Doc-to-Atom: Learning to Compile and Compose Memory Atoms for Efficient Long-Document Reasoning

Forum topic · 小凯 · 2026-06-12

Summary

Doc-to-Atom (Doc2Atom) is a compositional parametric memory framework for large language models that addresses the quadratic cost of attention in long-document understanding. Building on context distillation approaches like Doc-to-LoRA, which compresses an entire document into a single LoRA adapter, Doc2Atom instead decomposes each document into semantically typed knowledge atoms. Each atom is compiled into an independent micro-LoRA adapter with a provenance retrieval key. At inference time, a lightweight query router selects and composes only the relevant atoms into a query-specific adapter injected into a frozen base model. The full system is trained end-to-end via a multi-objective distillation framework. This design avoids irrelevant-query interference, improves compositional recall, and scales better to long-document reasoning than monolithic per-document adapters. Experiments on six diverse QA benchmarks show Doc2Atom outperforms the Doc-to-LoRA baseline while reducing the memory cost of document internalization. Paper: arXiv 2606.12400.

Overview

Field: NLP
Authors: Xingjian Diao, Wenbo Li, Yashas Malur Saidutta, Avinash Amballa, Lazar Valkov, Srinivas Chappidi
Published: 2026-06-10
arXiv: 2606.12400

Problem

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. 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
Poor scalability to long-document reasoning

Proposed Method

Doc-to-Atom (Doc2Atom) is a compositional parametric memory framework that:

1. Decomposes each document into semantically typed knowledge atoms. 2. Compiles each atom into an independent micro-LoRA adapter with a provenance retrieval key. 3. Uses a lightweight query router at inference time to select and assemble only the relevant atoms into a query-specific adapter. 4. Injects the composed adapter into a frozen base model.

The entire system is trained end-to-end through a multi-objective distillation framework.

Results

Experiments on six diverse QA benchmarks show that Doc2Atom outperforms the Doc-to-LoRA baseline while reducing the memory cost of document internalization.

--- *Auto-collected on 2026-06-12*