Research on scaling large language models (LLMs) has primarily focused on model parameters and training data size, overlooking the role of vocabulary size. We investigate how vocabulary size impacts LLM scaling laws by training models ranging from 33M to 3B parameters on up to 500B characters with various vocabulary configurations. We propose three complementary approaches for predicting the compute-optimal vocabulary size: IsoFLOPs analysis, derivative estimation, and parametric fit of the loss function. Our approaches converge on the conclusion that the optimal vocabulary size depends on the compute budget, with larger models requiring larger vocabularies.

In neural network-based models for natural language processing (NLP), the largest part of the parameters often consists of word embeddings. Conventional models prepare a large embedding matrix whose size depends on the vocabulary size. Therefore, storing these models in memory and disk storage is costly. In this study, to reduce the total number of parameters, the embeddings for all words are represented by transforming a shared embedding. The proposed method, ALONE (all word embeddings from one), constructs the embedding of a word by modifying the shared embedding with a filter vector, which is word-specific but non-trainable.

Large language models are trained with tokenizers, and the resulting token distribution is highly imbalanced: a few words dominate the stream while most occur rarely. Recent practice favors ever-larger vocabularies, but it is unclear where the benefit comes from. To this end, we perform a controlled study that scales the vocabulary of the language model from 24K to 196K while holding data, computation, and optimization unchanged. We begin by quantifying the complexity of tokenized text -- formalized via Kolmogorov complexity -- and show that larger vocabularies reduce this complexity. Above 24K, every common word is already tokenized as a single token, so enlarging vocabulary only deepens the relative token-frequency imbalance. Word-level loss decomposition shows that larger vocabularies reduce cross-entropy loss almost exclusively by lowering uncertainty on the 2,500 most frequent words, even though loss on the rare tail rises. The same frequent words cover roughly 75% of tokens in downstream benchmarks, so this training advantage transfers intact. We further show that enlarging model parameters with a fixed vocabulary yields the same frequent-word benefit. Our results recast "bigger vocabularies help" as "lowering complexity of tokenized text helps," offering a simple, principled knob for tokenizer-model co-design and clarifying the loss dynamics that govern language model scaling in pre-training.

We introduce a new tokenizer for language models that minimizes the average tokens per character, thereby reducing the number of tokens needed to represent text during training and to generate text during inference. Our method, which we refer to as the Length-MAX tokenizer, obtains its vocabulary by casting a length-weighted objective maximization as a graph partitioning problem and developing a greedy approximation algorithm. On FineWeb and diverse domains, it yields 14--18\% fewer tokens than Byte Pair Encoding (BPE) across vocabulary sizes from 10K to 50K, and the reduction is 13.0\% when the size is 64K. Training GPT-2 models at 124M, 355M, and 1.3B parameters from scratch with five runs each shows 18.5\%, 17.2\%, and 18.5\% fewer steps, respectively, to reach a fixed validation loss, and 13.7\%, 12.7\%, and 13.7\% lower inference latency, together with a 16\% throughput gain at 124M, while consistently improving on downstream tasks including reducing LAMBADA perplexity by 11.7\% and enhancing HellaSwag accuracy by 4.3\%. Moreover, the Length-MAX tokenizer achieves 99.62\% vocabulary coverage and the out-of-vocabulary rate remains low at 0.12\% on test sets. These results demonstrate that optimizing for average token length, rather than frequency alone, offers an effective approach to more efficient language modeling without sacrificing -- and often improving -- downstream performance. The tokenizer is compatible with production systems and reduces embedding and KV-cache memory by 18\% at inference.

Sequence models for binary analysis are bottlenecked by byte-level tokenization: raw bytes waste precious context window capacity for transformers and other neural network architectures, and many existing text-oriented tokenizers fail on arbitrary 0x00--0xFF sequences. To address this issue, we introduce the Binary BPE tokenizer family, a set of cross-platform Byte Pair Encoding (BPE) tokenizers for executables trained on a large corpus of binaries spanning multiple platforms, architectures, and operating systems, including Linux, Windows, macOS, Android, and malware sources. We release trained tokenizers with vocabularies of 4K, 8K, 16K, 32K, and 64K tokens, enabling both systematic scaling studies and practical deployment from resource-constrained edge devices to high-throughput datacenters. These tokenizers discover interpretable patterns (ELF/PE headers, instruction sequences, cross-platform strings) while yielding multi-byte compression per token. On representative uncompressed executables (e.g., ELF/PE/Mach-O rather than compressed APKs), the Binary BPE tokenizers typically allow for roughly 2-3x more binary content per fixed-length transformer context window than raw bytes, enabling more efficient research and practical deployment for content identification, malware detection, reverse engineering, and optimization. We release the trained Binary BPE tokenizers on HuggingFace, providing a drop-in, open-source foundation for binary-focused language models and context-efficient agentic tools.

Training Large Language Models (LLMs) typically involves a two-stage pipeline at the output layer: hidden states are projected into vocabulary logits via a linear transformation (lm_head), followed by cross-entropy loss computation against target tokens. While conceptually simple, this design incurs substantial overhead. The intermediate logits tensor, with dimensions proportional to batch size, sequence length, and vocabulary size, must be fully materialized in GPU memory, even though only one target token per position is ultimately used. This leads to significant memory footprint and bandwidth comsumption, limiting scalability and slowing training throughput. In this work, we introduce a novel approach to integrates the output projection and loss prediction into a single operation. By directly computing the loss from hidden states and target tokens, our approach bypasses explicit logits materialization. This design reduces memory usage and alleviates bandwidth pressure. Experiments on LLM training demonstrate that our method achieves substantial memory savings and measurable speedups compared to the standard two-stage pipeline, enabling large batch sizes and longer sequences without sacrificing accuracy. Our work highlights the benefits of rethinking the boundary between projection and prediction, offering a practical systems optimization for efficient LLM training.

We propose a fast approximation method of a softmax function with a very large vocabulary using singular value decomposition (SVD).

Neural Information Processing Systems http://nips.cc/

Research on scaling large language models (LLMs) has primarily focused on model parameters and training data size, overlooking the role of vocabulary size.