AI Paper Atlas

Research map / Continual learning / Continual fine-tuning of LLMs

Orthogonal Model Merging

Sihan Yang, Kexuan Shi, Weiyang Liu

ICML 2026 · Continual learning › Continual fine-tuning of LLMs · approach: Synthetic data and merging for CLIP · 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
add magnitude correction to Lie-algebra averaging (ablation)OrthoMerge · OFT experts, Llama-3.1-8B in-domainaverage accuracy: 41.01 -> 46.25Improved
average in Lie algebra instead of averaging orthogonal matrices directly (ablation)OrthoMerge · OFT experts, Llama-3.1-8B in-domainaverage accuracy: 39.53 -> 41.01Improved
merge rotations by sequential product instead of Lie-algebra averaging (ablation)OrthoMerge · OFT experts, Llama-3.1-8B MATH500MATH500: 24.60 -> 8.60Hurt
merge OFT experts with OrthoMerge instead of TIES in weight spaceOrthoMerge · OFT experts, Llama-3.1-8B in-domainaverage accuracy: 44.61 -> 46.25Improved
merge with OrthoMerge instead of TSV-M on OFT expertsOrthoMerge · OFT experts, M-ARC and AGIEval avgaverage accuracy: 40.48 -> 41.80Improved
add OrthoMerge-G decoupling to Task ArithmeticTA · MergeBench Llama-3.2-3B task avgaverage score: 34.16 -> 35.48 (+1.32)Improved
add OrthoMerge-C decoupling to TIESTIES · MergeBench Llama-3.2-3B task avgaverage score: 39.16 -> 41.57 (+2.41)Improved
add OrthoMerge-G decoupling to TSV-MTSV-M · MergeBench Llama-3.2-3B task avgaverage score: 41.52 -> 40.03 (-1.49)Hurt
add OrthoMerge-C decoupling to TSV-MTSV-M · MergeBench Llama-3.2-3B task avgaverage score: 41.52 -> 42.07 (+0.55)Improved
add OrthoMerge-C decoupling to TIESTIES · MergeBench MMLUMMLU (excl. math/coding): 54.52 -> 55.36Improved
add OrthoMerge-C decoupling to TIESTIES · Qwen2.5-VL-7B task avgaverage accuracy: 63.54 -> 64.04 (+0.50)Improved
add OrthoMerge-G decoupling to TSV-MTSV-M · Qwen2.5-VL-7B IFEvalIFEval: 54.90 -> 57.86Improved

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