Research

Paper

AI LLM March 11, 2026

Word Recovery in Large Language Models Enables Character-Level Tokenization Robustness

Authors

Zhipeng Yang, Shu Yang, Lijie Hu, Di Wang

Abstract

Large language models (LLMs) trained with canonical tokenization exhibit surprising robustness to non-canonical inputs such as character-level tokenization, yet the mechanisms underlying this robustness remain unclear. We study this phenomenon through mechanistic interpretability and identify a core process we term word recovery. We first introduce a decoding-based method to detect word recovery, showing that hidden states reconstruct canonical word-level token identities from character-level inputs. We then provide causal evidence by removing the corresponding subspace from hidden states, which consistently degrades downstream task performance. Finally, we conduct a fine-grained attention analysis and show that in-group attention among characters belonging to the same canonical token is critical for word recovery: masking such attention in early layers substantially reduces both recovery scores and task performance. Together, our findings provide a mechanistic explanation for tokenization robustness and identify word recovery as a key mechanism enabling LLMs to process character-level inputs.

Metadata

arXiv ID: 2603.10771
Provider: ARXIV
Primary Category: cs.CL
Published: 2026-03-11
Fetched: 2026-03-12 04:21

Related papers

Raw Data (Debug)
{
  "raw_xml": "<entry>\n    <id>http://arxiv.org/abs/2603.10771v1</id>\n    <title>Word Recovery in Large Language Models Enables Character-Level Tokenization Robustness</title>\n    <updated>2026-03-11T13:49:14Z</updated>\n    <link href='https://arxiv.org/abs/2603.10771v1' rel='alternate' type='text/html'/>\n    <link href='https://arxiv.org/pdf/2603.10771v1' rel='related' title='pdf' type='application/pdf'/>\n    <summary>Large language models (LLMs) trained with canonical tokenization exhibit surprising robustness to non-canonical inputs such as character-level tokenization, yet the mechanisms underlying this robustness remain unclear. We study this phenomenon through mechanistic interpretability and identify a core process we term word recovery. We first introduce a decoding-based method to detect word recovery, showing that hidden states reconstruct canonical word-level token identities from character-level inputs. We then provide causal evidence by removing the corresponding subspace from hidden states, which consistently degrades downstream task performance. Finally, we conduct a fine-grained attention analysis and show that in-group attention among characters belonging to the same canonical token is critical for word recovery: masking such attention in early layers substantially reduces both recovery scores and task performance. Together, our findings provide a mechanistic explanation for tokenization robustness and identify word recovery as a key mechanism enabling LLMs to process character-level inputs.</summary>\n    <category scheme='http://arxiv.org/schemas/atom' term='cs.CL'/>\n    <published>2026-03-11T13:49:14Z</published>\n    <arxiv:primary_category term='cs.CL'/>\n    <author>\n      <name>Zhipeng Yang</name>\n    </author>\n    <author>\n      <name>Shu Yang</name>\n    </author>\n    <author>\n      <name>Lijie Hu</name>\n    </author>\n    <author>\n      <name>Di Wang</name>\n    </author>\n  </entry>"
}