Research map / Architectures and efficiency / Pruning and efficient tuning
GraLoRA: Granular Low-Rank Adaptation for Parameter-Efficient Fine-Tuning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes GraLoRA, which partitions each LoRA weight update into a k×k grid of sub-blocks with independent low-rank adapters, raising effective rank by k at equal parameter count to narrow the gap with full fine-tuning.
Problem
LoRA overfits or stagnates when the rank is increased and remains below full fine-tuning performance, which the authors attribute to gradient entanglement caused by input outlier channels.
Approach
Replace the single low-rank update with a block-wise grid of k^2 small LoRA adapters, each covering one input/output sub-block, so the effective rank becomes kr with the same parameter count and FLOPs as LoRA and gradients are localized to blocks touched by outlier channels.
Key results
- On LLaMA3.1-8B HumanEval+ at rank 128, GraLoRA reaches 64.3% Pass@1 vs 55.8% for LoRA (+8.5 points).
- On commonsense reasoning, GraLoRA averages 84.6% on LLaMA3.2-3B vs 81.3% for LoRA and 84.0% for RaSA.
- On MATH with Qwen2.5-1.5B at rank 128, GraLoRA gets 28.9% vs 24.7% for LoRA.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace LoRA with GraLoRA | LLaMA3.1-8B · HumanEval+ test, rank 128 | Pass@1: 55.8% -> 64.3%; +8.5% | Helped |
| Replace LoRA with GraLoRA | Qwen2.5-1.5B · Commonsense reasoning average | accuracy: 78.7% -> 79.8% | Helped |
| Replace LoRA with GraLoRA | Qwen2.5-7B · Commonsense reasoning average | accuracy: 85.6% -> 86.4% | Helped |
| Replace LoRA with GraLoRA | LLaMA3.2-3B · Commonsense reasoning average | accuracy: 81.3% -> 84.6% | Helped |
| Replace LoRA with GraLoRA | Qwen2.5-1.5B · MATH, rank 128 | accuracy: 24.7% -> 28.9%; 4.2% | Helped |
| Replace LoRA with Best GraLoRA | RoBERTa-base · GLUE | average accuracy: 84.2 -> 86.0; 1.8% | Helped |
| Replace LoRA with GraLoRA | SDXL · Personalized image generation | CLIP similarity: 91.4% -> 91.9%; 0.5% | Helped |
| Replace LoRA with GraLoRA | SDXL · Personalized image generation | DINOv2 similarity: 79.2% -> 81.3%; 2.1% | Helped |
| Use GraLoRA with k=2 instead of k=1 (vanilla LoRA) (ablation) | LLaMA3.1-8B · HumanEval+ test, rank 32 | Pass@1 | Helped |
| Use GraLoRA with k=4 instead of k=1 (vanilla LoRA) (ablation) | LLaMA3.1-8B · HumanEval+ test, rank 128 | Pass@1 | Helped |
| Allocate part of rank to the LoRA component in Hybrid GraLoRA (ablation) | LLaMA3.1-8B · HumanEval+ test, rank 16 | accuracy | Helped |
| Allocate rank to the LoRA component in Hybrid GraLoRA (ablation) | LLaMA3.1-8B · Code generation, larger ranks | accuracy | Hurt |
Surprising findings
- Hybrid GraLoRA allocation to LoRA helps at low rank (<=16) but degrades performance at higher ranks, suggesting LoRA-style sharing causes gradient entanglement there.
Limitations
- Uniform partitioning assumed; choice of k depends on rank with a heuristic (r/k^2 about 8).
- Sub-block rank shrinks to r/k, which can hurt performance at low ranks, requiring the hybrid variant.
- Experiments mostly on LLaMA/Qwen models at a few ranks; baseline numbers for some methods are taken from prior papers.
Experimental setup
Datasets: Magicoder-Evol-Instruct-110k BoolQ PIQA SIQA HellaSwag WinoGrande ARC-Challenge ARC-Easy OpenBookQA MetaMathQA GLUE Naruto-BLIP-Captions
Benchmarks: HumanEval+ BoolQ PIQA SIQA HellaSwag WinoGrande ARC-Challenge ARC-Easy OpenBookQA MATH GLUE Naruto-BLIP-Captions personalized generation
Metrics: Pass@1 Pass@5 Pass@10 accuracy average accuracy CLIP similarity DINOv2 similarity
Baselines: LoRA MoRA RaSA MELoRA BOFT LoRA-XS LoRA-SB VeRA FourierFT DoRA PiSSA rsLoRA
Builds on: LoRA MoRA RaSA VeRA
Closest papers by meaning
- Flat-LoRA: Low-Rank Adaptation over a Flat Loss Landscape (ICML 2025 · similarity 0.93)
- Beyond Higher Rank: Token-wise Input-Output Projections for Efficient Low-Rank Adaptation (NeurIPS 2025 · similarity 0.92)
- SDS-LoRA: Overcoming Anisotropic Gradient Scaling in Low-Rank Adaptation (NeurIPS 2026 · similarity 0.92)
- LoRA-GA: Low-Rank Adaptation with Gradient Approximation (NeurIPS 2024 · similarity 0.92)
- LoRA-Pro: Are Low-Rank Adapters Properly Optimized? (ICLR 2025 · similarity 0.92)
- The Expressive Power of Low-Rank Adaptation (ICLR 2024 · similarity 0.92)
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.