Research map / Architectures and efficiency / Pruning and efficient tuning
IGU-LoRA: Adaptive Rank Allocation via Integrated Gradients and Uncertainty-Aware Scoring
Xuan Cui, HuiYue Li, Run Zeng, Yunfei Zhao, Jinrui Qian, Wei Duan, Bo Liu, Zhanpeng Zhou
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 |
|---|---|---|---|
| replace instantaneous-gradient importance with parameter-space integrated gradients plus uncertainty-aware (EMA SNR) scoring | adaptive-rank LoRA (vs AdaLoRA) · GLUE average (RoBERTa-large) | average score: 88.68 -> 89.42 | Improved |
| IGU-LoRA vs AdaLoRA on reasoning suite | Qwen-2.5-0.5B · 5-task reasoning average | average accuracy: 57.67 -> 59.17 | Improved |
| IGU-LoRA vs full fine-tuning | Qwen-2.5-0.5B · GSM8K | accuracy: 34.64 -> 34.16 | Hurt |
| remove the sampled alpha integration coefficient (ablation) | IGU-LoRA · BoolQ+GSM8K avg | average accuracy: 57.82 -> 58.31 | Improved |
| use N=20 alpha candidates instead of N=4 (ablation) | IGU-LoRA · BoolQ+GSM8K avg | average accuracy: 57.93 -> 58.31 | Improved |
| use SNR (EMA mean / deviation) instead of AdaLoRA-style multiplicative sensitivity-uncertainty (ablation) | IGU-LoRA · BoolQ+GSM8K avg | average accuracy: 57.99 -> 58.31 | Improved |
| IGU-LoRA vs DoRA/LoRA on Llama-2-7B | Llama-2-7B · 5-task average | average accuracy: 89.91 (AdaLoRA) -> 90.44 | Improved |
| IGU-LoRA vs LoRA on DeepSeek-R1-Distill-Qwen-2.5-7B GSM8K | DeepSeek-R1-Distill-Qwen-2.5-7B · GSM8K | accuracy: 74.60 -> 74.28 | No clear effect |
| remove high-score modules (L3_K and L10_V) vs low-score modules (ablation) | IGU-LoRA rank allocation · BoolQ | accuracy: 82.45 -> 80.44 (high-rank removed); 82.45 -> 82.30 (low-rank removed) | Hurt |
| add integrated gradients and SNR scoring (extra overhead vs LoRA) | LoRA · Qwen-2.5-0.5B BoolQ training | training time (h): 0.42 -> 0.87 | Hurt |
| IGU-LoRA vs LoRA on MMLU | Llama-2-7B · MMLU average | average accuracy: 49.94 -> 51.07 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- DoRA — DoRA: Weight-Decomposed Low-Rank Adaptation (ICML 2024) · better on 2 of 2 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 0.93)
- Beyond LoRA vs. Full Fine-Tuning: Gradient-Guided Optimizer Routing for LLM Adaptation (NeurIPS 2026 · similarity 0.92)
- C-LoRA: Contextual Low-Rank Adaptation for Uncertainty Estimation in Large Language Models (NeurIPS 2025 · similarity 0.92)
- Unveiling LoRA Intrinsic Ranks via Salience Analysis (NeurIPS 2024 · similarity 0.92)
- Uni-LoRA: One Vector is All You Need (NeurIPS 2025 · similarity 0.92)
- AuroRA: Breaking Low-Rank Bottleneck of LoRA with Nonlinear Mapping (NeurIPS 2025 · similarity 0.91)
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.
