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
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 changed | Where | Measured | Effect |
|---|---|---|---|
| add balancing projection (BaLoRA) to LoRA | LoRA, Qwen-2.5-3B · DeepMind Mathematics rank 128 | final train loss: 1.030 -> 1.014 | Improved |
| add balancing projection (BaLoRA) to LoRA | LoRA, Qwen-2.5-3B · DeepMind Mathematics rank 8 | final train loss: 1.035 -> 1.026 | Improved |
| add balancing projection (BaLoRA) to LoRA | LoRA, Qwen-2.5-3B · DeepMind Mathematics rank 2 | final train loss: 1.050 -> 1.046 | Improved |
| use BaLoRA instead of RefLoRA | balanced LoRA · DeepMind Mathematics rank 128 | final train loss: 1.032 -> 1.014 | Improved |
| use BaLoRA instead of RefLoRA | balanced LoRA · Alpaca/CodeFeedback/OpenHermes/OpenOrca/WizardLM | eval loss: 1.350 -> 1.350 | No clear effect |
| add balancing (BaLoRA) to LoRA | LoRA, Qwen-2.5-3B · Alpaca | eval loss: 1.352 -> 1.350 | Improved |
| add balancing (BaLoRA) to LoRA | LoRA, Qwen-2.5-3B · MetaMathQA | test loss: 0.142 -> 0.144 | No clear effect |
| add balancing (BaLoRA) to LoRA | LoRA, Llama-3.2-3B · GSM8K epoch 2 | test loss: 0.492 -> 0.493 | No clear effect |
| add balancing (BaLoRA) to LoRA | LoRA, Llama-3.2-3B · Wikitext-2 | test loss: 2.278 -> 2.274 | Improved |
| add balancing (BaLoRA) to LoRA | LoRA, Qwen-2.5-3B · Wikitext-2 | test loss: 2.257 -> 2.251 | Improved |
| replace LoRA with LORO | LoRA · Alpaca | eval loss: 1.352 -> 1.504 | Hurt |
| replace LoRA with DoRA | LoRA · Wikitext-2 peak memory | memory GB: 32.63 -> 36.78 | Hurt |
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.
- RefLoRA — RefLoRA: Refactored Low-Rank Adaptation for Efficient Fine-Tuning of Large Models (NeurIPS 2025) · better on 2 of 8 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
- GoRA: Gradient-driven Adaptive Low Rank Adaptation (NeurIPS 2025 · similarity 0.94)
- Uni-LoRA: One Vector is All You Need (NeurIPS 2025 · similarity 0.94)
- Beyond Higher Rank: Token-wise Input-Output Projections for Efficient Low-Rank Adaptation (NeurIPS 2025 · similarity 0.94)
- RefLoRA: Refactored Low-Rank Adaptation for Efficient Fine-Tuning of Large Models (NeurIPS 2025 · similarity 0.93)
- LoRA-One: One-Step Full Gradient Could Suffice for Fine-Tuning Large Language Models, Provably and Efficiently (ICML 2025 · similarity 0.93)
- Make LoRA Great Again: Boosting LoRA with Adaptive Singular Values and Mixture-of-Experts Optimization Alignment (ICML 2025 · similarity 0.93)
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.
