CURLY CITATION: A Token-Efficient Citation Representation for Large Language Model Contexts

Abstract

Citations provide supporting evidence for claims in large language model (LLM) contexts, but existing citation formats can introduce input-token overhead in long documents containing hundreds of references or more. We propose Curly Citation, a compact citation representation that preserves the information required to resolve a cited source while reducing repetitive bibliographic and URL information. It represents each reference in a fixed format using curly braces, an index, a predefined parameter, an endpoint and a pair of backticks. Preliminary benchmarks across multiple tokenisers and associated models show that popular conventional citation formats like IEEE, Harvard, APA and 20 other formats require 31–59 input-tokens to represent a selected citation. Curly Citation, however, reduces the input-token cost representation by approximately 50.00%–68.18% compared to selected existing citation formats. In a documentscale benchmark containing 150 citations, replacing IEEE citations with Curly Citation reduced total document input-token consumption by 11.33%–12.37% across the evaluated tokenisers. These results indicate that Curly Citation may provide a practical approach for lowering citation-related token overhead in LLM document processing.

Keywords

• Large Language Models (LLMs)
• Token Efficiency
• Context Efficiency
• Citation Compression
• Citation Representation
• Token-Efficient Citation Format