Doc-to-Atom: New Framework Decomposes Documents into Semantic Knowledge Units for Efficient LLM Processing
Researchers introduced Doc-to-Atom, a new method that breaks documents into semantically typed knowledge units (atoms) to improve how large language models handle long input sequences. The approach addresses limitations of prior work like Doc-to-LoRA by creating independent micro-adapters for each atom rather than a single monolithic adapter, reducing interference and improving compositional reasoning. This matters because it could make document understanding and multi-step reasoning in LLMs faster and more memory-efficient while maintaining or improving accuracy.
Doc-to-Atom proposes a compositional parametric memory framework designed to overcome the quadratic computational cost of attention mechanisms when processing long document sequences in large language models. Rather than compressing an entire document into a single adapter, the method decomposes documents into semantically typed knowledge atoms, each compiled into an independent micro-LoRA adapter paired with a provenance retrieval key. During inference, a lightweight query router selects and assembles only the relevant atoms into a query-specific adapter that is injected into a frozen base model. The entire system is trained end-to-end using a multi-objective distillation framework. Experiments across six diverse question-answering benchmarks show that Doc2Atom outperforms the Doc-to-LoRA baseline while reducing memory costs associated with document internalization, addressing key challenges of irrelevant-query interference and limited compositional recall.
What's missing
The paper does not discuss potential limitations of the approach, such as how performance scales with extremely long documents, computational overhead of the query router, or failure cases where semantic decomposition may be ambiguous or difficult.
What different sources said
- arXiv cs.CLCenter
Doc-to-Atom: Learning to Compile and Compose Memory Atoms
Related
Genetic Drift, Not Selection, Drives Rapid Feather Color Evolution in Island Bird Radiation
A new study of an island bird radiation found that rapid evolution of feather coloration is driven primarily by genetic drift in small populations rather than sexual or ecological selection. The research integrated whole-genome data with detailed plumage measurements across complete species sampling to test whether signaling trait evolution correlates with speciation rates. The findings suggest that neutral demographic processes play a central role in generating phenotypic diversity during island radiations, challenging assumptions about the mechanisms driving rapid evolution.
New AI Model Improves Prediction of Therapeutic Peptide Function from Protein Sequences
Researchers developed a lightweight CNN classifier that predicts whether peptide sequences have therapeutic properties, trained on a database of 54,655 peptides across 48 functional categories. The model uses a novel negative sampling strategy to reduce false positive rates from over 60% in previous approaches to 2.1%. This advancement could accelerate drug discovery by enabling faster computational screening of peptide candidates before expensive experimental testing.
Study Shows Different Metabolic Stress Models Produce Distinct Effects on Human Neuronal Networks
Researchers tested three common in vitro metabolic stress models on human-derived neuronal networks and found each produced different patterns of neuronal activity and cell damage. The models tested were hypoxia alone, oxygen-glucose deprivation (OGD), and hypoxia combined with glutamate exposure. The findings suggest that choice of experimental model significantly affects results and that combining electrophysiological and structural analyses is important for accurately assessing metabolic stress in stroke research.