AI Paper Atlas

Research map / Architectures and efficiency / Pruning and efficient tuning

Navigating the Accuracy-Size Trade-Off with Flexible Model Merging

Akash Dhasade, Divyansh Jhunjhunwala, Milos Vujasinovic, Gauri Joshi, Anne-Marie Kermarrec

ICLR 2026 · Architectures and efficiency › Pruning and efficient tuning · approach: Low-rank compression and merging

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
double deployed size from 1x to 2x with FlexMergeTA on 8 vision tasks (ViT-B/32) · 8 vision tasksaverage accuracy: 67.5 -> >80Improved
grow merged size from ~3x to ~6xConsensus on 30 vision tasks (ViT-B/32) · 30 vision tasksaverage accuracy: 76 -> 84.5Improved
grow merged size from 1x to 3xTA on 11 PEFT models (T0-3B) · 11 PEFT tasksaverage accuracy: 59 -> 66.2Improved
grow merged size from 1x to 2xTA on 7 FFT models (T5-Large) · 7 FFT tasksaverage accuracy: +9% to 75Improved
grow merged size from 2.2x to 4xEMR-Merging on 7 FFT models (T5-Large) · 7 FFT tasksaverage accuracy: 85.2 -> >86Improved
block-wise greedy pairwise merging instead of Channel Merging K-MeansTA (ViT-B/32) · 8 vision tasks, integer sizesaverage accuracyImproved
use min similarity instead of max/average between groups (ablation)FlexMerge · 8 vision tasks (ViT-B/32)average accuracyImproved
greedy merging order instead of left-to-right or right-to-left (ablation)FlexMerge · 8 vision tasks (ViT-B/32)average accuracyImproved
cosine-similarity block selection instead of random selection (ablation)FlexMerge · 8 vision tasks (ViT-B/32)average accuracyImproved
increase size beyond 1x for ranking of algorithmsvanilla averaging vs TIES · vision tasksaverage accuracyMixed
larger merged size on inference latencyFlexMerge vs standard merging · ViT-B/32 inferenceinference time: 12.30 ms -> 12.21 msNo clear effect
larger merged size on inference latencyFlexMerge vs standard merging · ViT-L/14 inferenceinference time: 118.70 ms -> 120.53 msNo clear effect

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