AI Paper Atlas

Research map / Large language models / Instruction and fine-tuning

MergeMix: Optimizing Mid-Training Data Mixtures via Learnable Model Merging

jiapeng wang, Changxin Tian, Kunlong Chen, Ziqi Liu, Jiaxin Mao, Xin Zhao, Zhiqiang Zhang, JUN ZHOU

ICML 2026 · Large language models › Instruction and fine-tuning · approach: Efficient LLM fine-tuning systems · 0 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
use MergeMix-derived mixture instead of manually tuned ratios8B MoE mid-training · Mid-training average of 5 domainsaverage score: 57.9 -> 58.4Improved
use MergeMix mixture instead of uniform sampling8B MoE mid-training · Mid-training averageaverage score: 54.2 -> 58.4Improved
use MergeMix instead of Adapted RegMix8B MoE mid-training · Mid-training averageaverage score: 58.1 -> 58.4Improved
replace manual mixture search by merging-based searchmixture tuning · search costcost in tokens: 2T tokens -> 20B tokens (100x)Improved
weight interpolation of math/code experts as proxy for data mixingmath and code experts · math/code ratio sweepSpearman rho: 0.82 (math), 0.78 (code)Improved
transfer the mixture found on 8B proxy to 16B model16B MoE mid-training · cross-scale transferaverage scoreImproved
re-optimize mixture at 50% of training (dynamic) vs static (ablation)MergeMix · mid-training averageaverage scoreNo clear effect
MergeMix fine-grained weights instead of uniform within domains (ablation)hierarchical MergeMix · 16 sub-clustersaverage score: 57.6 -> 58.2Improved
MergeMix fine-grained weights instead of manual fine-grained (ablation)hierarchical MergeMix · 16 sub-clustersaverage score: 58.4 -> 58.2No clear effect
fully automated MergeMix hierarchy instead of fully manualmid-training mixture · 16 sub-clustersaverage score: 57.9 -> 58.2Improved
MergeMix mixture vs uni-domain and aggressive specialization mixturesmath and code targeting · math and code benchmarksaccuracy: math 60.14, 60.96 vs MergeMix 61.48Improved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy