4. The scaling laws are primarily supported by fitting curves over observed poisoning-rate ranges, without systematic out-of-sample prediction experiments (Figs. 1, 7–8). Although the fitted exponents vary across model sizes, training stages, and poisoning types, the paper does not establish how reliably these variations can be predicted from measurable properties. Potential applications, such as estimating contamination levels or determining tolerable poisoning fractions, are not investigated. No direct evidence found in the manuscript. Additionally, the non-commuting-limits result raises questions about whether the empirical scaling relationships remain valid at much smaller poisoning rates (Sec. 5; Appendix A.8).

  5. The expansion in Eq. (128) appears to omit checkpoint-offset interaction terms without explicitly establishing their negligibility, while Result 21 would benefit from clearer sign conditions for the claimed interior optimum (Appendix B.5–B.6). The empirical exponent estimates also lack sufficient uncertainty analysis across independent training seeds and fitting ranges (Figs. 1, 7–8). Furthermore, the curvature-spectrum evidence is mainly based on a single model checkpoint (Fig. 4; Appendix E), and the reported training budget for the largest model appears inconsistent across Table 2 and Fig. 8(d). Finally, the Abstract and Introduction introduce several technical exponents before adequately motivating their relevance to the empirical observations.

Questions for authors

There are several questions and suggestions for the authors:

  1. I recommend adding clean-only runs with reduced token budgets and benign data-mixing controls under matched training conditions. These experiments would help quantify how much degradation arises from reduced clean-data exposure versus poisoned content. The authors should also compare the power-law fits against simpler analytic models and examine whether fractional exponents persist over smaller poisoning-rate intervals.

  2. The authors should clarify the relationship between fixed-checkpoint empirical exponents and the theoretical predictions based on optimized truncation or crossover horizons. Measuring the predicted crossover behavior across poisoning rates and training stages would be valuable. More importantly, independently estimating the clean-bias and poison-response exponents (\beta) and (\gamma), followed by a comparison between predicted and observed LLM scaling exponents, would provide much stronger evidence for the proposed mechanism.

  3. I suggest adding trigger-free versions of the existing poisoning methods, together with additional contamination types such as benign distribution shifts, random corruption, and targeted behavioral poisoning. The authors should systematically compare both scaling exponents and coefficients across these settings. Such experiments could help determine which properties of poisoned data are responsible for the observed degradation and whether the proposed scaling behavior generalizes across different contamination mechanisms.

  4. Beyond fitting existing curves, the authors could estimate scaling parameters using a subset of poisoning rates and predict degradation at held-out rates. It would also be valuable to investigate whether model size, training horizon, or measured spectral properties can predict scaling coefficients across configurations. Discussing potential applications and limitations, particularly contamination tolerance and poison-rate estimation, would clarify the broader significance of the theoretical results.

3 comments

We thank the reviewer for the detailed review and the concrete list of questions.

W1/Q1 the poisoned schedule matched clean run. We trained these controls for 4 sizes (135M–1287M) and ε∈{0.074,0.167,0.333}\varepsilon\in\lbrace 0.074,0.167,0.333\rbrace https://anonymous.4open.science/r/pretrain-poisoning-6DE4/language_switching_1287M.png. At (1−ε)D(1-\varepsilon)D a control has seen as many clean tokens as the poisoned run at DD, at the same point of the schedule, giving the matched clean loss

Rcleanmatched((1−ε)D)=R0(D)+Rctl,ε((1−ε)D)−Rctl,0(D).\mathcal R_{\rm clean}^{\rm matched}((1-\varepsilon)D)=\mathcal R_0(D)+\mathcal R_{{\rm ctl},\varepsilon}((1-\varepsilon)D)-\mathcal R_{{\rm ctl},0}(D).

The paper's observable, Eq. (3), splits into dilution and a poison-specific part:

Δobs=eRε−R0−1,Δdil=eRcleanmatched−R0−1,Δps=Δobs−Δdil.\Delta_{\rm obs}=e^{\mathcal R_\varepsilon-\mathcal R_0}-1,\qquad\Delta_{\rm dil}=e^{\mathcal R_{\rm clean}^{\rm matched}-\mathcal R_0}-1,\qquad\Delta_{\rm ps}=\Delta_{\rm obs}-\Delta_{\rm dil}.

Our new figures are for language switching data poisoning and it plots Rε/Rcleanmatched−1=ln⁡[1+Δps/(1+Δdil)]/Rcleanmatched≈Δps/Rcleanmatched\mathcal R_\varepsilon/\mathcal R_{\rm clean}^{\rm matched}-1=\ln[1+\Delta_{\rm ps}/(1+\Delta_{\rm dil})]/\mathcal R_{\rm clean}^{\rm matched}\approx\Delta_{\rm ps}/\mathcal R_{\rm clean}^{\rm matched} against D/DHD/D_{\rm H}:

  • At D=DHD=D_{\rm H}, Δps\Delta_{\rm ps} is 31–58% of Δobs\Delta_{\rm obs} for 490M–1287M, and 53–71% at 135M; the rest is Δdil\Delta_{\rm dil}.
  • From D/DH=0.4D/D_{\rm H}=0.4 to 1, Δps/Rcleanmatched\Delta_{\rm ps}/\mathcal R_{\rm clean}^{\rm matched} grows.

Q2 / W2, independent validation of the clean-bias and poison-response laws. Our new experiments validate it https://anonymous.4open.science/r/pretrain-poisoning-6DE4/1287M_delta_dil.png and https://anonymous.4open.science/r/pretrain-poisoning-6DE4/language_switching_1287M_delta_ps.png. In the local linear model, giving the clean data weight 1−ε1-\varepsilon is the clean run slowed to time (1−ε)T(1-\varepsilon)T, so the dilution response scales with the clean bias B(T)∝T−βB(T)\propto T^{-\beta}, and Δdil\Delta_{\rm dil}, which measures it directly, decays like the clean excess risk: Δdil∝T−κ\Delta_{\rm dil}\propto T^{-\kappa} with κ=β\kappa=\beta if the train–eval mismatch dominates, otherwise κ=2β\kappa=2\beta. The poison-specific part is driven by the poison influence I(T)∝TγI(T)\propto T^{\gamma} (Result 19): the linear part of Δps\Delta_{\rm ps} grows as TγT^{\gamma} (its train–eval-gap term) and its quadratic part as T2γT^{2\gamma}. Both behave as predicted: Δdil\Delta_{\rm dil} decays with training (by 20–37% from D/DH=0.4D/D_{\rm H}=0.4 to 1); Δps\Delta_{\rm ps} grows (by 15–77%). We will update Fig. 4 with these plots in a revised version of the paper and add explanation in Sec. 5. However, we do not extract numerical β\beta and γ\gamma from the LLM runs, measuring them carefully is worthy of a full new paper like arxiv: 2607.21716. We request the reviewer to consider these as limitations.

W4/Q4. held-out prediction on 1287M model. the power law Δobs=Cεa\Delta_{\rm obs}=C\varepsilon^a fitted to the six fractions 0.074–0.333 predicts the five at 0.020–0.091 with rms error 0.169 in ln⁡Δobs\ln\Delta_{\rm obs}; Δobs=fε+gε2\Delta_{\rm obs}=f\varepsilon+g\varepsilon^2 gives 0.121 and the effective-data form Δobs=b[(1−ε)−k−1]\Delta_{\rm obs}=b [(1-\varepsilon)^{-k}-1] 0.128. Our theoretical prediction of the current paper is the form Δobs=fε+gε2\Delta_{\rm obs}=f\varepsilon+g\varepsilon^2. The held-out prediction gets better with bigger model size, see Fig. 5, 6. Please see our reply to reviewer AiMd for a detailed theoretical explanation of this.

W3/Q3. different poison types. The two poison types are compared in Appendix D. A detailed experimental study of model degradation due to data poisoning is the content of arxiv 2510.03313. Although the scaling law point of view is our own, the current paper is a theory first paper and hence we have limited experimental studies compared to a full length experimental paper. We would be grateful to the reviewer if the theoretical value of our work is appreciated.

If you find our reply satisfactory, please raise your score. Thank you again.

Thank you for the detailed experiments and explanation. However, I still have some concerns about how to transfer the theoretical analysis to LLM data contamination in practice, even though the possibility of deriving fractional exponents from heavy tails and adaptive truncation is interesting and has its unique contribution. Personally, I would find the paper considerably more compelling if the authors took the initiative to explain what these theoretical insights enable us to think about differently, rather than leaving the reader to infer the practical implications. At present, I can see several possible connections to data filtering and contamination monitoring. Since I'm not an expert, it is difficult for me to determine which of these are genuinely supported by their analysis. This would not necessarily require a new set of application experiments, but articulating more clearly how a reader and potential researchers/engineers should translate existing insights into conclusions about handling actual pre-training data poisoning.

Thank you for engaging with the conversation.

Our theoretical results suggested the existence of a scaling law for pre-training data poisoning. This motivated us to perform experiments on LLMs and validate them. As with the standard scaling law for pre-training that tells us how the model's clean data performance scales with compute, it gives us predictive power as the poison rate varies, on how the model degrades on clean data. This is the central practical output of the current theoretical considerations. Due to computational limitations, we could not perform an extensive study on 10B or larger models, but we believe our scaling law would remain valid on a 70B model up to poisoning rates that are realistic, e.g., 0.01%, for backdoor attacks (our initial Fig. 5, 6 in the Appendix show that as model size grows, the scaling law becomes more robust at small poisoning fraction).

Thank you again for the review.

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

This paper investigates how clean-data performance degrades with the poisoning rate during LLM pre-training. Motivated by empirical observations of non-integer power-law exponents, the authors establish an analyticity barrier and develop a solvable heavy-tailed regression model with adaptive truncation. Their main theoretical result identifies two distinct poisoning-sensitivity exponents depending on the order of the high-dimensional and small-poisoning limits. Numerical simulations validate these predictions, while OLMo-style pre-training experiments and curvature-spectrum measurements provide preliminary empirical support for the proposed mechanism.

Critical strengths
  1. The paper derives distinct poisoning-sensitivity exponents under non-commuting asymptotic limits, providing a novel perspective on how dimensionality and data availability influence robustness to poisoning.

  2. The proposed model enables explicit theoretical characterization of poisoning effects, and numerical simulations closely match the predicted exponents across different tail parameters and aspect ratios.

Critical weaknesses
  1. The theoretical results rely on optimized truncation and specific spectral assumptions, whereas the LLM experiments fit exponents at fixed training checkpoints. The proposed clean-bias and poison-response scaling laws are not independently validated, leaving the connection between the theory and observed LLM behavior largely hypothetical.

  2. The solvable model relies on a specialized poisoning construction whose label shift scales with the truncation threshold. Moreover, non-integer exponents fitted over a limited poisoning-rate range do not establish genuine non-analytic asymptotic behavior, as smooth crossover effects could yield similar empirical slopes.

Questions for authors
  1. Can the authors validate the predicted LLM scaling mechanism by independently estimating the clean-bias and poison-response exponents, rather than directly fitting the poisoning exponent?

  2. Does the theoretical scaling law persist when the poisoning distribution is fixed independently of the truncation threshold?

  3. Can the authors evaluate substantially smaller poisoning rates and compare the power-law fit against alternative analytic models to establish whether the fractional exponent reflects genuine asymptotic behavior?

3 comments

We thank the reviewer for the careful reading and for recognizing the non-commuting limits as the paper's novel perspective.

Q2 / W2, a poison fixed independently of τ\tau. The law does not persist, we explain it below. In the solvable model (PP dimensions, DD samples, aspect ratio ϕ=P/D\phi=P/D, covariate size UU with tail P(U>t)=t−q⋆\mathbb P(U\gt t)=t^{-q_\star}, truncation threshold τ\tau) we generalized the poison label to Yp=c+ηpτγ′Y_{\rm p}=c+\eta_{\rm p}\tau^{\gamma'} (0<c,ηp<10\lt c,\eta_{\rm p}\lt1; leverage exponent γ′\gamma', the paper has γ′=1\gamma'=1) and redid Results 6 and 8. The optimized exponents at fixed ϕ\phi and for ϕ→0\phi\to0 (ample data) are

aprop=q⋆q⋆+2γ′,aample=2(q⋆−1)q⋆−1+γ′.a_{\rm prop}=\frac{q_\star}{q_\star+2\gamma'},\qquad a_{\rm ample}=\frac{2(q_\star-1)}{q_\star-1+\gamma'}.

A τ\tau-independent poison (γ′=0\gamma'=0) gives the analytic exponents, a=2a=2 (ample) and a=1a=1 (fixed ϕ\phi). So the non-integer law requires: the adversary's leverage must grow with the trust radius. In the LLM sketch τ↔T\tau\leftrightarrow T, the training time; this is Assumption 18, which gives the poison influence I(T)∝TγI(T)\propto T^{\gamma}, Result 19. Simulations confirm both formulas: https://anonymous.4open.science/r/pretrain-poisoning-6DE4/solvable_gamma_prime.png. We thank the reviewer for this important question, we will add this discussion to the Appendix in a revised version of the paper.

Q1 / W1, independent validation of the clean-bias and poison-response laws. Our new experiments validate it https://anonymous.4open.science/r/pretrain-poisoning-6DE4/1287M_delta_dil.png and https://anonymous.4open.science/r/pretrain-poisoning-6DE4/language_switching_1287M_delta_ps.png (please see our reply to reviewer cvBf Q1 for notation). In the local linear model, giving the clean data weight 1−ε1-\varepsilon is the clean run slowed to time (1−ε)T(1-\varepsilon)T, so the dilution response scales with the clean bias B(T)∝T−βB(T)\propto T^{-\beta}, and Δdil\Delta_{\rm dil}, which measures it directly, decays like the clean excess risk: Δdil∝T−κ\Delta_{\rm dil}\propto T^{-\kappa} with κ=β\kappa=\beta if the train–eval mismatch dominates, otherwise κ=2β\kappa=2\beta. The poison-specific part is driven by the poison influence I(T)∝TγI(T)\propto T^{\gamma} (Result 19): the linear part of Δps\Delta_{\rm ps} grows as TγT^{\gamma} (its train–eval-gap term) and its quadratic part as T2γT^{2\gamma}. Both behave as predicted: Δdil\Delta_{\rm dil} decays with training (by 20–37% from D/DH=0.4D/D_{\rm H}=0.4 to 1); Δps\Delta_{\rm ps} grows (by 15–77%). We will update Fig. 4 with these plots in a revised version of the paper and add explanation in Sec. 5. However, we do not extract numerical β\beta and γ\gamma from the LLM runs, measuring them carefully is worthy of a full new paper like arxiv: 2607.21716. We request the reviewer to consider these as limitations.

Q3 / W2, non-analyticity and ε→0\varepsilon \to 0 limit. In the solvable model we have shown that the poison scaling exponent is different for ample data (d/n→0d/n\to 0 first) and higher dimensional (ϵ→0\epsilon \to 0 first) limit. d/nd/n controls the difficulty of the task and determines the smallest value of ϵ\epsilon for which ample data scaling law remains valid - as model becomes stronger, i.e., its effective d/nd/n becomes smaller we expect the ample data scaling law to be valid for smaller ϵ\epsilon. This holds true for LLMs as well as shown in Fig. 5, 6 in the Appendix. Currently we cover ϵ\epsilon from 1%-30% relevant for model degradation studies and we can see that for the smaller values variance increases presumably because of the discussion above. Due to the variance, we need several costly seed runs and we cannot currently systematically extend our study for LLMs for higher dimensional scaling law limit. For the study of backdoor attack relevant value of ϵ∼0.1\epsilon \sim 0.1% or smaller - to see the ample data scaling law at that level we need models of size 10B−20B10B-20B again that is too costly for us to perform experiment on. We will mention this in the limitation section. We sincerely hope that the reviewer sees merit on what we have already achieved given that scaling law is our original point of view along with the theoretical modeling.

Please see new the poisoned schedule matched clean run: reply to cvBf; uncertainty: reply to SR9q. If you find our reply satisfactory, please raise your score. Thank you again.

Since the response period is ending soon, we kindly ask the reviewer if we've addressed their concerns and, if so, if the reviewer would consider updating their score.

Thank you for the detailed rebuttal and additional experiments. While the new results strengthen the paper, two key concerns remain:

  • W1. Insufficient quantitative validation of the LLM mechanism. The new experiments support the predicted qualitative trends, but the key exponents β\beta and γ\gamma are not independently estimated. Consequently, the proposed mechanism has not been quantitatively validated against the observed LLM scaling exponents.

  • W2. Limited generality and insufficient theoretical–empirical connection. The authors confirm that the fractional exponents in the solvable model depend on a specialized poisoning construction. While the linear–quadratic crossover plausibly explains the observed LLM behavior, the proposed spectral mechanism and its predicted crossover conditions remain insufficiently validated.

I appreciate the improvements and recognize the theoretical contribution of the non-commuting limits. However, these central concerns remain unresolved, so I will maintain my original rating.

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

This paper presents a comprehensive study of how poisoned data in the pretraining dataset affect a model’s performance. It explains when the degradation scales quadratically with the poisoning fraction (R1), then shows how heavy-tailed data and adaptive truncation can produce fractional scaling in a convex regression model (R2). The paper further proposes a conditional connection to language-model pretraining through a local quadratic approximation. Empirically, experiments across multiple model sizes and poisoning mechanisms exhibit power law relationships between the poisoning fraction and degradation in clean validation perplexity.

Critical strengths

The paper is well motivated by empirical observations of poisoning in LLM pretraining. The empirical study and theoretical analyses are clearly presented, and the individual sections are well organized and easy to follow.

R1 and R2 appear sound under the stated assumptions. Experiments in the convex regression model also support the scaling laws derived for that setting.

Critical weaknesses
  • Could the authors clarify how the empirical setting in Figure 1 relates to the assumptions of Result 1 (R1) ? Currently, it seems like Figure 1 measures degradation relative to a clean model at checkpoints, but R1 considers the small-poisoning limit around a non-degenerate clean optimum and the connection between two components is under-explained. Is Figure 1 intended as evidence of escaping the analyticity barrier, or primarily as a motivation for the theoretical analysis?

  • Section 5 currently feels somewhat disorganized and is difficult to follow. The manuscript would benefit from a clearer distinction between assumptions, various results, empirical interpretations, as well as clearer explanations of how they connect to previous sections. For example, the section seems to derive a poisoning-dependent crossover time (Tϵ∝ϵ−1/(β+γ))(T_\epsilon\propto\epsilon^{-1/(\beta+\gamma)}) and assumes that the experimental training horizons are comparable to this scale. The authors may want to further justify or explain the assumption. In particular, could the authors clarify under what conditions the theoretical exponent in R21 is expected to match the exponents estimated at fixed training checkpoints?

Questions for authors
  • Since the clean and poisoned runs use the same total token budget, increasing the poison budget would also reduces exposure to clean tokens, is it possible that the degradation is caused by this reduction? Would it be possible to run a controlled experiment where no poisoned samples are present.
4 comments

We thank the reviewer for carefully reading our paper and the questions. We have run the central experiment you have asked as we explain below along with response to the other concerns.

Notation. ε\varepsilon: poisoned fraction of training tokens; DD: tokens seen by the poisoned run; DHD_{\rm H}: Kaplan horizon of an NN-parameter model; Rε(D)\mathcal R_\varepsilon(D): clean-validation cross-entropy of the run at fraction ε\varepsilon; Rctl,ε\mathcal R_{{\rm ctl},\varepsilon}: clean-only control run with token budget and cosine schedule both shrunk by (1−ε)(1-\varepsilon).

Q1, the poisoned schedule matched clean run. We trained these controls for 4 sizes (135M–1287M) and ε∈{0.074,0.167,0.333}\varepsilon\in\lbrace 0.074,0.167,0.333\rbrace https://anonymous.4open.science/r/pretrain-poisoning-6DE4/language_switching_1287M.png. At (1−ε)D(1-\varepsilon)D a control has seen as many clean tokens as the poisoned run at DD, at the same point of the schedule, giving the matched clean loss

Rcleanmatched((1−ε)D)=R0(D)+Rctl,ε((1−ε)D)−Rctl,0(D).\mathcal R_{\rm clean}^{\rm matched}((1-\varepsilon)D)=\mathcal R_0(D)+\mathcal R_{{\rm ctl},\varepsilon}((1-\varepsilon)D)-\mathcal R_{{\rm ctl},0}(D).

The paper's observable, Eq. (3), splits into dilution and a poison-specific part:

Δobs=eRε−R0−1,Δdil=eRcleanmatched−R0−1,Δps=Δobs−Δdil.\Delta_{\rm obs}=e^{\mathcal R_\varepsilon-\mathcal R_0}-1,\qquad\Delta_{\rm dil}=e^{\mathcal R_{\rm clean}^{\rm matched}-\mathcal R_0}-1,\qquad\Delta_{\rm ps}=\Delta_{\rm obs}-\Delta_{\rm dil}.

Our new figures are for language switching data poisoning and it plots Rε/Rcleanmatched−1=ln⁡[1+Δps/(1+Δdil)]/Rcleanmatched≈Δps/Rcleanmatched\mathcal R_\varepsilon/\mathcal R_{\rm clean}^{\rm matched}-1=\ln[1+\Delta_{\rm ps}/(1+\Delta_{\rm dil})]/\mathcal R_{\rm clean}^{\rm matched}\approx\Delta_{\rm ps}/\mathcal R_{\rm clean}^{\rm matched} against D/DHD/D_{\rm H}:

  • At D=DHD=D_{\rm H}, Δps\Delta_{\rm ps} is 31–58% of Δobs\Delta_{\rm obs} for 490M–1287M, and 53–71% at 135M; the rest is Δdil\Delta_{\rm dil}.
  • From D/DH=0.4D/D_{\rm H}=0.4 to 1, Δps/Rcleanmatched\Delta_{\rm ps}/\mathcal R_{\rm clean}^{\rm matched} grows.

Hence we conclude data poisoning causes significant damage. We will add an Appendix to the paper with the new plots discussing this. We thank the reviewer for asking this question.

W1, Fig. 1 and Result 1. Fig. 1 is the main motivation for the theoretical analysis of our paper. As a theory first paper, we build towards explaining Fig. 1 in steps. We totally agree with the reviewer that Result 1 concerns the population optimum, where the first-order term vanishes, while Fig. 1 uses finite-time checkpoints, where a linear term is allowed. However, Result 1, 2 gives the main conceptual pathway that is used later in Sec. 5 to explain Fig 1 for a given model across different training stages (not across models). Sec. 4 is our main theoretical contribution which gives an exactly solvable model based on Results 1, 2. Sec that we validate numerically on its own terms. Sec. 5 (Bias-dominated crossover) then shows that the same mechanism is needed to explain how scaling exponent changes during training of large enough language models even when the discussion is around finite time checkpoint. At the same time Sec. 5 (Gap-dominated crossover) shows that a different mechanism explains the same for smaller language models. In a revised version of the paper, we will explain this flow of logic in the beginning of Sec. 3 according to reviewer's suggestion.

W2, Section 5. We agree with the reviewer that Sec. 5 is hard to follow and in a revised version we will organize it as assumptions, theoretical results, empirical interpretations emphasizing the measured scaling in Fig. 1 law is a crossover between a linear and a quadratic term, not a genuine fractional law:

δR(ε,T)=Rε−R0=f(T)ε+g(T)ε2,\delta\mathcal R(\varepsilon,T)=\mathcal R_\varepsilon-\mathcal R_0=f(T)\varepsilon+g(T)\varepsilon^2,

ff collecting the linear terms (checkpoint offset, train–eval gap, clean bias), gg the pure-poison term. A fixed-checkpoint crossover fraction is given by ε×(T)=f/g\varepsilon_\times(T)=f/g. Its local log-slope is (1+2u)/(1+u)(1+2u)/(1+u), with u=ε/ε×(T)u=\varepsilon/\varepsilon_\times(T). Result 21's exponent equals the fixed-checkpoint slope at exactly one ε∗=β−γ2γε×(T)\varepsilon^\ast=\frac{\beta-\gamma}{2\gamma}\varepsilon_\times(T). At this value, ∂TδR(ε∗,T)=0\partial_T\delta\mathcal R(\varepsilon^\ast,T)=0. This says:

  • Use the clean run and poisoned runs at several fractions ε1<⋯<εn\varepsilon_1<\dots<\varepsilon_n.
  • For each fraction, estimate Di=∂TδRiD_i=\partial_T\delta\mathcal R_i at TT.
  • Compute yi=Di/εiy_i=D_i/\varepsilon_i and fit yi=A+B εiy_i=A+B\,\varepsilon_i by least squares.
  • ε∗(T)=−A/B\varepsilon^\ast(T)=-A/B, the point where the fitted line crosses zero.

We will add this discussion in the Appendix of the revised version of the paper.

Please see new theory: reply to AiMd; independent validation of the clean-bias and poison-response laws: reply to SR9q. If you find our reply satisfactory, please raise your score. Thank you again.

Since the response period is ending soon, we kindly ask the reviewer if we've addressed their concerns and, if so, if the reviewer would consider updating their score.

Could the authors please highlight the added / revised content in the manuscript?

I would like to read the modified sections in-context with the rest of the paper.

I thank the authors for the rebuttal, which has partly addressed my concerns. In particular, I appreciate the notation table and the detailed explanations of Figure 1, Result 1, and the connections between sections. I would still like to read the revised manuscript to assess how the proposed changes fit within each section.

I also share the concern raised by another reviewer about generalization beyond the current experimental setting. Could the authors clarify what evidence supports extending the proposed scaling law to larger models and model families beyond the OLMo-style models studied here? I think that would help to distinguish what has been empirically validated from the theoretical expectation. For these reasons, I will maintain my score for now.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.