Is LLaMA-MoE superseded?

LLaMA-MoE (Mixture-of-experts routing): superseded — cited as a baseline and beaten by newer methods. 2 paper(s) critique it, 1 beat it on benchmarks — #27 of 1370 most-superseded. Sub-problem: cluster led by LLaMA-MoE. Newer alternatives in the same sub-problem include DOT-MoE, ZEDA, ExpertWeaver, Mixture-of-Experts (MoE) Adaptation, Elastic Mixture-of-Experts (EMoE).

Mixture-of-experts routing

LLaMA-MoE

LLaMA-MoE: Building Mixture-of-Experts from LLaMA with Continual Pre-training

Superseded baseline#27 of 1,370 most-superseded · first seen Jun 24, 2024

Superseded — cited as a baseline and beaten by newer methods

2 papers critique it · 1 beat it on benchmarks

What papers say

Verbatim critique sentences, each from a paper that cites LLaMA-MoE as a baseline.

require extensive continual training (200B and 7B tokens, respectively)

— CMoE: Converting Mixture-of-Experts from Dense to Accelerate LLM Inference

“Unlike existing methods cmoe, llamamoe, llamamoev2 which freeze the assignment and then train the router, DOT-MoE allows the router and the expert assignment to co-adapt.”

— DOT-MoE: Differentiable Optimal Transport for MoEfication

Beaten on benchmarks

Head-to-head results where a newer method reports beating LLaMA-MoE. Values are copied from the source paper's tables — verify against the cited paper.

CMoE beats LLaMA-MoE
49.86 vs 32.40
ARC-C · [Llama-2 70B]
CMoE: Converting Mixture-of-Experts from Dense to Accelerate LLM Inference

What to use instead

Recent methods in the same sub-problem, not yet superseded in the knowledge base — arXiv benchmark leaders, not vetted production recommendations.