IR3DE: A Lightweight Linear Routing Scheme for Domain Expert Large Language Models

This article introduces IR3DE, a lightweight router based on ridge regression designed to select the most suitable domain expert large language model for each prompt. Its core advantages include low-cost and high-efficiency inference, and support for dynamic addition and removal of expert models without retraining. This scheme was proposed by the Gensyn team, and the paper was published on arXiv (link: http://arxiv.org/abs/2606.06098v1, published on 2026-06-04).

The Fragmentation Dilemma of Large Language Models

With the development of large language model technology, the number of general-purpose models and domain expert models has surged, requiring users to balance performance, cost, and latency. Traditional single models are not optimal for handling all tasks—for example, code models perform mediocrely in legal analysis, while medical models struggle with mathematical reasoning.

Limitations of Existing Routing Schemes

• Weak-to-strong cost optimization category: Assumes a weak-to-strong model spectrum and only optimizes cost, but cannot handle domain expert models (ability distribution differences are not simply weak or strong).
• Domain expert routing category: Requires large amounts of data and computing resources to train the router; adding/removing expert models requires retraining, leading to heavy operation and maintenance burdens.

Ridge Regression: A Simple and Efficient Choice

IR3DE adopts the ridge regression algorithm with L2 regularization, with the following advantages:

• Extremely low computational overhead: Inference only requires one matrix multiplication and addition
• Fast training speed: Closed-form solution without iteration
• Strong generalization ability: Regularization prevents overfitting
• Good interpretability: Weights reflect feature importance

Dynamic Expert Management: Plug-and-Play

When adding an expert model, only need to calculate its performance on a small amount of validation data and update the regression coefficients without retraining; when removing an expert model, only need to delete the corresponding coefficient column, enabling dynamic adjustment of the expert pool.

The research team evaluated IR3DE in three scenarios:

1. General Domain CLM: Expert models trained on data from different domains, IR3DE's performance is comparable to complex baselines.
2. Hybrid Domain CLM: Expert models fine-tuned for different downstream tasks, IR3DE still maintains robustness comparable to baselines.
3. Reasoning Tasks: Expert models handling different reasoning types (mathematics, logic, common sense), IR3DE outperforms baselines, achieving 98.4% normalized performance.

• Model service providers: Low-cost and efficient multi-model scheduling, significantly reducing deployment and operation costs.
• Enterprise users: Flexible adjustment of expert model pools; adding domain models does not require service interruption or retraining.
• Researchers: Prove that simple methods (such as ridge regression) are more effective than complex models in specific tasks, prompting attention to the essential structure of problems.

Limitations

• The linear model assumes an approximate linear relationship between input features and targets; performance is affected when domain boundaries are blurred or non-linear.
• Relies on the quality of prompt encoding; if the encoder cannot capture key features, routing accuracy decreases.

Future Directions

• Explore more efficient feature encoding methods
• Combine active learning to optimize routing decisions
• Extend to multi-modal model routing scenarios