Low-rank adaptation
Algorithms AIMech:LoRA DRAFT
Aliases: LoRA
A parameter-efficient adaptation method that freezes pretrained weights and learns additive low-rank matrix updates.
Identity Notes & Concept Scope
A method for adapting model parameters, not a pretrained checkpoint or the LoRA software package. AIO:TransferLearning is a broader task description and does not specify this weight-update factorization.
Claims & Grounded Evidence
2 Claims
Mechanism
#mechanism
An adapted linear transformation uses W0x + BAx, with W0 fixed and the lower-rank factors A and B trainable.
Limitation
#merged-adapter-batching
Merging a task-specific BA update into the base weights complicates batching inputs requiring different adapters in one forward pass.
⚠️ Curated Limitations & Negative Results
- The inspected paper focuses its experiments on attention-weight adaptation; no universal task-performance claim is made.
- Scope is original LoRA; quantized, adaptive-rank and other later variants need separate identity assessment.
Artificial Intelligence Ontology (AIO) Grounding
Mapped AIO Classes
| AIO Target Class | Mapping Relation | Mapping Rationale |
|---|---|---|
AIO:TransferLearning
Transfer Learning
|
related_match | LoRA implements an adaptation procedure; the AIO target denotes the general transfer-learning task rather than this factorization algorithm. |
Evidence Sources & Primary Literature
1 Source| Source ID | Publication / Document | Type / Access | Reference & Checksum |
|---|---|---|---|
original-paper
|
Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen
|
primary paper
full_text
|
ID: arXiv:2106.09685v2
sha256:e9a0d3128767…
|
Curation Provenance
Curator: Codex
Method: ai_assisted
Model: GPT-6
Created: 2026-10-11T03:02:19Z
Updated: 2026-10-11T03:02:19Z
Source File:
data/curated/LoRA.yaml