# LG AI Research — [CVPR 2026] An upcycling method for creating true experts within a Mixture-of-Experts structure

- Company: LG AI Research (lgresearch.ai)
- Announced: 2026-06-26T00:00:00+00:00
- Category: research-paper
- Subject: EXAONE
- Source: https://www.lgresearch.ai/blog/view?seq=659
- Record: https://forck.live/items/4745-cvpr-2026-an-upcycling-method-for-creating-true-experts-within-a-mixture-of

LG AI Research proposes a cluster-aware upcycling method that initializes experts and routers based on activation clustering of a pre-trained dense model, aiming to improve expert specialization in Mixture-of-Experts structures.

## Evidence

Verbatim from https://www.lgresearch.ai/blog/view?seq=659:

> LG AI Research analyzed how the final training results and performance can vary depending on the initialization method used when upcycling pre-trained dense models into MoE models, and proposed an initialization technique for more effective training.

---

Record: https://forck.live/items/4745-cvpr-2026-an-upcycling-method-for-creating-true-experts-within-a-mixture-of
Catalogue: https://forck.live/llms.txt
Feed: https://forck.live/feed.md