ARIADNE adapter routing for PEFT, recovers 97.44% on Llama 3.2 1B

ARIADNE is a training-free, adapter-agnostic routing framework for selecting task adapters at inference time, submitted to arXiv on 17 Jun 2026 by Enrico Cassano, Michał Brzozowski, Zuzanna Dubanowska, Paolo Mandica and Neo Christopher Chung. The method represents each adapter using centroids computed from embeddings of that adapter's training set and picks an adapter for an unlabeled input by measuring proximity to those centroids in latent space.

How does ARIADNE select an adapter?

ARIADNE chooses adapters by comparing an input's embedding to precomputed centroids for each adapter, selecting the adapter whose centroids are closest in latent space. The paper describes computing centroids from embeddings of an adapter's training set, performing routing entirely in the input embedding space so no adapter internals or additional router training are required. Because routing uses only input embeddings, ARIADNE is compatible with arbitrary parameter-efficient fine-tuning methods and needs no modifications to adapters or to their training procedures.

How well does ARIADNE perform in evaluations?

On the Llama 3.2 1B Instruct backbone across 23 diverse NLP tasks, ARIADNE recovers 97.44% of the upper bound performance reported in the paper. When the authors scale the experiment to 44 tasks, ARIADNE achieves 89.7% average selection accuracy, again without additional training or access to adapter internals. Those two figures, 97.44% recovery on 23 tasks and 89.7% average selection accuracy on 44 tasks, are the specific performance points provided by the submission.

Why it matters

ARIADNE removes two common bottlenecks in routing: reliance on adapter internals and the need to train a separate router. That makes it portable across adapter formats and scalable as adapter pools grow, because new adapters only require their training-set embeddings to be processed into centroids. For deployments where queries arrive unlabeled and many task-specialized adapters coexist on a single backbone, a training-free, agnostic router reduces integration friction while retaining most of the upper-bound performance in the authors' experiments.

What to watch

Measure whether the centroid-only approach maintains similar recovery or selection-accuracy figures when applied to larger backbones or to adapter pools with different task distributions. The paper provides results for Llama 3.2 1B and for 23 and 44 task pools; the next confirmation would be comparable selection accuracy on other backbones or larger, more heterogeneous adapter collections.