AI & ML New Capability

Enables the prediction of an adapter's task, performance, and attributes directly from its LoRA weights without any inference or data access.

arXiv · March 18, 2026 · 2603.15990

Xiaolong Han, Ferrante Neri, Zijian Jiang, Fang Wu, Yanfang Ye, Lu Yin, Zehong Wang

The Takeaway

By resolving factorization ambiguity via a canonical form, the authors show that weight-space embeddings are highly informative. This allows for massive-scale adapter retrieval and routing in multi-tenant LLM systems without the cost of running the base model.

From the abstract

Each LoRA checkpoint compactly stores task-specific updates in low-rank weight matrices, offering an efficient way to adapt large language models to new tasks and domains. In principle, these weights already encode what the adapter does and how well it performs. In this paper, we ask whether this information can be read directly from the weights, without running the base model or accessing training data. A key obstacle is that a single LoRA update can be factorized in infinitely many ways. Witho