Breaking the Token Ceiling: Distilling Smaller, Stronger Byte Models

• URL: http://arxiv.org/abs/2609.12303v1
• Date: Fri, 11 Sep 2026 00:17:22 GMT
• Title: Breaking the Token Ceiling: Distilling Smaller, Stronger Byte Models
• Abstract summary: We present the first large scale study of overtraining decoder-only dense transformer models varying two dimensions simultaneously.<n>We find that Token-1B models outperform byte models (End-Of-Token-1B and Bytes-1B) in the low-FLOP regime but eventually plateau.<n>By operating over a small vocabulary of 256 bytes instead of on the order of 100K tokens, they circumvent the need for top-kruncation during logit dumping.
• Score: 67.72077328700088
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Small models are made more capable through distillation from a larger one that shares their tokenization scheme. However, do distilled byte and token models behave similarly in terms of scaling trends as compute and data increases? To enable this comparison, we introduce two variants to efficiently convert token logits to Byte Logits: 1) approximate: Marginalize-It, and 2) exact: End-Of-Token. We then present the first large scale study of overtraining decoder-only dense transformer models varying two dimensions simultaneously: the tokenization scheme (Tokens, Bytes, Bytes w/ eot) and the training objective (Distillation vs. Cross-Entropy), sweeping layer-parameter-matched models with roughly 1 billion parameters up to 1 trillion bytes of data. Across eight benchmarks spanning three categories: Multiple Choice QA, Language Generation, and Machine Translation, we find that Token-1B models outperform byte models (End-Of-Token-1B and Bytes-1B) in the low-FLOP regime but eventually plateau; byte models start worse yet surpass Token-1B models with more compute, reaching a higher downstream task performance ceiling. Extrapolating the average top-1 error vs. validation BPB scaling laws predicts that, asymptotically, distilled End-Of-Token-1B outperforms distilled Token-1B by up to 4%. They are also far more data efficient, matching the performance of distilled Token-1B using only one-sixth of the training data. Moreover, by operating over a small vocabulary of 256 bytes instead of on the order of 100K tokens, they circumvent the need for top-k truncation during logit dumping, while also reducing logit storage costs to roughly one-fifth. Finally, our downstream performance scaling laws predict that our distilled End-Of-Token-1B models asymptotically surpass the Llama 3.2-1B, Gemma-3-1B-pt, and Gemma 2B models on averaged downstream tasks by up to 6.5%, 8.1%, and 2.1%, respectively.

Related papers

• Compute Optimal Tokenization [79.3815358070537]
  We investigate how the information granularity of tokens, controlled by the compression rate, affects scaling trends.<n>We train 988 latent tokenized models (BLT) ranging from 50M to 7B parameters that enable setting the desired compression rate.<n>Our experiments reveal that in compute-optimal configurations, model parameter counts scale proportionally to data size measured in bytes, not in tokens as commonly perceived.
  arXiv  Detail & Related papers  (2026-05-02T01:53:22Z)
• OverFill: Two-Stage Models for Efficient Language Model Decoding [68.68408155020568]
  Large language models (LLMs) excel across diverse tasks but face significant deployment challenges due to high inference costs.<n>We propose OverFill, which decouples prefill and decode stages to optimize accuracy-efficiency tradeoffs.<n>Our 3B-to-1B OverFill configuration outperforms 1B pruned models by 83.2%, while the 8B-to-3B configuration improves over 3B pruned models by 79.2% on average.
  arXiv  Detail & Related papers  (2025-08-11T20:07:34Z)
• DataDecide: How to Predict Best Pretraining Data with Small Experiments [67.95896457895404]
  We release models, data, and evaluations in DataDecide -- the most extensive open suite of models over differences in data and scale.<n>We conduct controlled pretraining experiments across 25 corpora with differing sources, deduplication, and filtering up to 100B tokens, model sizes up to 1B parameters, and 3 random seeds.
  arXiv  Detail & Related papers  (2025-04-15T17:02:15Z)
• Byte Latent Transformer: Patches Scale Better Than Tokens [101.10994909832063]
  Byte Latent Transformer (BLT) encodes bytes into dynamically sized patches, which serve as the primary units of computation.<n>For fixed inference costs, BLT shows significantly better scaling than tokenization-based models, by simultaneously growing both patch and model size.
  arXiv  Detail & Related papers  (2024-12-13T05:33:32Z)
• Focus on the Core: Efficient Attention via Pruned Token Compression for Document Classification [6.660834045805309]
  Pre-trained transformers such as BERT suffer from a computationally expensive self-attention mechanism. We propose integrating two strategies: token pruning and token combining. Experiments with various datasets demonstrate superior performance compared to baseline models.
  arXiv  Detail & Related papers  (2024-06-03T12:51:52Z)
• Language models scale reliably with over-training and on downstream tasks [121.69867718185125]
  Scaling laws are useful guides for derisking expensive training runs. However, there remain gaps between current studies and how language models are trained. In contrast, scaling laws mostly predict loss on inference, but models are usually compared on downstream task performance.
  arXiv  Detail & Related papers  (2024-03-13T13:54:00Z)
• Token Dropping for Efficient BERT Pretraining [33.63507016806947]
  We develop a simple but effective "token dropping" method to accelerate the pretraining of transformer models. We leverage the already built-in masked language modeling (MLM) loss to identify unimportant tokens with practically no computational overhead. This simple approach reduces the pretraining cost of BERT by 25% while achieving similar overall fine-tuning performance on standard downstream tasks.
  arXiv  Detail & Related papers  (2022-03-24T17:50:46Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.