AI Paper Atlas

Research map / Architectures and efficiency / Pruning and efficient tuning

Decentralized Instruction Tuning: Conflict-Aware Splitting and Weight Merging

MINSIK CHOI, Geewook Kim

ICML 2026 · Architectures and efficiency › Pruning and efficient tuning · approach: Low-rank compression and merging · 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
replace centralized joint training (1 ep) with PCA conflict-aware 3D split and one-shot token-weighted mergeMERIT · 8-benchmark average, Vision-FLANaverage score: 54.3 -> 57.0Improved
use conflict-aware PCA split instead of random partitioning (8 groups) (ablation)MERIT · 8-benchmark average, Vision-FLANaverage score: 54.5 -> 57.0Improved
increase number of PCA dimensions from 1D to 2D to 3D (ablation)MERIT · 8-benchmark average, Vision-FLANaverage score: 55.2 -> 55.7 -> 57.0Improved
increase number of random groups from 2 to 8 (ablation)random partition and merge · 8-benchmark average, Vision-FLANaverage score: 54.6 -> 54.4 -> 54.5No clear effect
even random partitioning plus merge versus joint 1 ep (ablation)split and merge · 8-benchmark average, Vision-FLANaverage score: 54.3 -> 54.6Improved
use PCA split instead of greedy conflict-induced split (2 groups) (ablation)MERIT 1D · 8-benchmark average, Vision-FLANaverage score: 54.9 -> 55.2Improved
use uniform soup of 2 models on full mixture instead of 1D MERIT (ablation)soup baseline · 8-benchmark average, Vision-FLANaverage score: 55.4 vs 55.2Mixed
train 7B from LLaVA-style base with MERIT 2D instead of joint FFTBase VLM 7B · 8-benchmark average, 1.6M mixtureaverage score: 54.9 -> 55.4Improved
MERIT 2D instead of joint FFT on stronger 3.6M baseScaled base VLM 7B · 8-benchmark average, 1.6M mixtureaverage score: 60.9 -> 61.5Improved
use MERIT 2D instead of joint FFT on stronger baseScaled base VLM 7B · LLaVA-Wscore: 50.2 -> 66.2Improved
MERIT 2D instead of joint FFT on stronger baseScaled base VLM 7B · MathVistascore: 49.7 -> 43.0Hurt
MERIT 2D instead of 1-epoch joint training on text-only FLANQwen2.5-3B · Text-only FLAN, 8-benchmark averageaverage score: 57.6 -> 58.4Improved

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