Language Model Pretraining

Overview

Language model pretraining builds general-purpose models by learning from large collections of text, commonly through next-token prediction or masked-token prediction. Researchers study how data curation, filtering, mixing, repetition, and multilingual or domain-specific corpora affect what models learn, alongside changes to tokenization, architecture, training objectives, and optimizers. Work also aims to make training cheaper through methods such as sparse attention, low-precision computation, parameter sharing, and more efficient data selection. A central tension is that lower training loss or greater data volume does not always bring better downstream abilities: data quality, diversity, exposure, and evaluation choices all matter, and their effects can vary with model scale and task.

Latest papers

480 papers52 papers in the last four weeks, down 27% on the four weeks before. 0.4% of all new papers.

Oct 14, 2026

Language Exposure Matters: Distribution and Order in Multilingual BabyLM Pre-training

We investigate two aspects of multilingual pre-training with limited data, namely how language exposure is distributed across English, Dutch, and Chinese, and how that exposure is ordered during training. In particular, we compare balanced multilingual exposure to language-dominant exposure, where one language is exposed with more training data than the others. We also compare mixed training, where multilingual training data is used throughout training, with curriculum training, where monolingual exposure of the dominant language is followed by balanced multilingual exposure. Our results show that multilingual BabyLM performance varies with both the amount and order of language exposure. Mixed training outperforms curriculum learning across all 15 matched configurations, indicating that simultaneous multilingual exposure is more effective than staged exposure in our setting. Several language-dominant mixed training configurations achieve higher scores than the balanced configuration, suggesting that a balanced distribution is not necessarily the highest-performing configuration in our setting. Interestingly, the Dutch-primary configurations achieve the best overall performance.
Oct 14, 2026

AFCS: Atomic-Fact-Conditioned Synthesis for Factual Knowledge Pretraining

Synthetic data is increasingly used in language-model pretraining. Existing knowledge-corpus synthesis pipelines, including WRAP and Nemotron-CC, usually take a source passage or document segment as input and generate rewritten or QA-style samples from that text. This can improve surface form and prompt diversity, but it does not explicitly require fine-grained source facts -- especially secondary entities, list/table entries, and precise temporal attributes -- to appear in the synthesized corpus. We propose AFCS (Atomic-Fact-Conditioned Synthesis), a pipeline that first extracts source-supported atomic facts and then uses each retained fact to drive generation. AFCS uses structure-aware extraction over the document tree, source-grounded verification, local context expansion, and multi-format generation. In continued-pretraining experiments with OLMo-3-7B and a 32B base on the same FineWiki source corpus, AFCS outperforms WRAP and Nemotron-CC on SimpleQA, SimpleQA Verified, PopQA, and HotpotQA. Granularity ablations show that fact-level synthesis improves over chunk-level, paragraph-level, and entity-level synthesis; and a 100B-token mid-training integration improves factual benchmarks and raises the 12-task general benchmark average by +1.09.
Oct 14, 2026

CAST: History-Adaptive Spectral Clipping for LLM Pretraining

Soft spectral clipping (SSC) reduces gradient singular values, but its effect depends on the gradient scale and clipping threshold. We introduce CAST (Clipping with Adaptive Scales and Thresholds), a history-adaptive extension of SSC. AdamW's previous second moment sets row and column scales. An exponential moving average of the clipped output's estimated spectral norm sets each matrix's threshold. CAST restores the original coordinates before the AdamW update. We train a mixture-of-experts (MoE) model with 19.08B total parameters for 83.89B tokens. CAST achieves lower training and mean validation losses than nine gradient-processing baselines, with the fewest relative gradient spikes. It reaches the best baseline's final smoothed training loss in 7.2% fewer steps.
Oct 14, 2026

PA1-OPD: One-Step Prefix-Augmented On-Policy Distillation for Language Model Pretraining

Knowledge distillation (KD) improves language-model pretraining by matching teacher distributions at corpus prefixes, but provides no direct supervision at contexts induced by student predictions. We introduce PA1-OPD, which augments corpus-path distillation with teacher feedback after one student-sampled transition per selected prefix. This expands the supervised states while retaining next-token prediction and corpus-path KD. Data-anchored roots support parallel sampling, and differentiable tree attention shares prefix computation without changing inference. In two continued-pretraining settings, full coverage improves held-out cross-entropy over KD across C4, PG-19, and Pile at every evaluated checkpoint. In Stage2, selecting the 10% of roots with highest student surprisal adds 48.85% to KD's Macro cross-entropy gain over next-token training. Gains remain when branch supervision excludes successor states already on the corpus path, supporting the value of feedback beyond observed contexts. Additional depth or breadth improves aggregate results at unchanged per-term supervision weight, with benefits varying by corpus. These findings support local student-induced supervision as a complement to pretraining distillation.
Oct 13, 2026

Qwenfant: Train Big Character-level BabyLMs on Concise Parallel Text

This work concerns our submission for the Multilingual track of the 2026 BabyLM Challenge. We improve on our baseline using data that has been summarized and translated by an LLM, showing that higher-quality data can come from synthetic sources. We further improve with 2 simple changes: using a larger model, and using a character-level vocabulary. Both of these changes require LoRA-style fine-tuning for realizing their full potential. We dub our final Qwen3-based model Qwenfant, which, at the time of submission, ranks 1st in the Multilingual track.
Oct 13, 2026

Copy-Paste Language Models

Transformers are intuitively considered to be efficient at copying from their inputs. We actually show that the ability of LLMs to copy in context can be improved through explicit modeling. We propose Copy-Paste Language Models (CPLMs), which add a lightweight Copy-Paste head to the conventional language modeling (LM) head. CPLMs improve training efficiency, downstream performance, and offer new operating points in long-context and retrieval setups. We validate our approach by pretraining models from 135M to 1.7B parameters, showing improvements across all scales and backbones with negligible training and inference overheads. We explain the success of CPLMs by showing that they partially mitigate the LM head gradient bottleneck through an alternative gradient path.
Oct 13, 2026

Does Pretraining Teach Chemical Language Models Chemistry?

Chemical language models (CLMs) are pretrained on large corpora of molecules with the aim of learning representations that transfer to diverse downstream tasks. Here, we systematically study the effects of pretraining using a series of such CLMs across 64 molecular property prediction datasets. We find that the downstream gains of pretraining in CLMs do not scale and performance saturates below gradient-boosting baselines. Across the regimes evaluated, only supervised pretraining on traditional physico-chemical descriptors consistently improves downstream performance for the CLMs tested whereas some forms of pretraining result in models worse than a randomly initialized encoder. While models can resolve molecular structure well, they do so no better than a randomly initialized model, suggesting that, at best, CLMs learn descriptor-like properties in addition to low-level token statistics during pretraining, but surprisingly little chemistry beyond what classical molecular features already capture.
Oct 13, 2026

The Geometry of Language Model Logits: Generic Structure and Learned Departures

Autoregressive language models produce, at every prediction step, a vocabulary-sized logit vector whose structure determines the conditional next-token distribution. We empirically identify that the ordered vocabulary-head logits follow a remarkably consistent log-linear relation z(r)≈μ−αlog⁡r.z_{(r)} \approx μ- α\log r. across prediction contexts, architectures, and scales, emerging progressively during pretraining. To our knowledge, this is the first systematic analysis of the logit geometry underlying the distributional assumptions used in the prior decoding work. While controlled order-statistic baselines spanning bounded, light-tailed, and heavier-tailed distributions can largely explain the observed log-linear geometry, they do not explain the systematic shift in the slope αα: trained language models exhibit α≈1.4α\approx 1.4--1.71.7, sitting well above the initialization baseline (α≈0.2α\approx 0.2) . Moreover, αα correlates with predictive performance, suggesting that it captures learned structure beyond generic extreme-value effects.
Oct 13, 2026

Agaram 2.5B: The First Tamil-First Language Model Trained from Scratch with Tool Calling. A Playbook for Under-Represented Languages

We present Agaram 2.5B, a Tamil-first language model trained from scratch that converses in Tamil script, romanized Tanglish and English, and calls tools. Our first model, Agaram 1.5B, guided the new base, tokenizer and fine-tuning recipe: we pretrained a new 2.53B-parameter base on 50 billion tokens, 58% of them Tamil, with a tokenizer that needs 1.55 tokens per Tamil word where Qwen2.5's needs 9.44. A four-day public preview of our first 2.5B chat model showed where it fell short with real users, and we rebuilt supervised fine-tuning around what we saw, on 222,662 records. The released model scores above Sarvam-1 in all six translation directions we measured (FLORES chrF 56.9 into Tamil and 55.4 into English), chooses the expected tool on all 25 prompts of a tool set that was used during checkpoint selection and whose schemas appear in training, with exact arguments on 16, and was preferred over the preview model by an LLM judge on 100 sealed human-written prompts (net preference +30 points, 95% interval +14.5 to +45.0). On FLORES it also scores above Google's Gemma 4 E4B-it and our fine-tuned variant of it in both directions, with 2.53B parameters against E4B's 4.5B effective parameters. Factual knowledge and reliable refusal of harmful requests remain open problems, which we set out as future work. The base and chat weights and GGUF files are released under Apache-2.0, and this report gives a full account of the data, training and evaluation as a playbook for building language models for under-represented languages.
Oct 13, 2026

Mixed Precision Scaling Laws for Quantized Pretraining

Since scaling laws were introduced for modern LLMs (Kaplan et al., 2020), several works have extended them to account for quantization, tracking the emerging low-precision data types supported on AI accelerators. We develop a more accurate scaling law for quantized pretraining with mixed precision across the forward and backward passes, grounded in optimization theory and validated empirically. Our formulation separates the effects of quantization on model capacity and on optimization efficiency, and explicitly models the persistent optimization error from forward and backward quantization identified through convergence analysis. To select the regimes for fitting and evaluating our law, we first decompose existing 4-bit training recipes into their forward and backward components, isolating those that most affect loss. Our Qwen3 experiments under selected 4-bit regimes, together with published Llama-style experiments, show that our law predicts loss with up to 11×11\times lower MSE than existing laws as model size and token count scale. Finally, we extend our law to layer-wise mixed precision and identify the optimal quantized-pretraining recipe, i.e., model size, training tokens, and per-layer forward/backward precision, under a hardware-aware compute budget. Unlike prior precision scaling laws, our law predicts that the optimal precision increases with the compute budget.
Oct 13, 2026

Revisiting the 16x Pretraining Compute Gap between Diffusion and Autoregressive Language Models

Diffusion language models (DLMs) have attracted wide attention for their fast, parallel decoding, but they are believed to be far less efficient to train than autoregressive language models (ARMs): a widely cited study reports that a DLM needs roughly 16x the training compute of an ARM to match its likelihood, a figure that has made pretraining DLMs from scratch look prohibitive at scale. We find that the methodology behind this figure has fallen out of step with current DLM practice: it scores the model with its evidence lower bound (ELBO) and trains a full-sequence, decoder-only model. We therefore revisit the scaling laws of DLMs, adopting the exact likelihood under the model's own decoding order, block diffusion training, an encoder-decoder architecture, and a sweep over the block sizes used in practice. Under this new protocol, we study across ten model variants, about 1,700 models in total, and find that pretraining DLMs is far cheaper than previously thought. Concretely, scoring with the exact likelihood reduces the compute multiple to 12.8x; block diffusion training and an encoder-decoder architecture further reduce it to 7.8x at block size 32, and smaller block sizes narrow it to 4.6x at block size 8 and 3.0x at block size 4. We complement these scaling analyses with an empirical rule for the optimal decoder depth of encoder-decoder block diffusion, a held-out compute budget beyond the fitting range at which the gap between variants persists, and downstream benchmark results where likelihood improvements carry over to accuracy. Overall, the 16x figure substantially overstates the pretraining cost gap between DLMs and ARMs, and pretraining DLMs from scratch is far more viable than previously assumed.
Oct 12, 2026

Distilling Text LMs into Speech-to-Speech LMs by Matching Internal Text Predictions

Text language models (LMs) acquire broad linguistic and factual capabilities through large-scale pretraining, while comparable capabilities remain difficult to learn from speech alone. Recent work has shown that knowledge distillation from pretrained text LMs offers a promising way to transfer their linguistic and factual capabilities to speech models. However, existing distillation approaches have largely focused on speech-input, text-output models. This idea has yet to be explored for {\it speech-to-speech} models, which operate on both speech input and speech output. In this work we introduce a speech-to-speech pretraining method that creates an internal text interface for knowledge transfer while preserving the speech-only input and output. The model first consumes speech context without its transcription, switches to text prediction, and then uses the predicted text to lead future speech generation word by word. It also provides an interface for knowledge distillation: the internal text predictions can be directly distilled from the base text LM. To the best of our knowledge, this is the first application of text LM distillation to a speech-to-speech LM. Under matched training conditions, the proposed sequence construction substantially improves the generative perplexity of spoken continuations over speech-only and interleaved speech-text models, approaching the performance of a text-leading inner-monologue model with external ASR prompt transcription. More importantly, distillation improves factual performance under both text and speech evaluation, indicating better retention of knowledge from the base text language model.
Oct 12, 2026

Monarch-Muon: Structure-Aware Muon for Fast and Memory-Efficient Pretraining

Training large language models is constrained by both wall-clock time and memory footprint, and recent matrix-aware optimizers sharpen this tradeoff: Muon can improve optimization but its dense polar update remains expensive. We propose Monarch-Muon, a structured matrix-aware optimizer for weights parameterized by Monarch factors. The resulting structure reduces weights, gradients, and optimizer state, and exposes small independent update subproblems in place of a single dense orthogonalization; we formalize this through a blockwise trust-region interpretation. In quality runs, Monarch-Muon improves over Monarch-AdamW at fixed structure: 3.06 vs. 3.09 in the dense 257M model and 2.68 vs. 2.77 in the 1B MoE model; the two-block MoE variant also matches dense AdamW while using 15% fewer active and 25.4% fewer total parameters. In 6.89B-parameter systems measurements, Monarch-Muon reduces peak memory from 68.9 to 52.9 GB, Muon optimizer state from 14.60 to 6.77 GB, and the Muon optimizer step from 1288 to 304 ms.
Oct 11, 2026

Latent Taxonomies: How Property Induction Develops in Language Model Pre-training

When told that robins have sesamoid bones, people are far more prone to extend the property to sparrows than to whales. This behavior, called category-based property induction, is well documented in humans, but whether and how it develops in language models trained only to predict the next token is unclear. To shed light on this, we construct a hierarchical taxonomy of synthetic entities grouped by the properties they share. We then inject facts from this latent taxonomy into the continued pre-training of Pythia and OLMo 2 models from 70M to 7B parameters, and track the development of inductive behaviors. We find that taxonomically aligned induction emerges from next-token prediction alone: a property learned for one entity transfers more strongly within its taxonomic category than across categories. Internally, entities that share a deeper category become more similar in activation space, and clustering recovers the full taxonomic tree from entity representations. Contrary to the usual expectation that scale helps, we find that induction weakens with model size when the data budget is fixed, suggesting that larger models memorize individual facts rather than compress them into shared structure. Induction also strengthens with more pre-training before the taxonomy injection, transfers to novel in-context properties, and can be overridden by a latent taxonomy implied by the prompt. Together, these results show that a model's inductive generalization reflects the latent organization of its training data, with implications for data curation and for predicting how far a learned fact will spread.
Oct 11, 2026

In-Place Indic Tokenizer Expansion and Continual Pretraining of Nemotron-3-Nano-30B-A3B

We adapted Nemotron-3-Nano-30B-A3B for eight Indian languages without changing its backbone architecture. We expanded the vocabulary in place from 131,072 to 163,072 entries using continued byte-pair encoding and AdaptMix rebalancing, reducing tokens per word by 38.6%. We initialized new input and output rows by averaging their original constituent rows, then trained them on 90% Indic and 10% English text before updating the backbone. Continual pretraining then progressed through broad training, high-quality annealing, and long-context training over approximately 1.27 trillion tokens. The annealed checkpoint gained 12.9 points on the Indic MILU benchmark but lost 3.4 on English MMLU. We also report practical negative results. Direct token addition underperformed continued byte-pair encoding. AdamW weight decay changed rows intended to stay fixed. Perplexity on an Indic-heavy holdout did not select the learning rate that best preserved English performance.
Oct 11, 2026

A Sovereign, Indic-First, Reasoning Model

Gnani-Evon-v3.3-30B-A3B is a long-context hybrid Mamba-Transformer mixture-of-experts model for reasoning, instruction following, and tool use in English and eight Indian languages. Starting from NVIDIA's Nemotron-3-Nano-30B-A3B checkpoint, it retains the 52-layer backbone and changes the vocabulary, training data, and post-training objectives. An in-place expansion from 131,072 to 163,072 tokens reduces mean Indian-language fertility by 38.6%. The new embedding rows are initialized from their upstream decompositions and warmed before the backbone is updated. Continual pretraining then adds broad English-Indic data, high-quality annealing, and direct training at a 131,072-token sequence length. The central result appears after post-training. Reinforcement learning across four verifiable task environments improves mathematics, graduate-level science, and instruction following, but degrades Indic extractive question answering because the rewards do not require an answer in the requested language. The share of Indic extraction items answered in English rises from 8.7% to 91.7%. Accuracy conditional on answering in the target script roughly doubles, which is consistent with preserved reading ability but is measured on a self-selected subset. Two unreleased studies test repairs. A scoped Indic reward restores the language convention. A 26B-A4B generative judge outperforms a 235B-parameter model on RM-Bench, and distillation restores 81% of the capabilities displaced by optimizing against it. Relative to the architecture-identical starting checkpoint, the released model gains 5.2 points on an Indic aggregate while giving up 1.7 points of English retention. The unreleased Indic-aware branch trades further, gaining 14.6 Indic points at a cost of 5.4 in English. A planned FP8 release reduces the artifact to 32GB with near-zero mean change across 32 paired benchmarks.
Oct 10, 2026

UltraBERT: Beyond the Limit of BERT Pre-Training

Recent bidirectional encoders have been modernized with innovations from causal large language models (LLMs). However, the pretraining recipe itself remains inherited from small-scale conventions, leaving encoders underoptimized, with no single model performing competitively across both classification and retrieval tasks. In this work, we investigate how far encoder performance can be pushed by revisiting core design choices in training, revealing three key findings: (1) Encoders saturate on downstream tasks beyond a few trillion tokens, indicating that additional compute past this point does not help. (2) Across model sizes, lower masking ratios benefit classification while higher ones benefit retrieval, contradicting prior work suggesting larger ratios for larger models. (3) Distillation with pruning improves classification at low masking ratios but degrades retrieval at high ones. Guided by these findings, we present UltraBERT, a single encoder that assigns separate LM heads to low and high masking ratios within a batch and restricts the distillation objective to the low ratio. To our knowledge, UltraBERT is the first encoder to achieve state of the art on both classification and retrieval tasks, reaching 91.0 on GLUE and 47.7 on BEIR. Together, these results show how far encoder performance can be pushed and provide a principled basis for compute allocation, masking ratio design, and knowledge transfer in bidirectional models.
Oct 9, 2026

Enhancing Legal Reasoning in Pre-trained Language Models via Knowledge Graph-Guided Multi-Task Pre-training

Legal language understanding requires reasoning over semi-structured regulatory documents, yet standard pre-training objectives do not explicitly target this structural capability. Because current legal language models rely solely on masked language modeling, they treat statutory text as flat token sequences and ignore the underlying hierarchical constraints. To bridge this gap, we present a multi-task pre-training approach for German traffic law (StVO). Our approach combines three complementary objectives: Next Sentence Prediction (NSP), Knowledge Graph (KG) Guided Clustering (CLU), and Knowledge Graph (KG) Aware Paragraph Classification (CLS). This approach explicitly encodes structured legal knowledge directly into the model weights. This encoding enables zero-shot legal reasoning at inference time and removes the need for discrete graph traversal. We evaluate our approach across nine Transformer-based encoders and 324 experimental runs. On synthetic StVO question answering, our models demonstrate consistent relative improvements between 10.3% and 155.0% over baselines. Furthermore, relative to the established LexGLUE baselines (Chalkidis et al., 2022), our models improve over prior results on four of seven legal benchmarks. It also achieves consistent zero-shot transfer gains between 3.4% and 15.8% on Austrian driving regulations. These experimental results indicate that our training approach generalizes successfully beyond the initial training legal domain.
Oct 9, 2026

Simulating Training Spikes of Deep Neural Networks by Layer-wise Random Feature Models

Intermittent ``spikes'' in loss, during the pre-training of Large Language Models (LLMs), remains a mysterious phenomenon both theoretically and empirically. In this paper, we show that simple quadratic models with heavy-tailed, power-law distributed inputs can replicate these spikes. We introduce a theoretical framework and simulation algorithm to model per-layer training dynamics of Deep Neural Networks (DNNs), using such quadratic models. In a case study on an Adam hyperparameter configuration (where β2β_2 is set close to 11) known to induce spikes in Transformer LMs, our simulation successfully predicts bumps indicating spiky behavior from a given checkpoint. Furthermore, we derive an appropriate β2β_2 schedule from observed metrics, which empirically improves performance.
Oct 9, 2026

Host-Aware Web Crawling for LLM Pretraining Data

Web crawling is a major source of large language model (LLM) pretraining data, yet data-selection pipelines discard a large share of crawled pages. We investigate how a crawler should allocate its budget across and within hosts to collect more high-quality pretraining data. Combining document-level quality classifier scores, web-graph centrality metrics, and Common Crawl's per-host crawl coverage, we find that roughly half of the variation in classifier scores occurs between hosts. In the crawls we analyze, a few pages from a host also provide a useful estimate of the quality of its remaining pages. Common Crawl uses harmonic centrality, a graph centrality metric, to guide its crawl allocation. In our simulations, using centrality to choose which hosts to crawl yields limited gains over random ordering and performs worse for some classifiers. For deciding how deeply to crawl each host, sampling a few pages and then fetching more from hosts with higher-scoring samples collects several times more high-quality pages per fetch than uniform depth or centrality-based allocation. Finally, Common Crawl's observed per-host coverage has little correlation with the evaluated quality signals. We estimate that tens of millions of high-quality pages remain among known but uncrawled URLs in each crawl examined, with most concentrated on a small share of hosts.
Oct 8, 2026

Adapting English Quality Classifiers for Multilingual LLM Pretraining Data Selection

Recent advances in large language model (LLM) pretraining highlight the role of high-quality training data in improving performance. While model-based filtering has proven effective in selecting high-quality subsets from web-scale corpora, especially for high-resource languages, low-resource languages face challenges due to limited availability of annotated data. This work explores extending quality filtering to over 100 languages by proposing a multilingual adaptation approach that converts an existing English quality classifier into a multilingual variant. Our approach proposes training a small multi-layer perceptron on top of Transformer encoder-only model embeddings, using multilingual text as input and scores obtained from English classifiers applied to machine-translated text as labels. Our 1B, 3B and 8B scale experiments show that our approach maintains the downstream LLM benchmark performance of existing multilingual model-based filtering baselines, without harming regional and cultural knowledge benchmarks. To further evaluate cross-lingual generalization, we compare classifier scores of high-quality synthetic data and web samples, and the correlation of classifier scores with LLM-based ones, revealing that the classifier can learn the scoring criteria of its original English variant, even for languages not included in its training data.
Oct 7, 2026

OpenEBFM: Energy-based Foundation Models Are Scalable Architecture

Energy-based models have shown promising properties for domain-specific tasks but remain challenging in modeling general-purpose scenarios. To bridge this gap, we design energy-steered models (ESMs) for foundational language modeling. ESMs integrate energy steering and token generation within a single architecture to leverage the energy-guided search to refine token generation. Specifically, ESMs first optimize a latent state over the energy landscape, then leverages the refined latent state for token generation. By performing such energy-guided pre-training on large-scale corpora, ESMs learn generally effective energy landscape where favorable tokens correspond to lower energy values. To explore the scaling behavior of ESMs, we pre-train ESMs on three large-scale corpora, then evaluate across 26 downstream tasks spanning a broad range of domains. Our results demonstrate that ESMs have promising scaling behaviors with increasing model size and training data, and outperform GPTs on most downstream tasks. These findings highlight ESMs as a promising architecture for large-scale pre-training and general purpose problem solving. We fully release the model weights and training infrastructure to enable the community to completely reproduce and extend our results. GitHub: https://github.com/datamllab/openebfm HuggingFace: https://huggingface.co/collections/guan-wang/openebfm
Oct 7, 2026

Sequential Pretraining Favors Large Models

Large neural networks often acquire capabilities that small models fail to learn. Does this stem from large models learning more representative features, or from being more robust to unaccounted-for adverse effects introduced during training? We define and quantify one such adverse effect, primacy bias, as the extent to which exposure to early data distributions impairs later learning. We show that small models can allocate learning capacity inefficiently toward early distributions, whereas sufficiently overparameterized models are robust to this effect. This inefficiency is particularly consequential in pretraining, where foundation models often encounter heterogeneous data distributions sequentially rather than jointly. As a result, small foundation models can struggle to learn distributions encountered late in training, which is particularly harmful when later data emphasizes desirable capabilities such as code, mathematics, and reasoning. Motivated by these findings, we introduce Exposure Therapy (ET), a simple regularization that promotes more efficient allocation of learning capacity during sequential pretraining. We demonstrate that ET improves foundation models' performance on late data distributions as well as overall capability in models up to the billion-parameter scale. Overall, our results suggest that some benefits of large foundation models may arise from greater robustness to adverse training effects, rather than from learning more representative features, and that improved training algorithms can recover some of these advantages in smaller models.
Oct 6, 2026

Wieszcz-XIX: A 3.1-Billion-Word Corpus of Pre-1918 Polish and Temporally Bounded Language Models Trained From Scratch

Historical Polish is well documented as a language but annotated in machine-readable form only to about a million words for the period this paper covers; the rest sits behind optical character recognition of variable quality. We present Wieszcz-XIX, a corpus of 6.75 billion tokens (about 3.1 billion words) in 294,369 documents, most of them periodical issues, of Polish published from 1800 to 1918, assembled from Wolne Lektury and the Internet Archive by a pipeline that filters, deduplicates, audits for post-1918 leakage and splits at the document level. It is over three orders of magnitude larger than the annotated corpus of the same period, and we quantify its defects: recognition corruption against a false-positive floor, near-identical duplication, which is removed, and post-1918 leakage, which is excluded from the training corpus itself down to a known residue of 0.04 to 0.38% of its bytes, found in the transcribed source, so the published corpus is the trained one document for document. On a hand-corrected sample the character error rate is 0.68% where the text is legible, and 45% of the sampled passages cannot be corrected. On it we train a ladder of decoder-only models from 47M to 349M parameters from scratch, and measure their temporal boundedness. Against two modern Polish base models, one far larger, the 349M shows a crossover, as does the 107M against the comparator of its size: post-1918 vocabulary costs them about 3.1 bits per byte more than period vocabulary, a gap the comparators do not show, and period vocabulary costs them fewer bits than it costs the comparators. Shown period text, the models keep its spelling and the comparators only partly. Adding parameters gains about twice as much as a second pass over the data. We release the corpus, code and weights. Content warning: the models reproduce period prejudice, including antisemitic statements.
Oct 6, 2026

Tiny-Scale Chinese BERT Pretraining: A Controlled Comparison of MLM, WWM, and MacBERT Strategies

Pretraining strategies significantly impact the quality of language models, yet existing comparisons of Masked Language Modeling (MLM), Whole Word Masking (WWM), and MacBERT-style replacement have focused primarily on base-scale models (>=110M parameters). This paper presents a controlled comparison of these three strategies on a tiny-scale Chinese BERT model (4 layers, 256 hidden dimensions, 8.7M parameters). Under identical architecture, corpus (1.29M sentences from Chinese Wikipedia), and hyperparameters, we train three models from scratch and evaluate them across five intrinsic dimensions: perplexity, MLM hit rate, semantic discrimination, grammatical judgment, and contextual sensitivity. At tiny scale, MLM achieves the best overall intrinsic performance (winning 3 of 5 dimensions), while WWM excels in both perplexity (1.27 vs. 2.10, a 39.5% improvement) and MLM hit rate (22% vs. 16%). Notably, MacBERT under a severely limited synonym dictionary (222 entries, 3.3% coverage) exhibits severe perplexity degradation (47.23, 22x higher than MLM), yielding a ranking (MLM > WWM >> MacBERT) that differs markedly from the established base-scale conclusion (MacBERT > WWM > MLM). We further identify a critical evaluation pitfall: MacBERT achieves the lowest training loss (2.17) yet the highest perplexity (47.23), revealing that training loss alone is unreliable under mixed replacement strategies. All models and corpus are publicly available at https://huggingface.co/eebyp.
Oct 5, 2026

Domain adaptation of Russian ModernBERT for long legal documents

We investigate whether continued pretraining on Russian legislative documents improves a Russian ModernBERT encoder on legal text. The adapted model, RuModernBERT-ruLaw, was trained on a corpus reported to contain 304,382 legislative documents and 194,425,905 corpus tokens. Corpus token counts are distinguished from positions produced by the model tokenizer. We compare the original and adapted encoders on a fixed external collection of 1,031 court-decision segments. Both models receive the same hidden positions in each of five masking realizations. At maximum input lengths of 512, 2,048, and 8,192 tokens, mean masked-token cross-entropy decreases by 0.10942, 0.07052, and 0.06604 natural-log units, respectively. The reported 95% intervals summarize sensitivity to masking on this fixed collection; they do not quantify uncertainty across document collections. A second evaluation addresses legal-entity extraction. The original and adapted models achieve entity-level F1 scores of 0.99852 and 0.99820. However, 99.95% of test spans have the same normalized surface form and class in the training split. This evaluation therefore provides limited evidence about transfer to previously unseen forms. The paper explains the masking objective, overlapping windows, averaging rules, and exact entity-boundary scoring using editable diagrams and clearly marked illustrative examples. The comparison supports lower masked-token prediction loss for the studied pair of models and collection. It does not isolate the contribution of distant context or establish practical legal utility.
Oct 5, 2026

Introducing Code-Switched Contexts to Cognitively-Inspired Bilingual Model Training

During language acquisition, bilingual children are regularly exposed to code-switched input and use it as a cognitive scaffold to accelerate vocabulary growth and cross-linguistic syntactic mapping. In contrast, computational bilingual models are conventionally pretrained on interleaved monolingual corpora. While introducing synthetic code-switching during pretraining has become a promising strategy to enhance cross-lingual alignment and downstream performance, the structural and developmental parameters governing the success remain poorly understood. In this work, we investigate the efficiency of training with synthetic code-switched data across two typologically distinct language pairs by controlling two key variables: the structural location of code-switches and the dynamic switching rate across training stages. Our results show that training with code-switched data improves cross-lingual alignment for typologically close languages.
Oct 5, 2026

LightMTP: Lightweight Latent Multi-Token Prediction

Next-token prediction (NTP) is the standard pretraining objective for large language models, yet it provides an explicit training signal only for the immediate next token, which can lead models to exploit local patterns instead of capturing longer-range structure and ideas. Multi-token prediction (MTP) addresses this by training models to predict several future tokens. However, existing MTP methods often introduce a large number of new parameters with limited improvements in downstream performance. Latent MTP approaches address this efficiency issue by encoding future tokens into a vector representation. However, these approaches usually rely on external helper models for future token encoding. We propose LightMTP, a lightweight, i.e., parameter-efficient, latent MTP approach that bootstraps the future token representations from the model's own hidden states. Our two LightMTP variants extend supervision to more future tokens without requiring the additional computational overhead of conventional MTP nor the external supervision latent MTP normally relies on. LightMTP adds at most 1% extra parameters, retains better performance on general language modeling benchmarks, and achieves similar gains in planning, coding, and reasoning.
Oct 5, 2026

Learning Decision-Stump Thresholds in Context: Dynamics of Softmax Attention

Estimating a decision threshold requires locating observations near an unknown boundary. We study how gradient-based pretraining learns this statistical rule in a two-parameter softmax-attention model with a fixed feature and inequality direction. Pretraining uses labeled contexts and their true thresholds; a fresh threshold must be inferred from context alone. Under a large-resolution initialization, constant-step gradient descent on mm tasks with nn examples each produces a frozen estimator with error O~((m∧n)−1+N−1)\widetilde O((m\wedge n)^{-1}+N^{-1}) for each fixed interior threshold and every fresh-context size NN. The two terms separate finite-pretraining accuracy from fresh-context localization. The mechanism is coordinated parameter divergence: population training calibrates the relative label and feature scores, then increases the attention scale as t1/4t^{1/4}, giving population threshold error O(t−1/4)O(t^{-1/4}). To transfer this mechanism to a fixed finite corpus, we control gradient errors relative to the shrinking directions of progress at successive parameter scales. This certifies a growing training interval without requiring long-time tracking of the population trajectory. We also identify the boundary limitation of the one-head model and explain statistically what a reflected symmetrization could achieve.
Oct 5, 2026

Learning to Learn a Language

We present the Prior-Fitted Language Model (PFLM), a 300M-parameter byte-level transformer pretrained only on samples from a synthetic non-linguistic prior. Given a prefix of real text, it learns to predict the language in context with frozen weights, having never seen a word of any real language. Every training sequence is generated by a recurrent structural causal model drawn fresh from a distribution over such models. The model never sees the same language twice during training, so the only way to predict the continuation is to infer the language from the prefix. Samples from this prior share the statistical signatures of natural text: Zipfian frequencies, slow entropy-rate convergence, and long-range dependence. On Wikipedia in six languages, bits per byte fall from the uniform eight to between 0.9 and 2.4 at one million bytes of context. Given numerals instead of text, PFLM learns to count, to compare magnitudes, and to add approximately. It predicts deterministic sequences like Rudin-Shapiro or the prime indicator, and it compresses six non-text domains, from source code to speech, below gzip and PPMd. The model has not learned a language. It has learned to learn one.