Recent language models have shown impressive multilingual performance, even when not explicitly trained for it. Despite this, there are concerns about the quality of their outputs across different languages. In this paper, we show how disparity in the treatment of different languages arises at the tokenization stage, well before a model is even invoked. The same text translated into different languages can have drastically different tok-enization lengths, with differences up to 15 times in some cases. These disparities persist even for tokenizers that are intentionally trained for multilingual support.

Tokenization is widely used in large language models because it significantly improves performance. However, tokenization imposes several disadvantages, such as performance biases, increased adversarial vulnerability, decreased character-level modeling performance, and increased modeling complexity. To address these disadvantages without sacrificing performance, we propose SpaceByte, a novel byte-level decoder architecture that closes the performance gap between byte-level and subword autoregressive language modeling. SpaceByte consists of a byte-level Transformer model, but with extra larger transformer blocks inserted in the middle of the layers. We find that performance is significantly improved by applying these larger blocks only after certain bytes, such as space characters, which typically denote word boundaries. Our experiments show that for a fixed training and inference compute budget, SpaceByte outperforms other byte-level architectures and roughly matches the performance of tokenized Transformer architectures.

As a new paradigm in information retrieval, generative retrieval directly generates a ranked list of document identifiers (docids) for a given query using generative language models (LMs).How to assign each document a unique docid (denoted as document tokenization) is a critical problem, because it determines whether the generative retrieval model can precisely retrieve any document by simply decoding its docid.Most existing methods adopt rule-based tokenization, which is ad-hoc and does not generalize well.In contrast, in this paper we propose a novel document tokenization learning method, GenRet, which learns to encode the complete document semantics into docids.GenRet learns to tokenize documents into short discrete representations (i.e., docids) via a discrete auto-encoding approach.We develop a progressive training scheme to capture the autoregressive nature of docids and diverse clustering techniques to stabilize the training process.Based on the semantic-embedded docids of any set of documents, the generative retrieval model can learn to generate the most relevant docid only according to the docids' semantic relevance to the queries.We conduct experiments on the NQ320K, MS MARCO, and BEIR datasets.GenRet establishes the new state-of-the-art on the NQ320K dataset.Compared to generative retrieval baselines, GenRet can achieve significant improvements on unseen documents.Moreover, GenRet can also outperform comparable baselines on MS MARCO and BEIR, demonstrating the method's generalizability.

While there has been a large body of research attempting to circumvent tokenization for language modeling (Clark et al. 2022, Xue et al. 2022), the current consensus is that it is a necessary initial step for designing state-of-the-art performant language models. In this paper, we investigate tokenization from a theoretical point of view by studying the behavior of transformers on simple data generating processes. When trained on data drawn from certain simple $k^{\text{th}}$-order Markov processes for $k > 1$, transformers exhibit a surprising phenomenon - in the absence of tokenization, they empirically are incredibly slow or fail to learn the right distribution and predict characters according to a unigram model (Makkuva et al. 2024). With the addition of tokenization, however, we empirically observe that transformers break through this barrier and are able to model the probabilities of sequences drawn from the source near-optimally, achieving small cross-entropy loss. With this observation as starting point, we study the end-to-end cross-entropy loss achieved by transformers with and without tokenization. With the appropriate tokenization, we show that even the simplest unigram models (over tokens) learnt by transformers are able to model the probability of sequences drawn from $k^{\text{th}}$-order Markov sources near optimally. Our analysis provides a justification for the use of tokenization in practice through studying the behavior of transformers on Markovian data.

In multilingual settings, non-Latin scripts and low-resource languages are usually disadvantaged in terms of language models' utility, efficiency, and cost. Specifically, previous studies have reported multiple modeling biases that the current tokenization algorithms introduce to non-Latin script languages, the main one being over-segmentation. In this work, we propose MAGNET-- multilingual adaptive gradient-based tokenization--to reduce over-segmentation via adaptive gradient-based subword tokenization. MAGNET learns to predict segment boundaries between byte tokens in a sequence via sub-modules within the model, which act as internal boundary predictors (tokenizers). Previous gradient-based tokenization methods aimed for uniform compression across sequences by integrating a single boundary predictor during training and optimizing it end-to-end through stochastic reparameterization alongside the next token prediction objective. However, this approach still results in over-segmentation for non-Latin script languages in multilingual settings. In contrast, MAGNET offers a customizable architecture where byte-level sequences are routed through language-script-specific predictors, each optimized for its respective language script. This modularity enforces equitable segmentation granularity across different language scripts compared to previous methods. Through extensive experiments, we demonstrate that in addition to reducing segmentation disparities, MAGNET also enables faster language modeling and improves downstream utility.

This paper investigates the impact of using morphologically-informed tokenizers to aid and streamline the interlinear gloss annotation of an audio corpus of Yoloxóchitl Mixtec (YM) using a combination of ASR and text-based sequence-to-sequence tools, with the goal of improving efficiency while reducing the workload of a human annotator. We present two novel tokenization schemes that separate words in a nonlinear manner, preserving information about tonal morphology as much as possible. One of these approaches, a Segment and Melody tokenizer, simply extracts the tones without predicting segmentation. The other, a Sequence of Processes tokenizer, predicts segmentation for the words, which could allow an end-to-end ASR system to produce segmented and unsegmented transcriptions in a single pass. We find that these novel tokenizers are competitive with BPE and Unigram models, and the Segment-and-Melody model outperforms traditional tokenizers in terms of word error rate but does not reach the same character error rate. In addition, we analyze tokenizers on morphological and information-theoretic metrics to find predictive correlations with downstream performance. Our results suggest that nonlinear tokenizers designed specifically for the non-concatenative morphology of a language are competitive with conventional BPE and Unigram models for ASR. Further research will be necessary to determine the applicability of these tokenizers in downstream processing tasks.

Automatic speech recognition for low-resource languages remains fundamentally constrained by the scarcity of labeled data and computational resources required by state-of-the-art models. We present a systematic investigation into cross-lingual continuous pretraining for low-resource languages, using Perso-Arabic languages (Persian, Arabic, and Urdu) as our primary case study. Our approach demonstrates that strategic utilization of unlabeled speech data can effectively bridge the resource gap without sacrificing recognition accuracy. We construct a 3,000-hour multilingual corpus through a scalable unlabeled data collection pipeline and employ targeted continual pretraining combined with morphologically-aware tokenization to develop a 300M parameter model that achieves performance comparable to systems 5 times larger. Our model outperforms Whisper Large v3 (1.5B parameters) on Persian and achieves competitive results on Arabic and Urdu despite using significantly fewer parameters and substantially less labeled data. These findings challenge the prevailing assumption that ASR quality scales primarily with model size, revealing instead that data relevance and strategic pretraining are more critical factors for low-resource scenarios. This work provides a practical pathway toward inclusive speech technology, enabling effective ASR for underrepresented languages without dependence on massive computational infrastructure or proprietary datasets.

Tokenization strategies shape how models process electronic health records, yet fair comparisons of their effectiveness remain limited. We present a systematic evaluation of tokenization approaches for clinical time series modeling using transformer-based architectures, revealing task-dependent and sometimes counterintuitive findings about temporal and value feature importance. Through controlled ablations across four clinical prediction tasks on MIMIC-IV, we demonstrate that explicit time encodings provide no consistent statistically significant benefit for the evaluated downstream tasks. Value features show task-dependent importance, affecting mortality prediction but not readmission, suggesting code sequences alone can carry sufficient predictive signal. We further show that frozen pretrained code encoders dramatically outperform their trainable counterparts while requiring dramatically fewer parameters. Larger clinical encoders provide consistent improvements across tasks, benefiting from frozen embeddings that eliminate computational overhead. Our controlled evaluation enables fairer tokenization comparisons and demonstrates that simpler, parameter-efficient approaches can, in many cases, achieve strong performance, though the optimal tokenization strategy remains task-dependent.

Foundation models have revolutionized various fields such as natural language processing (NLP) and computer vision (CV). While efforts have been made to transfer the success of the foundation models in general AI domains to biology, existing works focus on directly adopting the existing foundation model architectures from general machine learning domains without a systematic design considering the unique physicochemical and structural properties of each biological data modality. This leads to suboptimal performance, as these repurposed architectures struggle to capture the long-range dependencies, sparse information, and complex underlying ``grammars'' inherent to biological data. To address this gap, we introduce BioArc, a novel framework designed to move beyond intuition-driven architecture design towards principled, automated architecture discovery for biological foundation models. Leveraging Neural Architecture Search (NAS), BioArc systematically explores a vast architecture design space, evaluating architectures across multiple biological modalities while rigorously analyzing the interplay between architecture, tokenization, and training strategies. This large-scale analysis identifies novel, high-performance architectures, allowing us to distill a set of empirical design principles to guide future model development. Furthermore, to make the best of this set of discovered principled architectures, we propose and compare several architecture prediction methods that effectively and efficiently predict optimal architectures for new biological tasks. Overall, our work provides a foundational resource and a principled methodology to guide the creation of the next generation of task-specific and foundation models for biology.

Deep learning research for binary analysis faces a critical infrastructure gap. Today, existing datasets target single platforms, require specialized tooling, or provide only hand-engineered features incompatible with modern neural architectures; no single dataset supports accessible research and pedagogy on realistic use cases. To solve this, we introduce Binary-30K, the first heterogeneous binary dataset designed for sequence-based models like transformers. Critically, Binary-30K covers Windows, Linux, macOS, and Android across 15+ CPU architectures. With 29,793 binaries and approximately 26.93% malware representation, Binary-30K enables research on platform-invariant detection, cross-target transfer learning, and long-context binary understanding. The dataset provides pre-computed byte-level BPE tokenization alongside comprehensive structural metadata, supporting both sequence modeling and structure-aware approaches. Platform-first stratified sampling ensures representative coverage across operating systems and architectures, while distribution via Hugging Face with official train/validation/test splits enables reproducible benchmarking. The dataset is publicly available at https://huggingface.co/datasets/mjbommar/binary-30k, providing an accessible resource for researchers, practitioners, and students alike.