The paper investigates how to effectively pre‑train language models when the data budget is limited but compute is plentiful. It shows that increasing model size only improves performance up to an optimal point, after which overfitting degrades generalization, and that this optimal size varies with both the data budget and downstream tasks. To overcome the inefficiencies of standard Transformers in this regime, the authors propose recursive Transformers that reuse a shared block across depth and employ factorized embeddings, achieving better results than standard models on 10M–100M word pre‑training budgets and competitive performance with BabyLM Challenge 2025 winners.
Scaling Large Language Models (LLMs) has been driven mainly by enlarging the Transformer backbone, but for an already-strong model this requires another round of costly pretraining. We study whether an existing backbone can keep improving by allocating more computation to each token while leaving the Transformer backbone fixed.
The paper proposes an information-weighted cross‑entropy loss that rescales token contributions using TF‑IDF statistics, thereby emphasizing semantically informative tokens and down‑weighting ubiquitous ones. Experiments on five decoder‑only language models (1.1B–13B parameters) show consistent reductions in memorized substring length while maintaining perplexity and downstream performance. The method is architecture‑agnostic, adds less than 3% computational overhead, and can be integrated into existing training pipelines.
arXiv:2606. 13647v1 Announce Type: cross Abstract: We introduce SkMTEB, the first comprehensive MTEB-style text embedding benchmark for Slovak, a low-resource West Slavic language, comprising 31 datasets across 7 task types -- nearly 4$\times$ the depth of existing multilingual benchmark coverage for Slovak.
The paper proposes a modular tokenizer framework for multilingual large language models, allowing the creation of language‑specific subtokenizers that match monolingual compression quality. It introduces a pretraining strategy that samples these subtokenizers to limit predictions to relevant vocabularies, enabling efficient training and inference. This approach reduces memory usage and speeds up inference without compromising performance.
Large language models are typically trained under uniform token weighting, which allows frequent and low-information tokens to dominate learning and can increase the tendency to memorize surface-level...
arXiv:2601. 21461v3 Announce Type: replace-cross Abstract: Modern sparse language models typically achieve sparsity through Mixture-of-Experts (MoE) layers, which dynamically route tokens to dense MLP "experts.
Multilingual Large Language Models (LLMs) traditionally rely on a single vocabulary shared by all supported languages, which can lead to uneven compression across them. Moreover, their large embedding...
The paper introduces TokenAdapt, a model‑agnostic tokenizer transplantation method that uses a hybrid heuristic to initialize new token embeddings, and a novel pre‑tokenization learning approach for multi‑word Supertokens to improve compression. TokenAdapt combines local subword decomposition and global semantic similarity to preserve semantics while reducing retraining needs. Empirical results show that TokenAdapt outperforms existing baselines such as Transtokenizer and ReTok, achieving lower perplexity ratios and significant compression gains.
arXiv:2607.00171v2 Announce Type: replace Abstract: Text embeddings are standard for semantic similarity tasks, yet their evaluation remains an open challenge. Current benchmarks are static, cover on...
The study compares an English-only and a bilingual decoder-only model, each 310 M parameters, trained on eight diverse languages while controlling for English exposure, compute, and document overlap. After aligning on shared English vocabulary, the authors find that token embeddings appear similar, but the deeper hidden states used for prediction diverge across models. This hidden‑state mismatch grows through middle transformer layers and persists despite controls, indicating that contextual processing differs between the models. "whyItMatters":"The findings show that embedding alignment can conceal significant internal representation differences, which is crucial for any downstream work that assumes aligned multilingual models are interchangeable."
The paper examines how new vocabulary tokens are added to language models for generative recommendation tasks. It shows that the common practice of initializing these tokens as the mean of existing embeddings collapses them into a degenerate subspace, hindering fine‑tuning. The authors propose Grounded Token Initialization (GTI), which places new tokens at semantically meaningful positions in the pretrained embedding space using linguistic supervision, and demonstrate that GTI outperforms mean initialization and other adaptation methods across several benchmarks.
The study evaluates synthetic pre‑pretraining (PPT) across a wide range of models (500 M–7 B parameters) and training budgets (up to 100 B tokens). Results show that PPT consistently improves downstream performance and token efficiency, saving at least 21 B tokens at the 3 B scale, but these gains do not appear to stem from a grammatical prior. Instead, PPT benefits arise from tasks that enhance long‑range retrieval, and the improvements remain robust across diverse data mixtures, diminishing only when web text is omitted.