Language Model Pretraining

Follow this topic Ask about this topic

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

448 papers72 papers in the last four weeks, up 57% on the four weeks before. 0.4% of all new papers.

Apr 29, 2026

Decoupling the Benefits of Subword Tokenization for Language Model Training via Byte-level Simulation

Subword tokenization is an essential part of modern large language models (LLMs), yet its specific contributions to training efficiency and model performance remain poorly understood. In this work, we decouple the effects of subword tokenization by isolating them within a controlled byte-level pretraining pipeline. We formulate and test hypotheses across various dimensions, including sample throughput, vocabulary scaling, and the linguistic prior of subword boundaries. By simulating these effects in a byte-level setting, we refine our understanding of why subword models outperform raw byte models and offer insights to improve the pretraining of future byte-level and subword models. Specifically, our experiments highlight the critical role of increased training throughput and the integration of subword boundaries as either explicit priors or inductive biases.

Théo Gigant, Bowen Peng, Jeffrey Quesnelle

Nous Research

Language Model Pretraining Efficient Language Model Training Language Modeling Subword Tokenization

Apr 27, 2026

Staged Factorial Screening for Budget-Constrained Micro-Pretraining

Budget-constrained micro-pretraining often requires triaging many candidate recipes on a shared accelerator before larger search budgets are spent. We study whether a staged fractional-factorial workflow can recover stable early effect structure in this setting. On a fixed autoresearch-derived single-GPU training loop, we run 613 experiments across pilot and follow-up screens at 2, 5, and 10 minutes; full 16-condition seeded reruns at 5 and 10 minutes; targeted seeded anchor checks; same-host greedy and matched-cost random baselines; a 60-minute bridge package; and bounded Windows A100 and Linux L40S anchor continuations through 24 hours. Main penalties from total batch, depth, and width are largest at short budgets and relax as budget increases. Within the predeclared seeded full-screen families, D, A, B, and C retain non-zero estimates at 5 and 10 minutes after within-budget Benjamini-Hochberg correction, while E does not. Random search can reach strong incumbents in this 32-condition space, but repeatedly in the same low-penalty region and without factor attribution. The 60-minute bridge anchor has the lowest mean, although that package does not separate workflow refinement from the larger bridge model's capacity advantage. In bounded 12-hour and 24-hour three-anchor continuations on both hosts, the bridge has the lowest sample mean while the non-bridge ordering stays host-sensitive. We therefore present a bounded methods result: use short designed screens to identify high-penalty directions, confirm promising anchors under repeated runs, and refine locally inside the reduced space. The evidence supports a bridge-centered recommendation through 24 hours on two hosts, not hardware-invariant ranking or general hyperparameter-optimization superiority.

Felipe Chavarro Polania

Hewlett Packard Enterprise

Language Model Pretraining Hyperparameter Optimization

Apr 27, 2026

GEM: Geometric Entropy Mixing for Optimal LLM Data Curation

LLM pre-training efficacy increasingly depends on data composition rather than sheer volume. Yet, optimal mixing is hindered by categorization flaws: human taxonomies suffer from ontological misalignment, and Euclidean clustering fails to address embedding anisotropy. We introduce GEM (Geometric Entropy Mixing), a framework reformulating data curation as a variational problem on the hypersphere augmented with a mixing-balance regularizer. By decoupling the generative prior and optimizing the objective via a provable MM (Minorize-Maximize) algorithm, GEM effectively counteracts the cluster collapse to discover balanced semantic structures invisible to Euclidean heuristics. We employ teacher-student distillation to scale this geometric fidelity to web-scale corpora and introduce the Geometric Influence Score (GIS) for interpretable taxonomy generation. Experiments with 1.1B-parameter models demonstrate that GEM establishes a new state-of-the-art when integrated into mixing strategies like DoReMi and RegMix, improving average downstream accuracy by up to 1.2% and offering a robust coordinate system for predictable data mixing.

Yue Min, Ziyun Qiao, Ruining Chen +1

Wizard Quant, Beijing, China · The Hong Kong University of Science and Technology, Hong Kong SAR, China · Peking University, Beijing, China +1

Language Model Pretraining Training Data Curation Geometric Representation Learning Data Mixture Optimization

Apr 23, 2026

PermaFrost-Attack: Stealth Pretraining Seeding(SPS) for planting Logic Landmines During LLM Training

Aligned large language models (LLMs) remain vulnerable to adversarial manipulation, and their reliance on web-scale pretraining creates a subtle but consequential attack surface. We study Stealth Pretraining Seeding (SPS), a threat model in which adversaries distribute small amounts of poisoned content across stealth websites, increasing the likelihood that such material is absorbed into future training corpora derived from sources such as Common Crawl. Because each individual payload is tiny, diffuse, and superficially benign, the attack is difficult to detect during dataset construction or filtering. The result is a latent form of poisoning that remains largely invisible under standard evaluation, yet can later be activated by a precise trigger such as <00TRIGGER00>. We call this attack PermaFrost, reflecting its latent and reactivatable nature. We study it through PermaFrost-Attack, a controlled framework for latent conceptual poisoning, together with three geometric diagnostics: Thermodynamic Length, Spectral Curvature, and the Infection Traceback Graph. Across multiple model families and scales, we show that this controlled SPS proxy can induce persistent unsafe behavior that often remains hidden under standard evaluation. Our results identify SPS as a practical and underappreciated threat to future foundation models. This paper introduces a novel geometric diagnostic lens for systematically examining latent model behavior, providing a principled foundation for detecting, characterizing, and understanding vulnerabilities that may remain invisible under standard evaluation.

Harsh Kumar, Rahul Maity, Tanmay Joshi +4

Manipal University Jaipur · NIT, Karnataka · Pragya Lab, BITS Pilani Goa, India +2

Language Model Pretraining LLM Backdoor Attacks LLM Auditing Data Contamination in Language Models Data Poisoning Attacks

Apr 23, 2026

Decoupled DiLoCo for Resilient Distributed Pre-training

Modern large-scale language model pre-training relies heavily on the single program multiple data (SPMD) paradigm, which requires tight coupling across accelerators. Due to this coupling, transient slowdowns, hardware failures, and synchronization overhead stall the entire computation, wasting significant compute time at scale. While recent distributed methods like DiLoCo reduced communication bandwidth, they remained fundamentally synchronous and vulnerable to these system stalls. To address this, we introduce Decoupled DiLoCo, an evolution of the DiLoCo framework designed to break the lock-step synchronization barrier and go beyond SPMD to maximize training goodput. Decoupled DiLoCo partitions compute across multiple independent

learners'' that execute local inner optimization steps. These learners asynchronously communicate parameter fragments to a central synchronizer, which circumvents failed or straggling learners by aggregating updates using a minimum quorum, an adaptive grace window, and dynamic token-weighted merging. Inspired by

chaos engineering'', we achieve significantly improved training efficiency in failure-prone environments with millions of simulated chips with strictly zero global downtime, while maintaining competitive model performance across text and vision tasks, for both dense and mixture-of-expert architectures.

Arthur Douillard, Keith Rush, Yani Donchev +14

1Google DeepMind · 2Google Research

Language Model Pretraining Distributed Optimization High-Performance Computing Fault-Tolerant Distributed Training Distributed Training

Apr 23, 2026

Listen and Chant Before You Read: The Ladder of Beauty in LM Pre-Training

We show that pre-training a Transformer on music before language significantly accelerates language acquisition. Using piano performances (MAESTRO dataset), a developmental pipeline -- music

\to

poetry

\to

prose -- yields a

17.5\%

perplexity improvement over random initialization (

p < 0.001

, 5 seeds), with music and poetry improving orthogonal model components (internal computation and embeddings, respectively). Convergence tests confirm that this is not a transient head start: at

d\!=\!64

, multi-seed validation (5 seeds) shows a persistent 5.5% gap at plateau (

p = 0.017

), with the pipeline converging faster and to a lower loss in every run. Real music matches the transfer ceiling of synthetic patterns with one-third the data, and scaling experiments reveal that optimal pre-training data volume shifts with model capacity (

-3\% \to +3\% \to +6\%

advantage of larger datasets from

d\!=\!16

to

d\!=\!64

). Across the scales we study (

d\!\in\!\{16,32,64\}

, up to

{\sim}400

K parameters), these results suggest a capacity-dependent data curation principle and indicate that structured human creative outputs can provide an efficient pre-training substrate for small language models; stronger conclusions at modern pre-training scale will require substantially larger experiments.

Yoshinori Nomura

Mirage Mountain Technologies Inc.

Language Model Pretraining Audio-Language Models Curriculum Learning

Apr 22, 2026

Toward Cross-Lingual Quality Classifiers for Multilingual Pretraining Data Selection

As Large Language Models (LLMs) scale, data curation has shifted from maximizing volume to optimizing the signal-to-noise ratio by performing quality filtering. However, for many languages, native high quality data is insufficient to train robust quality classifiers. This work investigates the idea that quality markers in embedding space may show cross-lingual consistency, which would allow high-resource languages to subsidize the filtering of low-resource ones. We evaluate various filtering strategies, including cross-lingual transfer, third quartile sampling (Q3), and retention rate tuning. Our results demonstrate that massive multilingual pooling frequently outperforms monolingual baselines in both rank stability and aggregate accuracy for a 1B model trained on 103B tokens, delivering gains for high resource languages (1.2% increase in aggregate normalized accuracy for French) and matching or exceeding monolingual baselines for low-resource languages. However, we find that scale alone does not guarantee stability. Furthermore, for high-resource languages like French, we show that refining the decision boundary through third quartile sampling (Q3) or tuning the retention rate is necessary to fully leverage the multilingual signal.

Yassine Turki, Vinko Sabolčec, Bettina Messmer +1

Machine Learning Optimization Lab Ecole Polytechnique Fédérale de Lausanne (EPFL) Lausanne, Switzerland

Language Model Pretraining Multilingual Language Models Low-Resource Language Processing Cross-Lingual Transfer

Apr 20, 2026

Heterogeneity in Formal Linguistic Competence of Language Models: Is Data the Real Bottleneck?

Large Language Models (LLMs) exhibit a puzzling disparity in their formal linguistic competence: while they learn some linguistic phenomena with near-perfect mastery, they often perform below chance on others, even after training on trillions of tokens. In this work, we investigate whether these failures stem from inherent architectural limitations or simply the scarcity of these specific grammatical constructions in web-scale corpora. We pre-train simple GPT-2 Small (124M) models on a 100M-token random sample of the FineWeb corpus and intervene by injecting a minimal amount (1%) of synthetic data targeting specific linguistic phenomena. We find that this targeted intervention substantially improves model performance in 8 out of the 9 worst-performing BLiMP paradigms - notably the accuracy on a specific paradigm, only_npi_scope, surges from 20.9% to 69.4%. Furthermore, we observe that these interventions generally preserve or slightly improve aggregate performance. However, while we also identify a resistant phenomenon, principle_A_c_command, whose performance remains below chance even after our data augmentation, our findings do serve as an optimistic existence proof that even small language models can substantially improve on those linguistic phenomena on which models typically perform poorly, provided the pre-training data contains sufficient exposure to them. This suggests that efforts towards human-scale language modeling may benefit greatly by focusing on data composition. The code to reproduce our results is open-sourced at https://github.com/kowndinya-renduchintala/heterogeneity-in-formal-linguistic-competence.

H S V N S Kowndinya Renduchintala, Sumit Bhatia

Media and Data Science Research Adobe Inc., India

Language Model Pretraining Language Modeling

Apr 19, 2026

Copy First, Translate Later: Interpreting Translation Dynamics in Multilingual Pretraining

Large language models exhibit impressive cross-lingual capabilities. However, prior work analyzes this phenomenon through isolated factors and at sparse points during training, limiting our understanding of how cross-lingual generalization emerges--particularly in the early phases of learning. To study the early trajectory of linguistic and translation capabilities, we pretrain a multilingual 1.7B model on nine diverse languages, capturing checkpoints at a much finer granularity. We use word-level translation as a testbed, introducing a novel dataset to trace how translation develops over training through behavioral analyses, model-component analysis, and parameter-based ablations. We find that the model quickly acquires basic linguistic capabilities in parallel with token-level copying, while translation develops in two distinct phases: an initial phase dominated by copying and surface-level similarities, and a second phase in which more generalizing translation mechanisms are developed while copying is refined. Together, these findings provide a fine-grained view of how cross-lingual generalization develops during multilingual pretraining.

Felicia Körner, Maria Matveev, Florian Eichin +3

MaiNLP, Center for Information and Language Processing, LMU Munich, Germany · Munich Center for Machine Learning (MCML) · Department of Mathematics, LMU Munich, Germany +2

Language Model Pretraining Multilingual Language Models Machine Translation Cross-Lingual Transfer

Apr 19, 2026

Rethinking Data Curation in LLM Training: Online Reweighting Offers Better Generalization than Offline Methods

Data curation is a critical yet under-explored area in large language model (LLM) training. Existing methods, such as data selection and mixing, operate in an offline paradigm, detaching themselves from training. This separation introduces engineering overhead and makes the curation brittle: the entire pipeline must be re-run under model/task shifts. Moreover, offline methods alter data size through hard filtering or resampling, often sacrificing data diversity and harming generalization. We propose to rethink data curation as an online reweighting problem, where sample importance is dynamically adjusted during training via loss weighting rather than static pre-processing. Specifically, we introduce ADAPT (Adaptive Data reweighting for Pretraining and FineTuning), a dynamic online framework that reweights training samples with adaptive per-sample learning rates guided by similarity-based quality signals, without changing the number of training samples. Unlike offline methods that enforce a static data distribution, ADAPT acts as an implicit curriculum learner, progressively shifting focus from coarse-grained patterns to fine-grained semantic distinctions as the model evolves. Experiments on both instruction tuning and large-scale pretraining show that ADAPT consistently outperforms offline selection/mixing and prior online methods, achieving stronger cross-benchmark generalization under equal FLOPs.

Wanru Zhao, Yihong Chen, Yuzhi Tang +6

University of Cambridge · OATML, University of Oxford · University of Toronto +2

Language Model Pretraining LLM Fine-Tuning Training Data Curation

Apr 17, 2026

In Search of Lost DNA Sequence Pretraining

DNA sequence encoding is fundamental to gene function prediction, protein synthesis, and diverse downstream biological tasks. Despite the substantial progress achieved by large-scale DNA sequence pretraining, existing studies have overwhelmingly emphasized pretraining scale and custom downstream evaluation datasets, while neglecting some essential components of the pretraining paradigm. In this paper, we reveal three critical yet heretofore overlooked problems in DNA pretraining: inappropriate downstream datasets, inherent flaws in the neighbor-masking strategy, and the lack of detailed discussion on vocabulary. Therefore, we undertake comprehensive investigations and propose principled guidelines, including selection criteria for evaluation datasets, guiding task design, and in-depth vocabulary analysis. Extensive experiments validate the significance of our identified problems and support the rationale behind our recommendations. Finally, we introduce a standardized testbed that enables reproducible and rigorous benchmarking of DNA pretraining methods to advance the development of genomic foundation models.

Zhijiang Tang, Jiaxin Qi, Yan Cui +3

Computer Network Information Center, Chinese Academy of Sciences, Beijing, China · Hangzhou Institute for Advanced Study, University of Chinese Academy of Sciences, Zhejiang, China

Language Model Pretraining Masked Language Modeling Benchmark Design

Apr 17, 2026

Target-Oriented Pretraining Data Selection via Neuron-Activated Graph

Everyday tasks come with a target, and pretraining models around this target is what turns them into experts. In this paper, we study target-oriented language model (LM) pretraining by introducing Neuron-Activated Graph Ranking (NAG-based Ranking), a training-free and interpretable framework for target pretraining data selection. Rather than using black-box representations, our approach directly characterizes each target input by a sparse set of high-impact neurons in any off-the-shelf LLMs. Concretely, we quantify neuron impact and select the most influential neurons across layers into a compact Neuron-Activated Graph (NAG), and rank candidate data by NAG similarity to target examples. We conduct experiments across six benchmarks, where our NAG-based Ranking improves target-oriented pretraining by 4.9% on average over random sampling, and also outperforms state-of-the-art baselines by 5.3% accuracy on HellaSwag. It also remains effective under a more applicable multi-target setting, where our best setup surpasses two baselines by 1.1% and 4.1%, respectively. Furthermore, we provide a comprehensive analysis on why and how our NAG works, e.g., deactivating NAG-selected neurons (only 0.12% of all) causes a 23.5% performance collapse, and restricting NAG to the final layer incurs a 4.1% average drop, indicating that NAG captures a sparse "functional backbone" for learning target features. We release the code at https://github.com/asillycat/NAG.

Zijun Wang, Haoqin Tu, Weidong Zhou +7

1ByteDance · 2UC Santa Cruz · 3UNC-Chapel Hill

Language Model Pretraining Training Data Selection

Apr 14, 2026

KoCo: Conditioning Language Model Pre-training on Knowledge Coordinates

Standard Large Language Model (LLM) pre-training typically treats corpora as flattened token sequences, often overlooking the real-world context that humans naturally rely on to contextualize information. To bridge this gap, we introduce Knowledge Coordinate Conditioning (KoCo), a simple method that maps every document into a three-dimensional semantic coordinate. By prepending these coordinates as textual prefixes for pre-training, we aim to equip the model with explicit contextual awareness to learn the documents within the real-world knowledge structure. Experiment results demonstrate that KoCo significantly enhances performance across 10 downstream tasks and accelerates pre-training convergence by approximately 30%. Furthermore, our analysis indicates that explicitly modeling knowledge coordinates helps the model distinguish stable facts from noise, effectively mitigating hallucination in generated outputs.

Yudong Li, Jiawei Cai, Linlin Shen

School of Computer Science and Software Engineering, Shenzhen University · Department of Electrical Engineering, Tsinghua University · School of Artificial Intelligence, Shenzhen University +1

Language Model Pretraining Knowledge Augmentation for Language Models

Apr 14, 2026

Nemotron 3 Super: Open, Efficient Mixture-of-Experts Hybrid Mamba-Transformer Model for Agentic Reasoning

We describe the pre-training, post-training, and quantization of Nemotron 3 Super, a 120 billion (active 12 billion) parameter hybrid Mamba-Attention Mixture-of-Experts model. Nemotron 3 Super is the first model in the Nemotron 3 family to 1) be pre-trained in NVFP4, 2) leverage LatentMoE, a new Mixture-of-Experts architecture that optimizes for both accuracy per FLOP and accuracy per parameter, and 3) include MTP layers for inference acceleration through native speculative decoding. We pre-trained Nemotron 3 Super on 25 trillion tokens followed by post-training using supervised fine tuning (SFT) and reinforcement learning (RL). The final model supports up to 1M context length and achieves comparable accuracy on common benchmarks, while also achieving up to 2.2x and 7.5x higher inference throughput compared to GPT-OSS-120B and Qwen3.5-122B, respectively. Nemotron 3 Super datasets, along with the base, post-trained, and quantized checkpoints, are open-sourced on HuggingFace.

NVIDIA, :, Aakshita Chandiramani +544

Hybrid Transformer-State Space Model Architectures Mixture-of-Experts Language Models Language Model Pretraining Hybrid Mamba-Transformer Architectures LLM Inference Acceleration

Apr 13, 2026

ks-pret-5m: a 5 million word, 12 million token kashmiri pretraining dataset

We present KS-PRET-5M, the largest publicly available pretraining dataset for the Kashmiri language, comprising 5,090,244 (5.09M) words, 27,692,959 (27.6M) characters, and a vocabulary of 295,433 (295.4K) unique word types. We assembled the dataset from two source classes: digitized archival and literary material, encompassing literature, news, biographies, novels, poetry, religious scholarship, and academic writing, recovered from the proprietary InPage desktop-publishing format using the converter of Malik~\cite{malik2024inpage}, and Unicode-native text collected from Kashmiri-language web sources. All text was processed through an eleven-stage cleaning pipeline that achieves a mean Kashmiri script ratio of 0.9965, reducing Devanagari contamination to 146 characters across the full dataset. We tokenized the dataset empirically using google/muril-base-cased, yielding a subword ratio of 2.383 tokens per word and a total of approximately 12.13 million subword tokens, substantially higher than prior estimates derived from non-Kashmiri Perso-Arabic analogues. KS-PRET-5M is released as a single continuous text stream under CCBY4.0 to support language model pretraining, tokenizer training, and computational linguistic research for Kashmiri.

Haq Nawaz Malik, Nahfid Nissar

Low-Resource Language Processing Language Model Pretraining

Apr 12, 2026

Advancing Polish Language Modeling through Tokenizer Optimization in the Bielik v3 7B and 11B Series

The development of the Bielik v3 PL series, encompassing both the 7B and 11B parameter variants, represents a significant milestone in the field of language-specific large language model (LLM) optimization. While general-purpose models often demonstrate impressive multilingual capabilities, they frequently suffer from a fundamental architectural inefficiency: the use of universal tokenizers. These tokenizers, typically designed to cover a broad spectrum of languages, often fail to capture the morphological nuances of specific languages like Polish, leading to higher fertility ratios, increased inference costs, and restricted effective context windows. This report details the transition from the universal Mistral-based tokenization to a dedicated Polish-optimized vocabulary for the Bielik v3 models, exploring the FOCUS-based embedding initialization, the multi-stage pretraining curriculum, and the subsequent post-training alignment involving Supervised Fine-Tuning, Direct Preference Optimization, and Reinforcement Learning through Group Relative Policy Optimization with verifiable rewards.

Krzysztof Ociepa, Łukasz Flis, Remigiusz Kinas +2

SpeakLeash · Azurro · ACK Cyfronet AGH +2

Language Modeling Subword Tokenization Language Model Pretraining LLM Fine-Tuning

Apr 12, 2026

Bridging Linguistic Gaps: Cross-Lingual Mapping in Pre-Training and Dataset for Enhanced Multilingual LLM Performance

Multilingual Large Language Models (LLMs) struggle with cross-lingual tasks due to data imbalances between high-resource and low-resource languages, as well as monolingual bias in pre-training. Existing methods, such as bilingual fine-tuning and contrastive alignment, can improve cross-lingual performance, but they often require extensive parallel data or suffer from instability. To address these challenges, we introduce a Cross-Lingual Mapping Task during the pre-training phase, which enhances cross-lingual alignment without compromising monolingual fluency. Our approach bi-directionally maps languages within the LLM embedding space, improving both language generation and comprehension. We further propose a Language Alignment Coefficient to robustly quantify cross-lingual consistency, even in limited-data scenarios. Experimental results on machine translation (MT), cross-lingual natural language understanding (CLNLU), and cross-lingual question answering (CLQA) show that our model achieves gains of up to 11.9 BLEU points in MT, 6.72 points in CLQA BERTScore-Precision, and more than 5% in CLNLU accuracy over strong multilingual baselines. These findings highlight the potential of incorporating cross-lingual objectives into pre-training to improve multilingual LLMs.

Weihua Zheng, Chang Liu, Zhengyuan Liu +5

Singapore University of Technology and Design, Singapore · ByteDance, Singapore · Agency for Science, Technology and Research, Singapore

Multilingual Language Models Cross-Lingual Representation Alignment Cross-Lingual Consistency in Language Models Language Model Pretraining

Apr 10, 2026

Nexus: Same Pretraining Loss, Better Downstream Generalization via Common Minima

The foundational capabilities of large language models are acquired during pretraining on internet-scale, highly heterogeneous data mixtures. In this work, we investigate an interesting geometric question regarding the converged state of pretraining: Does the model converge to a common minimizer across all data sources (e.g., \cref{fig:cwa_illustration:close}), or merely a minimizer of the summed loss (e.g., \cref{fig:cwa_illustration:distant})? We hypothesize that the geometric "closeness" of task-specific minima is intrinsically linked to downstream generalization. We reveal that standard optimizers (e.g., AdamW) often converge to points where task-specific minima are distant from each other. To address this, we propose the Nexus optimizer, which encourages the closeness of these minima by maximizing gradient similarity during optimization. Experiments across models ranging from 130M to 3B parameters, various data mixtures and hyperparameter schedules, show that Nexus significantly boosts downstream performance, despite achieving the same pretraining loss (see \cref{fig:demo:benchmark}). Notably, on the 3B model, Nexus reduces the out-of-distribution loss by 0.012 and yields up to a 15.0% accuracy improvement on complex reasoning tasks (e.g., GSM8k). This finding challenges the reliance on pretraining loss as the sole proxy for model evaluation and demonstrates the importance of implicit biases in unlocking downstream generalization.

Huanran Chen, Huaqing Zhang, Xiao Li +3

Tsinghua University · ByteDance Seed

Language Model Pretraining Neural Network Generalization Deep Learning Optimization Gradient Alignment OOD Generalization

Apr 9, 2026

HiFloat4 Format for Language Model Pre-training on Ascend NPUs

Training large foundation models at low numerical precision is one of the most promising directions for reducing the compute and memory cost of modern AI. Recent 4-bit floating-point formats such as MXFP4 and NVFP4 can be applied to linear GEMM operations in LLMs, but their limited dynamic range introduces numerical instability that prior work addresses by stacking stabilization mechanisms, typically executed at higher precision and partially eroding the efficiency gains that motivate FP4. In this work, we argue that numerical format design is itself a first-class lever for stable FP4 training, and present the first systematic study of FP4 LLM pretraining on energy-efficient Huawei Ascend NPUs. We compare the recently proposed HiFloat4 (HiF4) format against both MXFP4 and NVFP4, holding one recipe fixed across all three formats across dense (OpenPangu-1B, Llama3-8B) and Mixture-of-Experts (Qwen3-MoE-30B) architectures and executing all linear and expert GEMMs in FP4. At matched storage --- NVFP4 and HiF4 both spend 4.5 bits per value --- the three formats differ far more in what they require before they will train at all than in final accuracy. HiF4 reaches a relative loss of 1.55% with no stabilization, below fully stabilized MXFP4 (1.79%) and below NVFP4 carrying the per-tensor scaling it cannot train without (2.00%); NVFP4 diverges under every combination of stochastic rounding and Hadamard transform we tried. Our results suggest that stable, accurate FP4 training does not require an ever-growing stack of stabilization techniques; it requires the right numerical format.

Mehran Taghian, Yunke Peng, Xing Huang +22

Huawei

Low-Precision Training LLM Quantization LLM Training Language Model Pretraining

Apr 9, 2026

Static Program Slicing Using Language Models With Dataflow-Aware Pretraining and Constrained Decoding

Static program slicing is a fundamental software engineering technique for isolating code relevant to specific variables. While recent learning-based approaches using language models (LMs) show promise in automating slice prediction, they suffer from inaccurate dependency modeling and unconstrained generation, where LMs fail to capture precise data flow relations and produce slices containing hallucinated tokens and statements. To address these challenges, we propose Sliceformer, a novel approach that reformulates static program slicing as a sequence-to-sequence task using small language models such as CodeT5+. Sliceformer introduces two key innovations that directly target the identified limitations. First, to improve dependency modeling, we design dataflow-aware pretraining objectives that leverage data flow graphs (DFG) to teach models data dependencies through dataflow-preserving statement permutation and dataflow-aware span corruption. Second, to eliminate hallucination, we develop a constrained decoding mechanism that enforces both lexical and syntactic constraints. We evaluate Sliceformer on Java and Python program slicing benchmarks, demonstrating consistent improvements over state-of-the-art baselines with up to 22% gain in ExactMatch.

Pengfei He, Shaowei Wang, Tse-Hsun Chen +1

Department of Computer Science, University of Manitoba, Winnipeg, Canada · School of Engineering and Computer Science, Concordia University, Montreal, Canada · School of Computer Science, University of Windsor, Windsor, Canada

Code Language Models Language Model Pretraining Constrained Decoding

Apr 9, 2026

What do Language Models Learn and When? The Implicit Curriculum Hypothesis

Large language models (LLMs) can perform remarkably complex tasks, yet the fine-grained details of how these capabilities emerge during pretraining remain poorly understood. Scaling laws on validation loss tell us how much a model improves with additional compute, but not what skills it acquires in which order. To remedy this, we propose the Implicit Curriculum Hypothesis: pretraining follows a compositional and predictable curriculum across models and data mixtures. We test this by designing a suite of simple, composable tasks spanning retrieval, morphological transformations, coreference, logical reasoning, and mathematics. Using these tasks, we track emergence points across four model families spanning sizes from 410M-13B parameters. We find that emergence orderings of when models reach fixed accuracy thresholds are strikingly consistent (

ρ= .81

across 45 model pairs), and that composite tasks most often emerge after their component tasks. Furthermore, we find that this structure is encoded in model representations: tasks with similar function vector representations also tend to follow similar trajectories in training. By using the space of representations derived from our task set, we can effectively predict the training trajectories of simple held-out compositional tasks throughout the course of pretraining (

R^2 = .68

-

.84

across models) without previously evaluating them. Together, these results suggest that pretraining is more structured than loss curves reveal: skills emerge in a compositional order that is consistent across models and readable from their internals.

Emmy Liu, Kaiser Sun, Millicent Li +4

Language Technologies Institute, Carnegie Mellon University · Department of Computer Science, Data Science and AI Institute, Johns Hopkins University · Khoury College of Computer Science, Northeastern University +1

Emergent Abilities in Language Models Language Model Pretraining

Apr 9, 2026

Improving DNS Exfiltration Detection via Transformer Pretraining

We study whether in-domain pretraining of Bidirectional Encoder Representations from Transformer (BERT) model improves subdomain-level detection of exfiltration at low false positive rates. While previous work mostly examines fine-tuned generic Transformers, it does not aim to isolate the effect of pretraining on the downstream task of classification. To address this gap, we develop a controlled pipeline where we freeze operating points on validation and transfer them to the test set, thus enabling clean ablations across different label and pretraining budgets. Our results show significant improvements in the left tail of the Receiver Operating Characteristic (ROC) curve, especially against randomly initialized baseline. Additionally, within pretrained model variants, increasing the number of pretraining steps helps the most when more labeled data are available for fine-tuning.

Miloš Tomić, Aleksa Cvetanović, Predrag Tadić

School of Electrical Engineering, University of Belgrade, Serbia

Network Intrusion Detection Domain-Adaptive Pretraining Language Model Pretraining Transformer

Apr 9, 2026

Soro: A Lightweight Foundation Model and Chatbot for Tajik

We present Soro, a family of Tajik-specialized conversational large language models (LLMs) designed for real-world deployment under tight compute and connectivity constraints in Tajikistan. Starting from open-weight Gemma 3 checkpoints, we perform Tajik-only continual pretraining on a curated 1.9-billion-token corpus spanning filtered web text, PDF documents, and curriculum-aligned educational materials, followed by supervised instruction tuning on 40K Tajik teacher-style examples. To enable rigorous evaluation despite the limited coverage of Tajik in standard benchmarks, we introduce a suite of Tajik benchmarks covering general knowledge, linguistic competence, and school- and university entrance-exam domains, and we open-source them on Hugging Face. Across these Tajik benchmarks, Soro substantially outperforms same-size Gemma 3 baselines while retaining strong English performance on standard datasets. We further show that FP8 and INT4 quantization of Soro preserves most Tajik-language gains while reducing memory requirements for edge deployment, supporting an ongoing education-sector pilot and planned scale-out across schools in Tajikistan.

Stanislav Liashkov, Haitz Sáez de Ocáriz Borde, Azizjon Azimi +3

Low-Resource Language Processing LLM Fine-Tuning Language Model Pretraining

Apr 9, 2026

SAGE: Sign-Adaptive Gradient for Memory-Efficient LLM Optimization

The AdamW optimizer, while standard for LLM pretraining, is a critical memory bottleneck, consuming optimizer states equivalent to twice the model's size. Although light-state optimizers like SinkGD attempt to address this issue, we identify the embedding layer dilemma: these methods fail to handle the sparse, high-variance gradients inherent to embeddings, forcing a hybrid design that reverts to AdamW and partially negates the memory gains. We propose SAGE (Sign Adaptive GradiEnt), a novel optimizer that resolves this dilemma by replacing AdamW in this hybrid structure. SAGE combines a Lion-style update direction with a new, memory-efficient

O(d)

adaptive scale. This scale acts as a "safe damper," provably bounded by 1.0, which tames high-variance dimensions more effectively than existing methods. This superior stability allows SAGE to achieve better convergence. On Llama models up to 1.3B parameters, our SAGE-based hybrid achieves new state-of-the-art perplexity, outperforming all baselines, including SinkGD hybrid, while significantly reducing optimizer state memory.

Wooin Lee, Hyun-Tae Kim

London, United Kingdom · Seoul, Republic of Korea

LLM Input Optimization Memory-Efficient Optimization Sign-Based Optimization Language Model Pretraining

Apr 8, 2026

Learning is Forgetting: LLM Training As Lossy Compression

Despite the increasing prevalence of large language models (LLMs), we still have a limited understanding of how their representational spaces are structured. This limits our ability to interpret how and what they learn or relate them to learning in humans. We argue LLMs are best seen as an instance of lossy compression, where over training they learn by retaining only information in their training data relevant to their objective(s). We show pre-training results in models that are optimally compressed for next-sequence prediction, approaching the Information Bottleneck bound on compression. Across an array of open weights models, each compresses differently, likely due to differences in the data and training recipes used. However even across different families of LLMs the optimality of a model's compression, and the information present in it, can predict downstream performance on across a wide array of benchmarks, letting us directly link representational structure to actionable insights about model performance. In the general case the work presented here offers a unified Information-Theoretic framing for how these models learn that is deployable at scale.

Henry C. Conklin, Tom Hosking, Tan Yi-Chern +5

Princeton University · Cohere

Language Model Pretraining LLM Compression Information Bottleneck Representation Geometry in Language Models Neural Network Compression

Apr 5, 2026

How Does "English (US)" Become the Default? Triangulating Structural Bias Towards American English Across the LLM Pipeline

Large language models (LLMs) are increasingly embedded in educational, professional, and public infrastructure, yet widely used platforms expose "English (US)" as a primary English setting despite the global diversity of English. We ask: How does "English (US)" become the default? We study this question as structural bias, examining how geopolitical histories of data curation, digital dominance, and linguistic standardization intersect with the LLM development pipeline. Using British English as a controlled reference, we construct a curated resource of 1,813 matched American English (AmE)--British English (BrE) variants and introduce DiAlign, a dynamic, training-free method for estimating regional alignment from distributional evidence. We triangulate the AmE preference across data exposure --> representation --> generation, jointly examining pretraining and post-training data, tokenizer behavior and provenance, model prediction cost, and generated language across developer countries, prompt conditions, domains and sources, linguistic categories, and registers. AmE is consistently favored across all six audited pretraining corpora and 21 post-training datasets, is generally represented more compactly by tokenizers, and receives lower prediction cost. It also remains the dominant generation default under neutral English prompting; British-English prompting shifts this preference toward BrE but does not consistently eliminate the AmE default. To our knowledge, this is the first rigorous pipeline-wide study of structural bias across major phases of LLM development. Our findings show that contemporary LLMs privilege AmE as the de facto norm, raising concerns about linguistic homogenization, epistemic injustice, and inequity in global AI deployment, while providing a rigorous basis for targeted component-level intervention.

Mir Tafseer Nayeem, Davood Rafiei

Department of Computing Science University of Alberta

Language Model Pretraining LLM Auditing Language Model Bias Evaluation Linguistic Bias in Language Models Language Model Post-Training

Apr 3, 2026

Olmo Hybrid: From Theory to Practice and Back

Recent work has demonstrated the potential of non-transformer language models, especially linear recurrent neural networks (RNNs) and hybrid models that mix recurrence and attention. Yet there is no consensus on whether the potential benefits of these new architectures justify the risk and effort of scaling them up. To address this, we provide evidence for the advantages of hybrid models over pure transformers on several fronts. First, theoretically, we show that hybrid models do not merely inherit the expressivity of transformers and linear RNNs, but can express tasks beyond both, such as code execution. Putting this theory to practice, we train Olmo Hybrid, a 7B-parameter model largely comparable to Olmo 3 7B but with the sliding window layers replaced by Gated DeltaNet layers. We show that Olmo Hybrid outperforms Olmo 3 across standard pretraining and mid-training evaluations, demonstrating the benefit of hybrid models in a controlled, large-scale setting. We find that the hybrid model scales significantly more efficiently than the transformer, explaining its higher performance. However, its unclear why greater expressivity on specific formal problems should result in better scaling or superior performance on downstream tasks unrelated to those problems. To explain this apparent gap, we return to theory and argue why increased expressivity should translate to better scaling efficiency, completing the loop. Overall, our results suggest that hybrid models mixing attention and recurrent layers are a powerful extension to the language modeling paradigm: not merely to reduce memory during inference, but as a fundamental way to obtain more expressive models that scale better during pretraining.

William Merrill, Yanhong Li, Tyler Romero +19

Allen Institute for AI · ETH Zürich · Lambda ML Team +2

Hybrid Language Models Hybrid Transformer-State Space Model Architectures Hybrid Transformer Architectures Language Modeling Hybrid Attention Language Model Scaling+2

Apr 1, 2026

Test-Time Scaling Makes Overtraining Compute-Optimal

Modern LLMs scale at test-time, e.g. via repeated sampling, where inference cost grows with model size and the number of samples. This creates a trade-off that pretraining scaling laws, such as Chinchilla, do not address. We present Train-to-Test (

T^2

) scaling laws that jointly optimize model size, training tokens, and number of inference samples under fixed end-to-end budgets.

T^2

modernizes pretraining scaling laws with pass@

k

modeling used for test-time scaling, then jointly optimizes pretraining and test-time decisions. Forecasts from

T^2

are robust over distinct modeling approaches: measuring joint scaling effect on the task loss and modeling impact on task accuracy. Across eight downstream tasks, we find that when accounting for inference cost, optimal pretraining decisions shift radically into the overtraining regime, well-outside of the range of standard pretraining scaling suites. We validate our results by pretraining heavily overtrained models in the optimal region that

T^2

scaling forecasts, confirming their substantially stronger performance compared to pretraining scaling alone. Finally, as frontier LLMs are post-trained, we show that our findings survive the post-training stage, making

T^2

scaling meaningful in modern deployments.

Nicholas Roberts, Sungjun Cho, Zhiqi Gao +7

University of Wisconsin-Madison · Stanford University

Scaling Laws Language Model Scaling Laws Test-Time Scaling LLM Training Language Model Pretraining

Apr 1, 2026

GPT-NL Public Corpus: A Permissively Licensed, Dutch-First Dataset for LLM Pre-training

We present the GPT-NL Public Corpus, the biggest permissively licensed corpus of Dutch language resources. The GPT-NL Public Corpus contains 21 Dutch-only collections totalling 36B preprocessed Dutch tokens not present in any other LLM pretraining corpus. Additionally, the corpus includes roughly 207B English, 232B Code, and 48B German/Danish tokens taken from existing sets which we further curated for compliance. This corpus includes curated data from large existing corpora like Common Corpus and Common Crawl, as well as newly created Dutch-specific collections. Most newly created Dutch collections consist of content collected in collaboration with organisations or synthetically augmented content. All data is collected and evaluated with the aim of facilitating the creation of (commercial) language models that are lawful, useful and non-harmful. All data included in the GPT-NL Public Corpus is sourced from datasets with permissive licensing and is curated and redistributed under a CC-BY license. The full dataset is publicly available on the Hugging Face Hub.

Jesse van Oort, Frank Brinkkemper, Erik de Graaf +2

TNO, The Netherlands · GPT-NL, The Netherlands · Instituut voor de Nederlandse Taal, The Netherlands +1

Language Model Pretraining Low-Resource Language Processing

Apr 1, 2026

Scalable Pretraining of Large Mixture of Experts Language Models on Aurora Super Computer

Pretraining Large Language Models (LLMs) from scratch requires massive amount of compute. Aurora super computer is an ExaScale machine with 127,488 Intel PVC (Ponte Vechio) GPU tiles. In this work, we showcase LLM pretraining on Aurora at the scale of 1000s of GPU tiles. Towards this effort, we developed Optimus, an inhouse training library with support for standard large model training techniques. Using Optimus, we first pretrained Mula-1B, a 1 Billion dense model and Mula-7B-A1B, a 7 Billion Mixture of Experts (MoE) model from scratch on 3072 GPU tiles for the full 4 trillion tokens of the OLMoE-mix-0924 dataset. We then demonstrated model scaling by pretraining three large MoE models Mula-20B-A2B, Mula-100B-A7B, and Mula-220B-A10B till 100 Billion tokens on the same dataset. On our largest model Mula-220B-A10B, we pushed the compute scaling from 384 to 12288 GPU tiles and observed scaling efficiency of around 90% at 12288 GPU tiles. We significantly improved the runtime performance of MoE models using custom GPU kernels for expert computation, and a novel EP-Aware sharded optimizer resulting in training speedups up to 1.71x. As part of the Optimus library, we also developed a robust set of reliability and fault tolerant features to improve training stability and continuity at scale.

Dharma Teja Vooturi, Dhiraj Kalamkar, Dipankar Das +1

Parallel Computing Lab (India) Intel Corporation

Language Model Pretraining LLM Training Efficient Language Model Training Mixture-of-Experts Language Models