AI Paper Atlas

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

Low Kruskal-Rank Adaptation

Yixing Xu, Guanchen Li, Chao Li, Xuanwu Yin, Dong Li, Spandan Tiwari, Ashish Sirasao, Emad Barsoum

ICML 2026 · Large language models › Instruction and fine-tuning · approach: Low-rank adaptation variants

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 K-rank log-det penalty to LoRALoRA r=32 on LLaMA-7B · Commonsense reasoning (avg)average accuracy: 76.3 -> 78.7Improved
add K-rank log-det penalty to LoRALoRA r=32 on LLaMA-13B · Commonsense reasoning (avg)average accuracy: 79.9 -> 82.3Improved
add K-rank log-det penalty to LoRALoRA r=32 on LLaMA2-7B · Commonsense reasoning (avg)average accuracy: 77.6 -> 81.1Improved
add K-rank log-det penalty to LoRALoRA r=32 on LLaMA3-8B · Commonsense reasoning (avg)average accuracy: 80.9 -> 85.9Improved
add K-rank log-det penalty to LoRALoRA r=32 on Qwen3-8B · Commonsense reasoning (avg)average accuracy: 85.6 -> 88.9Improved
replace matrix product with Khatri-Rao product (LoKRA+) (ablation)LoKRA on LLaMA-7B · Commonsense reasoning (avg)average accuracy: 78.7 -> 79.0Improved
replace matrix product with Khatri-Rao product (LoKRA+) (ablation)LoKRA on LLaMA-13B · Commonsense reasoning (avg)average accuracy: 82.3 -> 83.1Improved
add K-rank penalty to RL with LoRATina on STILL-3-1.5B-preview · Math reasoning (6 datasets, avg)average accuracy: 48.16 -> 51.91Improved
add K-rank penalty to RL with LoRATina on Open-RS2 · Math reasoning (6 datasets, avg)average accuracy: 50.60 -> 50.95Improved
replace matrix product with Khatri-Rao product (LoKRA+) (ablation)LoKRA on DeepScalerR-1.5B-preview · Math reasoning (6 datasets, avg)average accuracy: 51.25 -> 52.40Improved
train LoRA for longer (same time as LoKRA+) (ablation)LoRA on LLaMA-7B · Commonsense reasoning (avg)average accuracy: 76.3 -> 72.7Hurt
train LoKRA for longer (same time as LoKRA+) (ablation)LoKRA on LLaMA-7B · Commonsense reasoning (avg)average accuracy: 78.7 -> 77.8Hurt

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