Back to AI Research

AI Research

ARIADNE: Agnostic Routing for Inference-time Adapte... | AI Research

Key Takeaways

  • ARIADNE: Agnostic Routing for Inference-time Adapter DyNamic sElection As language models become more specialized, developers often use "adapters"—small, mod...
  • The increasing deployment of parameter-efficient fine-tuning (PEFT) has led to model ecosystems in which a single backbone is paired with many task-specialized adapters.
  • In this setting, inference-time queries often arrive without task labels, requiring the system to automatically select the most appropriate adapter from a growing and heterogeneous adapter pool.
  • We introduce ARIADNE, a training-free, adapter-agnostic routing framework for dynamic adapter selection at inference time.
  • ARIADNE represents each adapter through a set of centroids computed from embeddings of its training set, capturing the data distribution associated with that adapter.
Paper AbstractExpand

The increasing deployment of parameter-efficient fine-tuning (PEFT) has led to model ecosystems in which a single backbone is paired with many task-specialized adapters. In this setting, inference-time queries often arrive without task labels, requiring the system to automatically select the most appropriate adapter from a growing and heterogeneous adapter pool. Existing routing methods either depend on access to adapter internals, such as weight decompositions or gradient-based statistics, or require additional router training, which limits scalability and portability as new adapters are added. We introduce ARIADNE, a training-free, adapter-agnostic routing framework for dynamic adapter selection at inference time. ARIADNE represents each adapter through a set of centroids computed from embeddings of its training set, capturing the data distribution associated with that adapter. Given an unlabeled input, it selects an adapter by measuring proximity to these centroids in latent space. Because routing is performed entirely in the input embedding space, ARIADNE is compatible with arbitrary PEFT methods and requires no modification to the adapters or training procedures. Primarily evaluated with Llama 3.2 1B Instruct on 23 diverse NLP tasks, ARIADNE recovers 97.44% of the upper bound performance. Scaling to 44 tasks, it achieves 89.7% average selection accuracy, without additional training or access to adapter internals.

ARIADNE: Agnostic Routing for Inference-time Adapter DyNamic sElection
As language models become more specialized, developers often use "adapters"—small, modular add-ons that allow a single base model to perform many different tasks. A major challenge in these ecosystems is "routing": when a user sends a query, the system must automatically select the correct adapter without knowing the task in advance. Existing solutions often require training extra components or digging into the internal weight structures of the adapters, which makes them difficult to scale or apply to different types of models. ARIADNE is a new, training-free framework that solves this by treating adapter selection as a simple input classification problem, allowing it to work with any adapter type without needing access to its internal design.

How ARIADNE Works

Instead of analyzing the complex internal math of an adapter, ARIADNE focuses on the input data itself. The researchers observed that inputs belonging to the same task naturally cluster together in the latent space of a standard text encoder. To leverage this, the system creates "centroids"—representative points—for each task by averaging the embeddings of its training data. When a new, unlabeled query arrives, ARIADNE converts it into an embedding and compares it to these pre-computed centroids. It then selects the adapter associated with the most similar centroid. Because this process happens entirely in the input space, it is completely independent of how the adapters were built or which base model is being used.

Performance and Scalability

The researchers tested ARIADNE using Llama 3.2 1B Instruct across 23 diverse natural language processing tasks. The results show that the framework is highly effective, recovering 97.44% of the performance that would be achieved by an "oracle" (a system that always knows the correct adapter). When compared to existing "spectral" routing methods—which rely on analyzing adapter weights—ARIADNE consistently achieved higher selection accuracy. Furthermore, the system remains stable as the library of adapters grows; in a test involving 44 different tasks, it maintained an average selection accuracy of 89.7%, proving that it can handle increasingly complex model ecosystems.

Graceful Failure

A key finding of the research is how the system behaves when it makes a mistake. Because ARIADNE routes based on the semantic meaning of the input, its errors are not random. When the system selects the wrong adapter, it typically chooses one that is semantically related to the correct task. This leads to "graceful degradation," where the model still provides useful results rather than failing completely. This behavior makes the system easier to diagnose and more reliable in practical, real-world applications.

Important Considerations

While ARIADNE is highly flexible and adapter-agnostic, it does have one primary limitation: it requires access to the training data associated with each adapter to calculate the initial centroids. This means it may not be suitable for decentralized environments where training data is kept private or is unavailable. However, the researchers suggest that future work could focus on extracting descriptive signals directly from the adapters themselves, potentially allowing the system to function without needing any training samples at all.

Comments (0)

No comments yet

Be the first to share your thoughts!