Research map / Large language models / Instruction and fine-tuning
NaRA: Noise-Aware LoRA for Parameter-Efficient Fine-Tuning of Diffusion LLMs
Shuaidi Wang, Zhan Zhuang, HUANG Ruping, Yu Zhang
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 noise-aware hypernetwork core matrix (NaRA vs LoRA) | LoRA on LLaDA-8B-Instruct · Commonsense reasoning | average accuracy: 80.18 -> 84.09 | Improved |
| NaRA vs LoRA on base model | LoRA on LLaDA-8B-Base · Commonsense reasoning | average accuracy: 76.10 -> 83.83 | Improved |
| NaRA vs LoRA | LLaDA-8B-Instruct · Math reasoning (4 benchmarks avg) | average accuracy: 80.07 -> 80.89 | Improved |
| NaRA vs LoRA | LLaDA-8B-Base · Math reasoning (4 benchmarks avg) | average accuracy: 76.29 -> 77.02 | Improved |
| NaRA vs LoRA | LLaDA-8B-Instruct · Code generation | pass@1 average: 38.32 -> 39.32 | Improved |
| use Fourier embedding instead of scalar noise input (ablation) | NaRA · Math reasoning (4 benchmarks avg) | average accuracy: 80.26 -> 80.89 | Improved |
| use Fourier embedding instead of MLP embedding (ablation) | NaRA · Math reasoning (4 benchmarks avg) | average accuracy: 80.39 -> 80.89 | Improved |
| share one hypernetwork instead of separate ones per module (ablation) | NaRA · Math reasoning (4 benchmarks avg) | average accuracy: 80.13 -> 80.89 | Improved |
| eta 0.1 instead of 1.0 (ablation) | NaRA · Math reasoning (4 benchmarks avg) | average accuracy: 80.39 -> 80.89 | Improved |
| eta 0.1 instead of 0.01 (ablation) | NaRA · Math reasoning (4 benchmarks avg) | average accuracy: 80.26 -> 80.89 | Improved |
| continuous hypernetwork instead of 4 separate noise-interval LoRAs (ablation) | Multi-LoRA · Math reasoning (4 benchmarks avg) | average accuracy: 74.83 -> 80.89 | Improved |
| increase LoRA rank from 32 to 40 (ablation) | LoRA · Math reasoning (4 benchmarks avg) | average accuracy: 80.07 -> 80.19 | No clear effect |
Closest papers by meaning
- Uncertainty Quantification for Large Language Diffusion Models (NeurIPS 2026 · similarity 0.89)
- Large Language Models to Diffusion Finetuning (ICML 2025 · similarity 0.88)
- Fast and Fluent Diffusion Language Models via Convolutional Decoding and Rejective Fine-tuning (NeurIPS 2025 · similarity 0.88)
- SemanticDLM+: Improving Diffusion LLMs through Bias-variance Trade-off in Transition Kernel Design (NeurIPS 2026 · similarity 0.88)
- Block Diffusion: Interpolating Between Autoregressive and Diffusion Language Models (ICLR 2025 · similarity 0.88)
- Noise Hypernetworks: Amortizing Test-Time Compute in Diffusion Models (NeurIPS 2025 · similarity 0.88)
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.
