Doc-to-Atom: Learning to Compile and Compose Memory Atoms

Xingjian Diao, Wenbo Li, Yashas Malur Saidutta, Avinash Amballa, Lazar Valkov et al.|June 10, 2026arXiv

Key Takeaway

Breaking documents into composable micro-adapters instead of monolithic ones improves both efficiency and accuracy for long-document reasoning in LLMs.

Summary

This paper introduces Doc-to-Atom, a system that breaks down documents into small, typed knowledge units called 'atoms' and converts each into a tiny adapter for language models. Instead of creating one large adapter per document, the system selectively combines only relevant atoms based on the query, reducing memory usage and improving performance on question-answering tasks.

efficiency training

Key Terms

lora context-distillation parameter-efficient-fine-tuning query-router knowledge-atoms