IR3DE: A Linear Router for Large Language Models

Hugging Face Daily Papers Papers

Summary

IR3DE is a ridge regression-based router that selects domain-expert LLMs for different tasks, achieving competitive performance while enabling dynamic addition or removal of experts without retraining.

Foundational Large Language Models (LLMs) demonstrate proficiency on a wide range of general tasks, and achieve remarkable results on various specialized tasks via domain-expert LLMs. With the ever-growing list of available LLMs, inference routers are being proposed to select the most appropriate LLM for each prompt. However, existing routing methods either optimize cost across weak-to-strong generalist LLMs or require substantial training to support domain-expertise routing. In this paper, we propose IR3DE, a Ridge Regression-based Router for Domain Experts that provides cheap and fast routing decisions for each prompt. We evaluate IR3DE in two Causal Language Modeling (CLM) settings where the tasks are next-token prediction for all domains, and one reasoning setting where each domain has its own distinct reasoning task. Despite being a linear router, IR3DE achieves performance comparable to the other baselines in both CLM settings, and surpassing them in the reasoning setting, with a normalized performance of 98.4%. Moreover, IR3DE enables the addition or removal of new domain experts without requiring the router to be retrained from scratch, allowing a dynamic set of LLMs to be served with minimal disruption to the router itself. Our code is available at: github.com/gensyn-ai/IR3DE.
Original Article
View Cached Full Text

Cached at: 06/10/26, 09:43 AM

Paper page - IR3DE: A Linear Router for Large Language Models

Source: https://huggingface.co/papers/2606.06098

Abstract

A ridge regression-based routing method achieves competitive performance in selecting domain-expert LLMs for different tasks while enabling dynamic addition/removal of experts without retraining.

Foundational Large Language Models (LLMs) demonstrate proficiency on a wide range of general tasks, and achieve remarkable results on various specialized tasks viadomain-expert LLMs. With the ever-growing list of available LLMs,inference routersare being proposed to select the most appropriate LLM for each prompt. However, existing routing methods either optimize cost across weak-to-strong generalist LLMs or require substantial training to support domain-expertise routing. In this paper, we propose IR3DE, aRidge Regression-based Router for Domain Experts that provides cheap and fast routing decisions for each prompt. We evaluate IR3DE in twoCausal Language Modeling(CLM) settings where the tasks arenext-token predictionfor all domains, and one reasoning setting where each domain has its own distinct reasoning task. Despite being alinear router, IR3DE achieves performance comparable to the other baselines in both CLM settings, and surpassing them in the reasoning setting, with a normalized performance of 98.4%. Moreover, IR3DE enables the addition or removal of new domain experts without requiring the router to be retrained from scratch, allowing a dynamic set of LLMs to be served with minimal disruption to the router itself. Our code is available at: github.com/gensyn-ai/IR3DE.

View arXiv pageView PDFProject pageGitHub2Add to collection

Get this paper in your agent:

hf papers read 2606\.06098

Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash

Models citing this paper0

No model linking this paper

Cite arxiv.org/abs/2606.06098 in a model README.md to link it from this page.

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2606.06098 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2606.06098 in a Space README.md to link it from this page.

Collections including this paper1

Similar Articles

Dynamic Latent Routing

Hugging Face Daily Papers

Dynamic Latent Routing (DLR) lets LLMs learn their own inner monologue by composing sub-policies via search, inspired by language compositionality. In low-data fine-tuning, DLR matches or outperforms standard supervised fine-tuning.