Synthetic Pre-pretraining Survives Scale, but Not as a Grammatical Prior

Abstract: Pre-pretraining (PPT) on synthetic non-natural language data improves token efficiency during LLM pre-training (PT). Prior work attributes this gain to a grammatical prior, i.e., a structural inductive bias learned during PPT that transfers to natural language grammar. However, PPT has only been tested on models of at most 1B parameters and PT budgets below 2B tokens on predominantly web text. It is unknown whether PPT is effective at larger scales and under more realistic PT data mixtures that combine diverse sources (e.g., code and math). We therefore present a comprehensive study on PPT spanning five PPT tasks, four PT data mixtures, four parameter scales (500M to 7B), and PT budgets of up to 100B tokens. Our results demonstrate that the downstream performance and token efficiency gains of PPT persist at scale, e.g., saving at least 21B PT tokens at the 3B scale. However, in contrast to prior work, we find no consistent evidence that these gains stem from a grammatical prior. Downstream performance does not consistently align with grammatical acceptability across model sizes. Instead, we find that downstream gains arise from PPT tasks that improve long-range retrieval. Finally, PPT performance gains are robust to how PT data mixtures are composed and diminish only when web text is absent. Overall, PPT is a low-cost addition to PT, and future PPT task design should target long-range retrieval rather than natural language grammar.

• The paper investigates whether synthetic pre-pretraining (PPT) enhances the performance of a neural network language model when the model's scale and data budget are increased to those characteristic of state-of-the-art pretraining, uncovering that downstream gains do not reliably stem from a grammatical prior but rather from improved retrieval of earlier sequence positions.
• PPT persists in improving the performance of language models across a range of sizes (500M to 7B parameters) and pretraining (PT) budgets (21B to 100B tokens), emphasizing the practical benefits of PPT over large scales and diverse PT mixtures.
• Even with bulky amounts of code or mathematics in PT, the practical benefits of PPT remain outstanding. Conversely, language modeling performance significantly drops when web text content is eliminated from PT, emphasizing web text as a key composite for effective PPT.

Research question and contribution

The paper investigates whether synthetic pre-pretraining (PPT) remains useful under model and data scales that approximate contemporary language-model pretraining. PPT inserts a short optimization phase on synthetic sequences before ordinary pretraining (PT), with the resulting parameters used to initialize PT. Earlier work reported improved token efficiency and attributed the effect to a transferable grammatical prior: a structural bias learned from formal languages such as $k$-Shuffle Dyck and subsequently applied to natural-language syntax. This study challenges that interpretation while preserving the practical conclusion that PPT can improve downstream performance.

The experimental scope is substantially broader than prior PPT evaluations. The authors vary model size from 500M to 7B parameters, PT budgets from 21B to 100B tokens, four PT mixtures, and five PPT tasks. The evaluation combines ten downstream benchmarks with BLiMP grammatical acceptability and verbatim long-range retrieval. The central claims are:

• PPT gains persist across model scaling and extended PT.
• The gains do not consistently track grammatical acceptability.
• Effective PPT tasks share a requirement to retrieve a particular earlier position in a sequence.
• Increasing code and mathematics content does not eliminate PPT benefits, whereas removing web text largely does.

The study therefore separates the empirical fact of PPT transfer from the proposed mechanism of grammatical transfer (2609.39827).

Experimental design

All models use the SmolLM3 architecture and tokenizer, with parameter counts of 500M, 1B, 3B, and 7B. The principal PPT phase consists of 500 optimization steps, corresponding to approximately 1.05B synthetic tokens. The subsequent PT phase uses a 4,096-token context, a global batch of 512 sequences, approximately 2.1M tokens per step, and a large-batch AdamW configuration. The standard PT budget is approximately 21B tokens; extended experiments use 100B tokens at 3B and 75.5B tokens at 7B.

The main synthetic task is $k$-Shuffle Dyck, a context-sensitive bracket language in which bracket dependencies can span intervening symbols. The task ablation includes MP-Struct Core, which makes matching relations explicit through structural markers; neural cellular automata (NCA), which encode repeated state transitions; and Set, which requires identifying whether a token has appeared previously but does not require locating a particular occurrence. A fifth condition, Control, applies the same 500-step warm-up to held-out natural-language PT data, thereby testing whether gains result merely from additional optimization.

The PT mixtures differ in domain composition. C4 contains only web text. SmolLM3 includes filtered web text, code, and mathematics. OLMo3 is more STEM-weighted and includes OCR-derived academic PDFs, code, and mathematics. Marin is predominantly web-based with smaller code and mathematics components. This design permits separate assessment of parameter scale, training duration, corpus composition, and synthetic-task structure.

Results are reported as paired differences between PPT and PT-Only over checkpoints from 5K to 10K PT steps. A gain is called stable when its mean exceeds its checkpoint standard deviation. This averaging procedure is important because BLiMP margins are highly checkpoint-sensitive: in the authors’ audit, half of the 12 BLiMP comparisons reverse sign when a single 10K checkpoint is replaced with the multi-checkpoint mean, whereas no downstream-average comparison reverses sign.

Downstream gains survive scale

The principal empirical result is that $k$-Shuffle Dyck improves general downstream performance across a broad range of model sizes and PT mixtures. Among the 12 combinations of three model scales and four PT mixtures, nine improve by at least 0.6 percentage points, with a mean gain of 1.6 points among those cases. The mean gains by scale are 0.8 points at 500M, 1.4 at 1B, and 1.3 at 3B. Thus, the benefit does not disappear when model capacity increases through 3B parameters.

The gains are distributed across task categories rather than concentrated in a single benchmark family. Reading comprehension improves in 11 of 12 scale–mixture configurations, science question answering in 10, commonsense reasoning in 9, and language modeling in 10. Language modeling shows the largest category-level improvements, with mean gains of 1.2, 1.8, and 1.7 points at 500M, 1B, and 3B, respectively. ReCoRD, HellaSwag, and LAMBADA are particularly consistent beneficiaries; each requires information from an earlier passage rather than only the immediately preceding sentence.

The Control condition remains close to PT-Only, with mean differences of $-0.2$, $+0.4$, and $+0.2$ points at 500M, 1B, and 3B. In contrast, $k$-Shuffle Dyck exceeds Control in 10 of 12 configurations, with margins of approximately 1.0, 1.0, and 1.1 points across the three scales. This comparison supports the conclusion that the effect is attributable to the synthetic training distribution rather than simply to 500 additional optimization steps.

Figure 1: Downstream score changes from PPT relative to PT-Only across scale and PT-mixture configurations; whiskers represent one standard deviation across checkpoints.

The extended-budget experiments directly address whether PPT is only a transient initialization advantage. At 3B, downstream performance remains higher for PPT at every evaluated budget from 21B to 100B tokens on both C4 and Marin. On Marin, the PPT model reaches an average downstream score of 62.3 after 63B tokens, while PT-Only reaches 62.0 only after 84B tokens. This corresponds to a saving of at least 21B PT tokens. At 100B tokens, PPT retains gains of 1.6 points on C4 and 1.2 points on Marin relative to PT-Only.

The effect is smaller but still positive at 7B. Across Marin budgets up to 75.5B tokens, PPT improves the downstream average at every evaluated point, with a mean gain of 0.6 points. The reduction relative to 3B should not be interpreted solely as a capacity effect, because the 7B experiments receive approximately 11 tokens per parameter compared with approximately 33 at 3B. Nevertheless, PPT remains beneficial after substantial optimization at a scale beyond previous studies.

The grammatical-prior account is not supported

The paper’s strongest conceptual claim is also its principal departure from prior work: the downstream benefits of PPT do not consistently arise through improved grammatical acceptability. If PPT creates a transferable grammatical prior, formal-language PPT should improve BLiMP, particularly its morphology and syntax subsets, which measure agreement, constituency, binding, and related dependencies.

The observed results do not meet this prediction. Although $k$-Shuffle Dyck improves overall BLiMP accuracy in 9 of 12 scale–mixture pairs, only one improvement is stable under the paper’s mean-versus-standard-deviation criterion. By contrast, 10 of 12 downstream-average improvements are stable. At 500M, the BLiMP effect is already mixed: OLMo3 and Marin improve, while SmolLM3 and C4 decline. The C4 degradation is especially consequential because earlier work reported a grammatical benefit in a comparable web-text setting. Increasing model size does not resolve the inconsistency; on Marin, the BLiMP delta changes from $+0.3$ at 500M to $+1.6$ at 1B and then to $k$0 at 3B.

The subgroup results further weaken the grammatical interpretation. Morphology and syntax—the categories most directly related to the hypothesized structural transfer—show neither consistent nor stable improvements. In the appendix analysis, $k$1-Shuffle Dyck changes in these categories remain at or below 1.0 point in 19 of 24 subgroup comparisons and are unstable in 20. Variation instead concentrates in semantics, where checkpoint standard deviations are much larger. This pattern is not what would be expected if formal bracket structure were transferring specifically to natural-language grammar.

Figure 2: BLiMP score changes from PPT relative to PT-Only; the inconsistent and checkpoint-sensitive margins contrast with the more reliable downstream gains.

The result is not that PPT never improves BLiMP. Rather, BLiMP improvements are too inconsistent, too small relative to checkpoint variation, and too weakly aligned with downstream gains to function as the principal explanatory variable. This distinction matters: the paper rejects a specific mechanistic interpretation, not the existence of structural effects in general.

Long-range retrieval as the common capability

Verbatim retrieval provides the contrasting result. The task measures mean NLL when the model must reproduce a noun list encountered earlier in a long context. $k$2-Shuffle Dyck lowers NLL in all 12 scale–mixture pairs, with stable reductions in 7. The magnitude is modest but substantially more regular than the BLiMP pattern. Control improves retrieval in 8 of 12 cases, but only 3 of these reductions are stable, indicating that the synthetic sequences provide a more specific benefit than generic warm-up optimization.

The authors relate this result to the computational structure of the synthetic tasks. In $k$3-Shuffle Dyck, predicting a closing bracket requires matching it to a relevant opening bracket across an arbitrary number of intervening symbols. This is a retrieval operation over an abstract vocabulary. The same interpretation explains why downstream gains are largest on LAMBADA, ReCoRD, and HellaSwag: these tasks often require recovering information from a preceding passage.

The claim is therefore not that PPT teaches a grammar-independent retrieval algorithm in a fully established causal sense. Rather, the convergent benchmark pattern indicates that the most reliable behavioral change induced by PPT is improved access to information at earlier sequence positions. The distinction between BLiMP and verbatim retrieval is central: BLiMP usually determines the answer from evidence within a short sentence, whereas verbatim retrieval requires locating and reproducing a prior span.

PT-mixture composition and web-text dependence

The data-mixture experiments test whether code and mathematics already provide the structural signals that PPT supplies. The results contradict the hypothesis that increasing these domains makes PPT redundant. On Marin at 3B, raising the mathematics share from 1.3% to 17.0%—a 13-fold increase—produces an average PPT gain of 2.0 points, essentially unchanged from the 1.9-point gain on the original mixture. Category-level gains also remain similar.

The analysis instead identifies the presence of web text as the relevant condition. DCLM alone preserves 1.5 of the original 1.9 downstream points, and FineWeb-Edu alone preserves 1.0 point. Removing web text entirely, leaving a mixture of approximately 82% source code and 18% mathematics, reduces the gain to 0.2 points. Language modeling follows the same pattern: PPT gains range from 1.5 to 3.1 points when web text is present but fall to 0.6 without it.

This result should be interpreted cautiously. The study establishes sensitivity to a web-text-free ablation, not a general causal law about web data. The source-code and mathematics-only condition is unlike the practical mixtures otherwise evaluated, and the authors do not identify which properties of web text—lexical diversity, discourse structure, document organization, or distributional mismatch—mediate the interaction. Still, the result shows that the relevant redundancy question is not simply whether PT contains structured domains.

The OLMo3 anomaly remains unresolved. OLMo3 is the only mixture that fails to benefit at more than one scale, despite having a web share close to the successful 17%-mathematics Marin variant. The paper explicitly leaves a direct OLMo3 composition ablation open, so the evidence does not support attributing its weak gains to either mathematics proportion or web proportion alone.

Which synthetic tasks transfer?

The task ablation demonstrates that formal grammar is neither necessary nor sufficient for effective PPT. At 3B, MP-Struct Core yields a 0.9-point average downstream gain, while the non-formal NCA task yields 1.2 points, comparable to the 1.3-point gain from $k$4-Shuffle Dyck. Each improves three of the four PT mixtures. NCA is particularly informative because its dependencies arise from repeated automaton states rather than nested linguistic constituents.

Set produces the opposite result: a 7.2-point average downstream degradation and no improvement on any PT mixture. The task requires determining whether a token has appeared previously, but the answer can be computed by maintaining a record of seen values; it does not require retrieving a specific earlier position. Set also worsens verbatim retrieval, with a mean NLL increase of 0.587.

The contrast between NCA and Set supports a more precise task-design criterion. Effective synthetic tasks should require retrieval of a determinate earlier location or state, not merely aggregate membership tracking. This explains why MP-Struct Core remains effective despite its formal-language status and why NCA transfers despite lacking a natural-language grammar. It also qualifies the paper’s headline distinction: the relevant alternative to a grammatical prior is not arbitrary sequence structure, but structure that operationalizes positional retrieval.

Limitations and open questions

The principal limitation is statistical coverage. Most configurations use a single random seed, with three seeds evaluated only for 3B Marin. In that setting, downstream gains are consistent across seeds, varying from 1.7 to 1.9 points, but BLiMP changes from $k$5 to $k$6, and verbatim retrieval improves in only two of three seeds. The broad downstream pattern is therefore better supported than the finer-grained mechanistic claims.

The study also uses one architecture, one tokenizer, a fixed 4,096-token context, and a single 500-step PPT budget. Consequently, it does not establish whether the retrieval interpretation transfers across architectures, tokenization schemes, context lengths, PPT durations, or optimizer schedules. The 7B analysis is restricted to Marin and $k$7-Shuffle Dyck, preventing a full task-by-mixture-by-scale comparison at that size.

The causal status of “long-range retrieval” remains partly inferential. The alignment among verbatim retrieval, downstream tasks requiring passage-level information, and effective PPT-task structure is substantial, but the paper does not manipulate retrieval distance, ambiguity, positional distribution, or memory demands independently. Such experiments would be needed to distinguish positional retrieval from related mechanisms such as improved attention allocation, reduced optimization difficulty, or altered representation geometry.

Finally, the weak OLMo3 result is unexplained, and the web-text ablation does not isolate the responsible corpus property. The paper leaves open whether the decisive factor is web-text diversity, document-level organization, lexical statistics, natural-language prevalence, or interaction with the OLMo3 mixture’s OCR and STEM components.

Conclusion

The paper establishes that synthetic PPT is not merely a short-horizon initialization artifact. Across 500M–7B models and PT budgets reaching 100B tokens, it yields persistent downstream improvements, including a measured saving of at least 21B PT tokens at 3B. However, the evidence does not support treating those gains as the transfer of a grammatical prior. BLiMP improvements are inconsistent and checkpoint-sensitive, while verbatim retrieval improvements are substantially more reliable.

The most coherent account supported by the experiments is that effective PPT tasks train models to retrieve specific earlier positions or states in long sequences. This property appears in formal and non-formal tasks, survives diverse PT mixtures, and is not made redundant by substantial code or mathematics content. The paper consequently reframes PPT task design around retrieval demands rather than grammatical resemblance, while leaving open the precise computational mechanism and the corpus properties responsible for the observed dependence on web text (2609.39827).