JevSoup: System-One Routing for Training-Free LoRA Composition

Xiuying Wang, Jiahua Cheng, Shuotian Li, Yufan Cheng, Bowen Deng +2 more
9/25/2026
cs.AI
lora-peft-merging

Official code: Leowang980/JevSoup (1 star, last commit 2026-09-25).

Abstract

Building adaptable AI systems requires effective coordination of specialized capabilities across diverse tasks. Low-rank adaptation (LoRA) enables modular expertise, but existing routing approaches may require auxiliary data, additional training, or autoregressive decoding. We propose JevSoup, a training-free framework separating System One expert routing from System Two execution. Using only the input and expert descriptions, Jev selects two experts through structured probabilities. JevSoup retains the leading expert's update, projects the second onto the orthogonal complement of the first update's row space, and combines them with equal weights. Across 14 PorTAL tasks and three Qwen3 scales, JepSoup achieves absolute gains of up to 1.19\% in task-macro and 1.21\% in sample-micro accuracy over the strongest evaluated external baselines. Our code is available at https://github.com/Leowang980/JevSoup.

View on arXivView PDF

Code Implementations(1)

Leowang980/JevSoupOfficial100%
10Shell, PythonSep 24, 20264 weeks ago

Similar Papers

Learning to Select, Not Relearn: Hard-Routed Mixtures of Reasoning LoRAs

85% match2026

Share First, Route What Remains: A Unified Framework for Token-Adaptive MoE Computation

84% match2026

MoE-Sieve: Routing-Guided LoRA for Efficient MoE Fine-Tuning

83% match2026

GaP-MoE : Explicit Expert Specialization in Mixture-of-Experts via Orthogonal Gradient Accumulation and Projection

83% match2026

Expert-Space Exploration in MoE Reinforcement Learning

82% match2026

Adaptive Shared Experts with LoRA-Based Mixture of Experts for Multi-Task Learning

82% match2025

Cite this paper

@article{wang2026jevsoup,
  title  = {JevSoup: System-One Routing for Training-Free LoRA Composition},
  author = {Xiuying Wang and Jiahua Cheng and Shuotian Li and Yufan Cheng and Bowen Deng and Zhexuan Bai and Yichen Li},
  year   = {2026},
  eprint = {2609.30922},
  archivePrefix = {arXiv},
  url    = {http://arxiv.org/abs/2609.30922v1}
}

Add a badge to your README

Shows whether official code is linked for this paper.

Code of Paper badge
[![Code of Paper](https://codeofpaper.com/badge/2609.30922.svg)](https://codeofpaper.com/paper/2609.30922)

Discussion