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.

Nov 17, 2026

Structure--Residual Inductive Biases for Language Model Pretraining

Next-token prediction is relied upon to learn both predictive behavior and internal representations. We ask whether representation structure itself should be an explicit inductive bias. Motivated by the two-part coding view of Kolmogorov complexity, we study a structure--residual principle in which a representation separates structure from realization detail. We instantiate this principle in two complementary ways. Boundary Bottleneck Concept (BBC) routes information from token segments through boundary carriers while preserving local autoregressive computation. Continuous Concept (CC) regularizes representation geometry and induces a spectral decomposition into low-frequency structure and high-frequency residual. We derive coding bounds for both views and show that multiplicity-corrected boundary attention becomes exact as the within-segment path energy vanishes. Empirically, BBC and CC exhibit strikingly different depth preferences: BBC works best in early layers, whereas CC is most effective in later layers. At 300M training tokens, combining BBC and CC improves NLL by 1.44% on GPT-2 Medium. These gains persist with training runs up to 20B tokens. Representation diagnostics show that BBC increases the relative predictive importance of boundary states, while CC reduces spectral residual energy. Together, these results suggest that structure--residual biases can improve language-model pretraining beyond next-token prediction. Code: https://github.com/haihuang-ml/structure-residual
Nov 16, 2026

Latent Feedback Is a Compute and Data Amplifier

When high-quality data runs out, a common practice is to train on it for several epochs, but each repeat is worth less than the last. Latent feedback, which feeds each token's top-layer state into the next token's input, is another way to spend that compute, yet it has not been tested against a vanilla Transformer that spends the same compute on more epochs. We run this test across a ladder of model sizes, unique data and epochs, tuning both models at every size and amount of unique data. Latent feedback makes every FLOP and every token count for more except when unique data is scarce relative to model size, a regime that today's models are far from. We find that latent feedback serves as an amplifier: with unique data fixed, it reaches a Transformer's loss with less compute; with epochs fixed, it reaches that loss with less unique data. Furthermore, both savings grow with unique data per parameter and with repetition, and downstream tasks agree. To go beyond our grid, we fit one scaling law to both models, in which latent feedback enters as a Transformer with more parameters; it predicts that both savings keep growing with unique data and repetition. Fitted on models from 60M to 271M parameters, the law predicts the losses at a held-out 778M size to within 0.3% and the gap between the two models to within 0.05%. At most, latent feedback saves 48% of the compute and 47% of the unique data. When data, not compute, is the limit, latent feedback is the better use of a FLOP for a model of a given size.
Nov 15, 2026

Bioformer-RetroMAE: retrieval-oriented pretraining of a compact biomedical language model for literature search

Dense retrieval is increasingly used for biomedical literature search and retrieval-augmented generation (RAG). Most biomedical bi-encoders are built on BERTBase-sized encoders pretrained with masked language modeling (MLM). These encoders are costly for corpus-scale indexing and latency-sensitive query encoding, and MLM does not explicitly train the [CLS] vector that a bi-encoder uses to represent a whole text. We present Bioformer-RetroMAE, a compact (43M-parameter) retrieval-oriented biomedical encoder obtained by continually pretraining Bioformer8L} with RetroMAE on approximately 32 million PubMed titles and abstracts. After bi-encoder fine-tuning with BM25 hard negatives, we evaluated it on BioASQ Task 12B test questions by retrieving from all PubMed abstracts. Bioformer-RetroMAE outperformed BioBERT and PubMedBERT bi-encoders fine-tuned under identical settings on all metrics, despite having about 40% of their parameters and producing 512-dimensional rather than 768-dimensional vectors. Compared with PubMedBERT, it improved mean average precision (MAP) by 37.0% and Recall@100 by 29.1%. Compared with the original Bioformer8L}, RetroMAE pretraining improved MAP by 53.6% and Recall@100 by 32.8% without changing model size or inference cost. In hybrid retrieval, fusing Bioformer-RetroMAE with BM25 through weighted reciprocal rank fusion improved MAP and NDCG@10 over BM25 alone by 4.9% and 4.6%, respectively, larger gains than those obtained by fusing BM25 with PubMedBERT or BioBERT.
Nov 15, 2026

Data Attribution Should Steer Pretraining

Pretraining data curation decides what a model learns from, yet what makes a document worth learning from depends on what the model already knows. Influence-driven selectors already rank documents to improve one run; we argue that training-data attribution should steer pretraining by identifying testable data properties that become reusable curation criteria, evaluated on new data against cheaper selectors at matched total cost, discovery included. A document ranking is consumed by the run it was computed for, whereas an explained data property can be tested across corpora and checkpoints, especially where familiar quality signals diverge from measured usefulness. Because historical contribution, an understandable explanation, and future data value require distinct evidence, we specify a feedback loop that interprets attributed sources, tests candidate data changes, and states for five claim types what evidence each requires and what none establishes. We close with a falsifiable comparison against perplexity, classifier, and mixing baselines at equal total compute.
Nov 14, 2026

Understanding the Role of Reasoning Data in Pretraining

Reasoning data expose models to multi-step reasoning, but reasoning traces may cover only a subset of the available facts. How should pretraining balance atomic facts with reasoning traces? We study this question by constructing a setting in which every fact provides an atomic training signal, while only a subset appears in reasoning supervision, enabling us to measure generalization to transitions observed atomically but never used in training traces. For a linear transformer, we derive the weight assigned to atomic training signals that minimizes held-transition squared error and show that its leading-order optimum increases with reasoning-chain length and decreases with the fraction of facts held out from reasoning traces. These trends persist in a full transformer layer with an MLP, softmax attention, and cross-entropy loss. We further validate the role of reasoning complexity and data coverage in two natural-language settings: controlled multi-hop reasoning and mathematical pretraining. Together, our results provide a principled view of how reasoning-chain structure and data coverage should shape the allocation of training data during pretraining.
Nov 13, 2026

Geometrically Disentangling Concept Learning from the Language Modeling Loss

Language models acquire a vast array of concepts and abilities from next-token prediction alone. We introduce an interpretability framework that exposes how this single loss relates to the development of different concepts at different points during pretraining. We propose Concept Energy, a measure of the concept-specific alignment between a concept subspace and the average gradient outer product (AGOP) of the next-token-prediction loss. The concept subspace captures where the concept currently lives in the model's representation; AGOP captures the directions to which the loss is most sensitive. Across 19 checkpoints of Pythia-410M and -1B and 12 lexical, syntactic, and semantic concepts, we find that the first peak in Concept Energy generally occurs at or before the steepest increase in probe decodability. Furthermore, whereas probe performance usually increases monotonically throughout training, Concept Energy is not confined to a single peak; some concepts exhibit sustained or repeated engagement across pretraining, suggesting that concept learning may require continual reinforcement or occur at multiple points. Our results indicate that the loss contributes to concept learning in a myriad of ways that simple probe performance alone cannot reveal.
Nov 13, 2026

LECAT: Measuring LLM Category Learning from Pretraining Data

Humans know that airplanes are fast, that this extends to red airplanes, but that it does not hold for broken airplanes. How do LLMs acquire category knowledge that humans learn and apply effortlessly? We introduce LeCat, a two-part framework to analyze LLM's Learning of Category knowledge. First, we use LeCat to evaluate models on their ability to learn category knowledge, distinguishing between three cases: memorization of in-domain category knowledge, generalization to unseen instances, and ability to flexibly use knowledge in appropriate contexts. Then, we use LeCat to trace how this knowledge can be learned in pretraining from text that may encode category knowledge in various ways, such as verbatim repetition or exceptions, which we refer to as forms of "knowledge expression". We first assess open-source pretrained models using 199 curated category-feature pairs and analyze their training data. Our findings suggest that strengthening context and category-feature co-occurrence likelihood (PMI) improve learning for larger models and verbatim rule-like statements diminish flexible category learning. To systematically isolate these effects, we then train a 68M model from scratch with precise control over the categorical information each sees in their training data, confirming that verbatim repetition reduces flexible learning but mixture of diverse knowledge expressions lead to more robust category learning. Together, our findings highlight that while contextual flexibility is persistently difficult for models, how category knowledge is expressed in pretraining greatly impacts models' acquisition of robust category knowledge.
Nov 13, 2026

BLiMP-RO: A Benchmark of Linguistic Minimal Pairs for Romanian

Evaluating grammatical knowledge in language models requires benchmarks that cover diverse linguistic phenomena and are validated against human judgments. We introduce BLiMP-RO, a Romanian minimal-pair benchmark covering 50 grammatical phenomena across ten domains. It comprises 493 linguist-constructed sentence pairs, evaluated by native speakers in five experiments, and 5,942 additional semi-automatically generated, human-validated pairs. We further investigate grammatical generalization under limited linguistic input by training GPT-2 and GPT-BERT models from scratch on Romanian BabyBabelLM data and a newly collected corpus of 10 million words from child-targeted video captions. Increasing caption training data yields approximately 9 percentage points of accuracy improvement per tenfold increase in corpus size. At matched data sizes, BabyBabelLM generally outperforms video captions, while GPT-BERT consistently outperforms GPT-2. Our findings highlight how training data quantity, composition, and model architecture shape grammatical generalization in Romanian.
Nov 12, 2026

Evaluating Pretrained Language Semantics for Structured EHR Prediction Tasks

Structured Electronic Health Record (EHR) models typically represent clinical concepts as dataset-specific tokens, discarding their associated natural-language semantics. Consequently, relationships between concepts must largely be learned from EHR co-occurrence without directly exploiting pretrained linguistic knowledge. We investigate the contribution of pretrained language semantics to longitudinal EHR prediction through tEHR, a framework that converts clinical events into natural-language descriptions and adapts pretrained autoregressive language models through EHR-specific continued pretraining and task-specific discriminative fine-tuning. We evaluate tEHR across four MIMIC-IV prediction tasks, where Mistral-7B achieves the strongest overall performance against structured-data, code-based, and LLM-based baselines. Controlled ablations demonstrate the importance of lexical grounding, numerical values, and both adaptation stages. A capacity-controlled experiment further shows that replacing clinical descriptions with atomic code tokens substantially reduces performance despite retaining the same LLM backbone. Alternative approaches incorporating pretrained semantic embeddings offer lower computational costs but do not match full textualisation. These results highlight the importance of both semantic representation and downstream adaptation when applying pretrained language models to structured EHR prediction.
Nov 12, 2026

Lara Mini Base: Language Diversity as a Resource, Not a Trade-off

General-purpose language models are trained mostly on English: over open models we review, at most, non-English pretraining data makes up only a 12% share. Models built for European languages give them far more, but score well below general-purpose models of the same size. A common concern is that, in small models, languages compete for capacity. We present Lara Mini base, a model trained from scratch on 6T tokens in English and 44 other languages, which together cover all 24 official languages of the European Union. As training progresses, the 44 other languages together come to outweigh English. To give these languages enough capacity at low cost, Lara Mini base uses a mixture of experts architecture with 7.2B total and 0.6B active parameters. Among open base models with up to 2B active parameters, it achieves the highest average score across standard English, multilingual, math, and code benchmarks, ahead of models trained on 4 to 6 times as many tokens. Its margin is largest on multilingual benchmarks in the languages it supports and on math benchmarks. It supports contexts of up to 128k tokens. We find that our MoE architecture reduces capacity competition between languages. For a single user on one GPU, without quantization, it generates text at least 1.8 times faster than any model with a higher average score. We release the weights of Lara Mini base under the CC BY-NC-SA 4.0 license. In addition, we will release T-HellaSwag, a higher-quality machine translation of the commonsense reasoning benchmark HellaSwag into 110 language variants.
Nov 12, 2026

Pretraining Data Statistics Shape the Phases of Learning Entity Comparison in Language Models

How does data shape language model (LM) behavior throughout pretraining? We investigate this question through a case study on entity comparison, e.g., Between France and Brazil, which country is larger?, across nine comparison tasks spanning five entity types. We begin with controlled experiments in which we train small LMs (163M parameters) on mixtures of natural text from pretraining corpora and synthetic task data. Our analysis identifies three distinct phases of learning: (1) an early phase where the LM selects entities by frequency, (2) a middle phase where the LM selects entities by position in the prompt (first vs. last mentioned entity), and (3) a late phase where the LM selects the entity that is the correct answer. We show that the emergence of these three phases is causally determined by statistical properties of the training data. With small amounts of task-specific synthetic data, we observe only the first two phases and the model fails to learn the task; with large amounts, the model jumps directly from the frequency-based heuristic to solving the task correctly. Moreover, changing the entity frequency distribution from a naturally occurring Zipfian distribution (a small number of entities are very common and the vast majority are rare) to a uniform one weakens the frequency phase and speeds up learning of the task. Lastly, we show that the three phases of training dynamics from our controlled experiments appear in the open source model OLMo-2 32B. Together, our findings demonstrate that properties of pretraining data are causal drivers of heuristic learning and show that small-scale synthetic experiments can predict training dynamics at larger scales.
Nov 10, 2026

Beyond Passive Imitation: On-Policy Mid-Training Distillation for Language Models

Continual pre-training (CPT) adapts large language models (LLMs) to specialized domains, yet its standard next-token prediction (NTP) objective treats every corpus token as a fixed target, regardless of its utility to the current model. In a controlled study on the same 1B-token mathematical corpus, we observe sharply different outcomes across base models: NTP improves Llama-3.2-3B's performance on mathematical benchmarks but substantially degrades Qwen3-1.7B-Base's performance, indicating that, under a fixed training recipe, the same training sequences can have model-dependent effects. To address this problem, we introduce On-Policy Mid-Training Distillation (OPMD), which reformulates mid-training from offline text imitation into active, query-driven knowledge acquisition. At selected anchor positions, OPMD samples trajectories from the evolving student policy and applies dense token-level reverse-KL rewards from a frozen teacher, training the model on states it actually visits rather than only on ground-truth prefixes. This on-policy objective reduces the training-generation mismatch of teacher-forced NTP. On MATH-500 and Minerva, OPMD improves the Qwen student over its base performance and outperforms the evaluated NTP baseline, which degrades on both benchmarks. OPMD integrates into CPT pipelines without added inference-time overhead and offers a model-aware alternative to passive domain adaptation.
Nov 9, 2026

An Analytical Model of Compute-limited Multistage Training Pipelines

Modern foundation models are trained through multiple stages, yet how these stages interact and how compute should be allocated across them remain poorly understood. We introduce an analytically tractable high-dimensional model combining unsupervised pretraining with a linear autoencoder, supervised fine-tuning (SFT) with noisy stepwise labels, and reinforcement learning (RL) with sparse episodic rewards. Using random matrix theory and an operator-projection closure, we derive approximate low-dimensional learning dynamics that closely match simulations. We show that finite pretraining data imposes a representation-dependent performance limit, with Bayes classification error scaling as the inverse square root of the pretraining sample ratio, and observe the same exponent in asymptotic SFT and RL errors. Better pretraining also reduces sensitivity to noisy supervision. SFT initialization shortens the initial RL plateau induced by sparse rewards, an effect we also observe in OLMo-2 1B on GSM8K. Longer SFT warm starts additionally improve asymptotic RL performance. Optimizing compute across all three stages reveals that the optimal allocation depends on supervision quality and total budget, with a non-monotonic dependence of the RL compute share on budget. Finally, a complementary compositional model shows how RL selectively strengthens task-relevant pretrained skills while largely preserving unused skills. Together, these results identify mechanisms linking representation quality, feedback structure, and compute allocation in multistage training.
Nov 8, 2026

Perturbing Inputs Improves Pretraining Distillation

Knowledge distillation trains a student language model to match a teacher's predictions on the training text. We ask whether learning the teacher's behavior beyond the training text can further improve the student. We investigate this using three input perturbations: additive token noise, blending two documents, and blending neighboring tokens. Specifically, for an early part of pretraining we perturb the inputs and train the student to match the teacher on them, and we spend the rest of the same token budget on ordinary distillation. When pretraining from scratch, the best method improves on standard distillation at both scales we tried, by 0.025 nats at 135M parameters and by 0.020 nats at 0.5B (mean of three seeds). At 135M this is 1.45 times the gain that distillation itself provides over next-token training. We also observe promising downstream transfer: at 0.5B, the best variant improves average zero-shot accuracy by about one percentage point across seven tasks. It also retains its validation-loss advantage after low-bit weight quantization. We also test a data-limited setting in which training data must be reused, and find that the gains persist. Finally, ablations suggest that, for additive noise, perturbing the teacher's inputs contributes most of the gain, while perturbing only the student's inputs yields much less.
Nov 7, 2026

No Second Run: Does RSI Stop at Frontier Pre-training?

Pre-training is where every foundation model begins; bringing recursive self-improvement (RSI) there would let AI improve the process that creates AI. Yet public RSI work has so far targeted only post-training, which starts from an existing base model, runs trials costing GPU-hours, and returns verifiable scores. Pre-training breaks all three conditions at orders-of-magnitude higher cost: a run consumes hundreds of thousands of GPU-hours over months and cannot be repeated; its objective, post-training headroom, cannot be read off any single score; and what must improve is not a model but the whole production system, from data pipelines and training infrastructure to the teams that run them. To our knowledge, we are the first to apply RSI to an entire frontier pre-training run from scratch and to report it in detail. RSI does not stop at frontier pre-training, but it changes form: since no end-to-end loop can close over a single unrepeatable run, we decompose it into smaller loops that can, each repairing a step the full run breaks. The result is 801-30B-A3B-Base, a 30B-A3B MoE trained from scratch on 11.8T tokens in three months by a newly formed team. We release it as a base model; the lightweight SFT reported here (6B tokens, no RL) only probes post-training headroom. As a base model, 801-30B-A3B outperforms Qwen3-30B-A3B and Nemotron-3.5-Lightning-30B-A3B, trailing only the larger Qwen3.5-35B-A3B, and keeps this position under identical SFT; with lightweight SFT alone, it achieves the highest average on four agentic benchmarks against fully post-trained models, including Qwen3.6-35B-A3B. Since post-training RSI presupposes a strong base model, bringing RSI into pre-training acts on the foundation every later stage builds on. We open-source 801-30B-A3B-Base and the lightweight SFT data at https://github.com/GAIR-NLP/801-30B .
Nov 7, 2026

Pruning DNA Sequence

Data pruning can substantially reduce training cost by removing examples that contribute little additional learning signal, but it has not been studied for DNA language-model pretraining. We introduce Sequence Pruning (SP), a dynamic pruning framework that removes genomic sequences as they become easy for the model to predict. SP uses per-sequence pretraining accuracy computed directly from the model's forward pass, requiring no separate scoring model or additional pass over the corpus. We evaluate SP across five pretraining settings spanning two corpora, three model families, and both MLM and CLM objectives, with downstream evaluation on five genomic benchmarks. SP reduces pretraining data processing by up to 96% while largely preserving downstream performance. At moderate savings, pruning improves performance by up to 1.2 MCC points over full-data pretraining; at the highest savings, pretraining uses only 4% of the baseline GPU hours with little performance loss. SP also outperforms the dynamic pruning methods InfoBatch and BPAS across most comparisons. Beyond accuracy, our framework supports alternative criteria such as perplexity and EL2N, although accuracy yields the strongest downstream performance at matched savings. These results show that genomic pretraining does not require repeatedly processing all available sequences and that dynamic sequence-level pruning can substantially improve its data and computational efficiency. Code is available at https://github.com/FaiasPromit/pruning-dna-sequence.
Nov 6, 2026

Compact Greek Language Models from Scratch: Pretraining 500M- and 1B-Parameter Decoder-Only Transformers

We present two compact decoder-only language models trained from scratch exclusively on Greek text: LexisLM-1B (1.09B parameters) and LexisLM-500M (502M parameters). Both use a dedicated 32k byte-level BPE tokenizer and a 22B-token pretraining corpus drawn primarily from Greek Wikipedia and FineWeb2. Both models complete the 22B-token schedule, with best held-out perplexities of 9.77 for LexisLM-1B and 11.31 for LexisLM-500M. Evaluation of both models on Greek-translated ARC, MMLU, and TruthfulQA shows above-chance performance on ARC-Easy, while ARC-Challenge and MMLU remain near chance. We also perform supervised fine-tuning of both models on Greek instruction data, primarily EuroAlpaca, reaching held-out instruction perplexities of 7.12 for LexisLM-500M and 6.37 for LexisLM-1B. The report describes the tokenizer, data preparation, training, and evaluation methods, providing a foundation for further work on compact Greek language models.
Nov 5, 2026

Different Corpora, Different Skills? A Paired Study of Small Language Models

Small language models are often trained under tight data budgets, where the choice of pretraining corpus may matter as much as model size or training recipe. We compare two contrasting corpora at matched 10-million- and 37-million-word budgets: the 1911 Encyclopaedia Britannica and the BabyLM 2026 strict pretraining corpus. For each corpus, we train teacher models, distilled 58-million-parameter Baby-Llama students, and no-distillation student ablations, using the same model families, training recipes, and evaluation protocol across corpora. Across BLiMP, BLiMP Supplement, EWoK, Entity Tracking, and COMPS, the corpora produce different skill profiles rather than a single quality ranking. BabyLM 2026 is consistently stronger on linguistic acceptability benchmarks, while Britannica is often stronger on Entity Tracking, a synthetic state-tracking task. The contrast appears across teacher and student models and at both budgets, with the cleanest cross-seed replication in the distilled student comparison. Targeted controls show that the central contrast persists after matching training exposure and at a longer diagnostic context; direct teacher ensembling also does not reproduce the 10M Britannica student's Entity Tracking behavior. Entity Tracking magnitudes nevertheless remain sensitive to training dynamics. For this corpus pair, corpus choice can therefore produce different relative skill profiles at small scale, motivating multi-benchmark evaluation when comparing pretraining data.
Nov 4, 2026

A Night at the LLM Zoo: Data Conditioned Flow Matching for Generating Pretrained LLM Weights

A challenge to Weight Space Learning as it expands from task specific and vision models to Large Language Models (LLMs) is the absence of a well curated and analyzed LLM population. To address this, we train such a model zoo for application to Weight Space Generation by systematically varying the data mixture composition between web text, books, mathematics, coding, and French across populations of 51M, 124M, and 355M parameter GPT-2 Models, each tier with count N = 100. We analyze these populations by cosine similarity, functional diversity, Leave-One-Mixture-Out modeling, and variance decomposition. We show that the zoo contains diverse models whose variation is structured by training mixture composition. We embed the training mixture composition to condition a Flow Matching model of complete pretrained LLM weights in Gram-PCA compressed weight space. By demonstrating zero-shot generalization to held out mixtures, measuring Continuous Ranked Probability Score skill, as well as finding minimal changes in downstream test-set perplexity and general capability benchmark scores, we show that our generator generalizes on the simplex of data mixtures to produce models that perform similarly to those in the zoo ex nihilo. Finally, we produce all source code and artifacts to support the continued expansion of Weight Space Learning and Generation into the LLM domain.
Nov 4, 2026

Loop Before You Leap: What Recurrent Depth Can and Cannot Buy in Data-Limited Pretraining

Multimodal models have often scored below text-only baselines in the BabyLM Challenge (Klerings et al., 2024; Takmaz et al., 2025). We test whether additional recurrent computation helps close this gap. Recurrence increases computation by reusing layers without adding parameters or training data. Using a single implementation, we compare standard and looped Transformers under a 100M-word budget, with three seeds per main condition. Each image is represented by one feature vector from a frozen DINOv2 encoder, used as a prefix token. Caption words count toward the budget, so multimodal training also changes the language data. In a pre-registered text-only comparison, a 12-layer stack repeated four times exceeds a standard 48-layer Transformer by 5.4 BLiMP percentage points at matched inference compute. However, multimodal training lowers BLiMP accuracy by about 3.8 points in both architectures, with no clear mitigation from recurrence. An exploratory loop-side no-image control, trained with the multimodal language data and batching schedule, scores 5.7 points below the text-only baseline. Adding the visual prefix raises accuracy by 1.85 points on average, but this contrast is uncertain with three seeds. These results suggest that changes in language data and its presentation are a major source of the drop. Diagnostic tests show that the models use visual information; its contribution to grammatical generalization remains uncertain under this single-vector interface.
Nov 4, 2026

Adapting Qwen3 to Tajik from 0.6B to 8B: Lower Text Loss Did Not Mean Better Understanding

Small LoRA adapters trained on Tajik news and books lowered Qwen3's Tajik text loss at every size we tried, from 0.6B to 8B, by 0.60 to 1.45 nats per token, but none of them improved Tajik reading comprehension where it can be measured. On Belebele the 1.7B and 8B adapters left accuracy unchanged and the 4B adapter lowered it by 3.7 points; at 0.6B the score rose by 4.2 points but sits near chance and coincides with a change in which answer letter the model prefers, so we do not read it as understanding. At 8B, continued pretraining on about 375 million tokens of raw Tajik text was followed by a Belebele gain of 9.9 points. After a further supervised fine-tuning stage, our main model, e3, is 11.4 points above Qwen3-8B on Belebele; its English MMLU changed by -1.2 points, and a loss larger than 3.3 points is excluded. The same heavy recipe applied to Qwen3-0.6B-Base brought its text loss close to that of the 8B model after books-only pretraining, while its multiple-choice answers stayed near chance and its English MMLU fell by 7.6 points. We report 15 trained models and their 5 starting models from June to October 2025, scored in 2026 on one harness with paired intervals, and we report what went wrong: lost training records, a thinking-mode dataset with almost no reasoning in it, later branches whose training data was never documented, and a 4-bit release that contains no transformer layers.
Nov 3, 2026

ARLoop: Fully-Open Pre-trained Looped Transformers with Attention Residuals

Looped transformers increase effective depth by repeatedly applying a shared set of transformer layers, offering a parameter-efficient way to scale computation. However, standard residual connections aggregate information across recurrent depth into a single stream, providing limited control over which intermediate representations from preceding layers should be reused. In this paper, we introduce ARLoop, a looped transformer architecture based on attention residuals, where the input of each layer is selectively aggregated through attention over model depth. Distinct from standard transformers, we find that looped transformers do not require access to the full residual history: sliding-window depth attention matches or outperforms full-history attention at larger recurrent depths. Across controlled experiments with multiple parameter scales (140M and 370M) and recurrent depths (R = 1 to 5), ARLoop consistently outperforms prior residual designs for looped transformers. More importantly, we further scale ARLoop to 0.6B and 1.2B parameters under a realistic LLM pretraining recipe, demonstrating that the gains transfer to realistic pretraining, with particularly strong improvements on math and coding tasks. Our 1.2B model, pretrained on only 0.5T tokens with three recurrences, outperforms Huginn-3.5B (the only fully-open looped LLM as of our release date) across all evaluated tasks and is competitive with similar-sized non-looped models such as Qwen3.5-2B despite being trained on substantially less data.
Nov 3, 2026

Near-Zero Token Embeddings Silently Stall Training in Gated Sequence Models

Among 291 surveyed open-weight language model repositories (280 distinct embedding tables), 52.6% of untied gated and hybrid checkpoints (20/38) contain reachable input embeddings with norms below 10−610^{-6}, compared with 0% of weight-tied models. Such near-zero rows can arise during pretraining when unseen rare or template tokens receive zero input gradients while AdamW weight decay shrinks untied embeddings. In tested Transformer configurations, non-initial positions or a BOS token prevent this failure mode. In gated sequence models (Mamba, Gated DeltaNet, xLSTM), a near-zero token can amplify input-embedding gradients by more than 30 orders of magnitude (up to 35). At zero input, current-token readout factors (SiLU gates or linear queries) vanish, suppressing historical-state readout. Residual RMS values can then fall below the square root of the RMSNorm floor across consecutive layers, causing Pre-RMSNorm to amplify backward gradients by up to 10310^3 per layer. When a gradient entry exceeds 1.84×10191.84 \times 10^{19}, FP32 global-norm clipping in PyTorch, Hugging Face Trainer, and FSDP2 can overflow to infinity and zero the current batch's gradients. In Telugu continual pretraining of Nemotron-3-Nano-4B, this discards updates on approximately 20% of steps and leaves triggered rows frozen. We propose two remedies. For released checkpoints, GroupClip decouples input-embedding clipping using a 64-bit or rescaled norm; its rescaled implementation eliminates update loss and heals triggered rows in one step with negligible measured end-to-end overhead. For pretraining, GateFloor (SiLU(z)+δ\mathrm{SiLU}(z) + δ, δ∈[0.01,0.1]δ\in [0.01, 0.1]) keeps the gate positive at zero input, reducing near-zero gradients by 28 orders of magnitude in experiments up to 370M parameters, with perplexity comparable to the baseline.
Nov 2, 2026

DepthGraft: Structural Regularization through Hierarchical Cross-Layer KV Reconstruction

Scaling depth is a key route to stronger large language models, but Pre-LN architectures often suffer from depth-wise information dilution, limiting the effective use of deep layers. While cross-layer connectivity has proven effective at mitigating this problem, existing designs usually introduce additional computation or memory access. We revisit cross-layer KV sharing as an efficient form of such connectivity. Beyond cache compression, each KV reuse edge creates a direct gradient path from deep layers to source KV projections, turning KV sharing into an implicit structural regularizer for KV representation geometry. We provide a systematic analysis of this mechanism, showing that proper sharing topologies sharpen key retrieval directions while enriching the value content subspace, whereas existing adjacent or global-anchor designs do not fully exploit this effect. Motivated by this, we propose DepthGraft, a hierarchical KV-sharing architecture that reconstructs deep-layer KV from stride-aligned shallow-middle source pairs. DepthGraft distributes non-local gradient flow across source layers, while preserving prefill early completion and reducing KV cache by roughly one third. Experiments on dense and MoE models from 0.5B to 16B parameters show consistent downstream improvements under both GQA and MLA. Notably, on the 16B MoE model, DepthGraft achieves a 7.0% absolute improvement on MMLU-Pro, along with gains of 6.6% on CMMLU and 6.4% on C3. These results validate the potential of DepthGraft as a scalable design for large-scale pretraining. The project page is available at https://github.com/jianke0604/DepthGraft.
Nov 2, 2026

Language Models Can Benefit from Domain-Specific Inductive Biases

Language models are typically trained with a single choice of architecture on diverse domains and data sources. Modern transformers have been extensively refined for such general-purpose learning, but are they equally effective across different subsets of the pretraining data? We investigate this question by training small language models from scratch on representative subsets of pretraining data such as web text, code, and mathematics. We introduce variants of transformers in which standard activation functions are replaced with learned splines tuned on specific datasets, providing a simple way to modify their inductive biases. Using a mix-and-match of architectures and datasets, we evaluate cross-domain transfer, i.e. how well does an architecture tuned for one domain perform on data from another? We find that the tuned transformer variants improve training efficiency or generalization on all seven of our tested datasets, but the gains are often domain-specific. For example, architectures tuned on a natural language dataset such as FineWeb perform well on TinyStories, whereas those tuned on mathematics perform poorly on natural language. We identify several mechanisms underlying these improvements, including wider loss basins and more abundant "lottery tickets". Our results show that different subsets of pretraining data can benefit from different learning mechanisms, highlighting a limitation of the prevailing trend towards a small set of general-purpose architectures and optimizers. This points to an opportunity to improve data efficiency by developing domain-specific learning strategies that retain the ability to train general-purpose models.
Nov 2, 2026

Accelerating LLM Pretraining with Zeroth-Order Curvature on Idle Serving Compute

LLM pretraining is increasingly compute-intensive, while substantial inference serving compute remains idle during off-peak periods. However, such serving hardware is difficult to use for conventional first-order training due to its memory, networking and other system requirements. Zeroth-order (ZO) optimization offers a natural interface to this forward-only compute resource, but its gradient estimates break down in pretraining. Unlike fine-tuning, which can confine updates to a small parameter subspace, pretraining must update every parameter, and the alignment of a ZO gradient estimate with the true gradient shrinks as the number of updated parameters grows. We make a different observation: noisy ZO estimates can still provide useful curvature information, even when they are too inaccurate to serve as gradients. Based on this insight, we propose ZO-Curv, which leverages idle forward-only devices to estimate curvature and uses it to adapt and refine the update used during first-order training, thereby accelerating the process. We provide convergence guarantees that characterize the effect of finite-query curvature estimation. Experiments on LLMs from 125M to 1.5B parameters show that ZO-Curv consistently accelerates pretraining. At 770M parameters, ZO-Curv reaches AdamW's 20K-step validation loss in 3.9 hours instead of 7.5 hours, while achieving a lower final loss of 2.70 versus 2.84 under the same step budget. These results show that idle model serving compute can accelerate LLM pretraining when zeroth-order queries are used to estimate curvature rather than gradients directly.
Oct 31, 2026

Do Orthogonality Constraints Help Language-Model Training Beyond Sphere Optimization?

A growing body of work reports benefits from orthogonality constraints in fine-tuning large language models (LLMs). These results raise an open question: do orthogonality constraints also help LLM pre-training? We conduct a systematic empirical study of decoder-only LLM pre-training with models ranging from 360M to 3.1B parameters, trained on nearly 200B tokens. We examine fourteen variants of orthogonality training: some constrain the weight matrix itself, while others use orthogonal factors to reparameterize it. Because a strictly orthogonal square matrix loses half of its degrees of freedom, we also study block-wise orthogonality, where every dd rows form a row-orthonormal (Stiefel) matrix and the block size dd is a hyperparameter. Among block sizes ranging from 1 to the shorter side of the matrix, no d>1d>1 used throughout the network reliably outperforms d=1d=1, and large blocks are clearly worse; the one placement that helps, symmetric positive-definite factors on the query and key projections, loses its small gain as models grow. At d=1d=1, the constraint reduces to sphere optimization, with each row direction on its own unit sphere, which by itself already improves on tuned Muon. The same two findings hold for a diffusion transformer trained on ImageNet, and after fine-tuning, downstream accuracy does not separate the parameterizations. Our conclusion is that, across the realizations we tried, orthogonality between rows does not help language-model pre-training beyond sphere optimization. Since this appears to conflict with several prior results, we re-examine five prior methods in pre-training and fine-tuning that report orthogonality improvements. In each case the reported gain is attributable to numerical precision, training horizon, optimizer realization, or per-row norm control rather than to the orthogonality the methods impose.
Oct 30, 2026

Saidzoda-Tajik-LFM-2.6B: Adapting a 2.6B Hybrid Model to Tajik Tool Calling, and What Continued Pretraining Takes Away

We release Saidzoda-Tajik-LFM-2.6B, a Tajik adaptation of the 2.6B-parameter hybrid model LFM2.5-2.6B-Base, built for tool calling in Tajik on modest hardware. The recipe has two stages: LoRA continued pretraining (CPT) on 1,485,319 cleaned Tajik web documents for 3 epochs, then one epoch of supervised fine-tuning (SFT) on a mix of machine-translated Tajik tool calls, English tool calls kept as replay, and Tajik instructions. On a 174-item tool-calling test set built from BFCL v3 and machine-translated to Tajik, strict full-call accuracy on Tajik prompts rises from 20.69 for the base model to 64.14 (paired difference +43.45 points, 95% interval +35.2 to +51.7); a lenient parser that also accepts the base model's own call formats puts the base at 31.72. English shows no detectable change (76.55 and 77.24). Every CPT checkpoint scores zero: CPT removes the native tool-call format and, as the lenient re-scoring shows, part of the skill itself. Outside tool calling, accuracy on a Tajik knowledge benchmark rises from 28.0 to 35.0, Belebele Tajik shows no significant change, and English MMLU falls from 60.8 to 52.5. The model is not a general Tajik assistant. It almost never declines when no tool fits (6.9% correct on Tajik, against 31.0% for the base under the lenient parser), it is weak at multi-turn calls, and its calls should be checked before they are executed. We describe the data, the training settings and every evaluation in enough detail to repeat them.
Oct 29, 2026

FARO: Function-Space Adaptive Return-to-Risk Optimization via Cross-Block Equal-Risk-Contribution Budget Allocation

Training a large language model is compute-bound, and current recipes divide the work between optimizers by hand: orthogonalized momentum for attention and MLP matrices, Adam-style updates for the rest. This division rests on human priors about which layer type deserves which geometry, and it fixes a single global update budget that every parameter block shares. We introduce FARO, a function-space adaptive optimizer that makes the partition of parameters into blocks the unit of decision. FARO gives each block a geometry by its realized counterfactual return, then freezes the routes and reallocates the frozen per-block budget by cross-block equal risk contribution (ERC), the risk-parity principle from portfolio theory. On a GPT-style model trained on FineWeb, the framework gives two results. First, its FARO-Adaptive arm reaches 3.2780 nats at the full schedule, matching NorMuonAndAdam (3.2794), the best hand-designed recipe: it recovers the human prior without hand-setting per-layer geometry. Second, FARO-ERC improves on that arm: it lowers the clean-data loss at every one of 6 seeds, by 0.0136 nats on average, a modest but robust gain. Its sign test, paired test, and confidence interval all agree, and it exceeds the measured run-to-run noise at no added time or memory. It works by trimming the riskiest blocks rather than expanding the safest, and it holds up under a corrupted stream, still lowering the mean loss and winning the seed majority. The source code is available at https://github.com/wxj630/FARO.
Oct 29, 2026

Rethinking Pre-pretraining

We report that allocating compute to synthetic pre-pretraining (PPT) never outperforms pure natural-language training in our compute-matched experiments. This holds across four PPT curricula (k-Dyck, neural cellular automata (NCA), declarative reference corpora, and code) and three training regimes: data-constrained, data-unconstrained, and a high token-to-parameter-ratio regime (881M parameters trained on 50B tokens). When natural language is scarce, repeated epochs over a 100M-token corpus beat every synthetic curriculum on natural-language validation loss, with no consistent downstream accuracy advantage from PPT. We find that PPT gains reported in prior work are sensitive to the weight initialization scheme. In an NCA reproduction, we recover a 5.4% perplexity improvement from PPT under the original initialization scheme. An alternative scheme improves both methods, but the no-PPT baseline more, closing the gap between them. However, when PPT receives additional compute rather than a share of the natural-language budget, gains persist at small natural-language budgets even with improved initialization, but disappear by 1B tokens. Finally, models forget what they learn during PPT in subsequent natural-language training; interleaving the same data throughout training prevents forgetting, but does not significantly improve natural-language performance. Our results suggest limited benefits of synthetic pre-pretraining as a distinct training stage and motivate rethinking how such data are used in language model training.