[Paper] Doc-to-Atom: Learning to Compile and Compose Memory Atoms
Source: arXiv - 2606.12400v1
Overview
Long input sequences are central to document understanding and multi-step reasoning in Large Language Models, yet the quadratic cost of attention makes inference both memory-intensive and slow. Context distillation mitigates this by compressing contextual information into model parameters, and recent work such as Doc-to-LoRA amortizes context distillation into a single forward pass that generates one LoRA adapter per document. However, producing a single monolithic adapter for all queries leads to irrelevant-query interference, limited compositional recall, and poor scalability to long-document reasoning. To address these challenges, we propose Doc-to-Atom (Doc2Atom), a compositional parametric memory framework that decomposes each document into semantically typed knowledge atoms. Each atom is compiled into an independent micro-LoRA adapter and a provenance retrieval key. At inference time, a lightweight query router selects and assembles only the relevant atoms into a query-specific adapter, which is then injected into a frozen base model. The entire system is trained end-to-end through a multi-objective distillation framework. Experiments on six diverse QA benchmarks demonstrate that Doc2Atom outperforms Doc-to-LoRA baselines while reducing the memory cost of document internalization.
Key Contributions
This paper presents research in the following areas:
- cs.CL
- cs.IR
Methodology
Please refer to the full paper for detailed methodology.
Practical Implications
This research contributes to the advancement of cs.CL.
Authors
- Xingjian Diao
- Wenbo Li
- Yashas Malur Saidutta
- Avinash Amballa
- Lazar Valkov
- Srinivas Chappidi
Paper Information
- arXiv ID: 2606.12400v1
- Categories: cs.CL, cs.IR
- Published: June 10, 2026
- PDF: Download PDF