AI Paper Atlas

Research map / Architectures and efficiency / Pruning and efficient tuning

Balanced LoRA: Removing Parameter Invariance to Accelerate Convergence

Valérie Castin, Kimia Nadjahi, Pierre Ablin, Gabriel Peyré

ICML 2026 · Architectures and efficiency › Pruning and efficient tuning · approach: LoRA variants and theory · 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 balancing projection (BaLoRA) to LoRALoRA, Qwen-2.5-3B · DeepMind Mathematics rank 128final train loss: 1.030 -> 1.014Improved
add balancing projection (BaLoRA) to LoRALoRA, Qwen-2.5-3B · DeepMind Mathematics rank 8final train loss: 1.035 -> 1.026Improved
add balancing projection (BaLoRA) to LoRALoRA, Qwen-2.5-3B · DeepMind Mathematics rank 2final train loss: 1.050 -> 1.046Improved
use BaLoRA instead of RefLoRAbalanced LoRA · DeepMind Mathematics rank 128final train loss: 1.032 -> 1.014Improved
use BaLoRA instead of RefLoRAbalanced LoRA · Alpaca/CodeFeedback/OpenHermes/OpenOrca/WizardLMeval loss: 1.350 -> 1.350No clear effect
add balancing (BaLoRA) to LoRALoRA, Qwen-2.5-3B · Alpacaeval loss: 1.352 -> 1.350Improved
add balancing (BaLoRA) to LoRALoRA, Qwen-2.5-3B · MetaMathQAtest loss: 0.142 -> 0.144No clear effect
add balancing (BaLoRA) to LoRALoRA, Llama-3.2-3B · GSM8K epoch 2test loss: 0.492 -> 0.493No clear effect
add balancing (BaLoRA) to LoRALoRA, Llama-3.2-3B · Wikitext-2test loss: 2.278 -> 2.274Improved
add balancing (BaLoRA) to LoRALoRA, Qwen-2.5-3B · Wikitext-2test loss: 2.257 -> 2.251Improved
replace LoRA with LOROLoRA · Alpacaeval loss: 1.352 -> 1.504Hurt
replace LoRA with DoRALoRA · Wikitext-2 peak memorymemory GB: 32.63 -> 36.78Hurt

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

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