Research map / Large language models / Instruction and fine-tuning
Geometry-Preserving Orthonormal Initialization for Low-Rank Adaptation in RLVR
Ruijia Zhang, Jiacheng Zhu, Hanqing Zhu, Laixi Shi
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 |
|---|---|---|---|
| use orthonormal principal-subspace init without singular-value scaling (LoRA-RLPO) instead of PiSSA | DeepSeek-R1-Distill-Qwen-1.5B LoRA RLVR · Average of 5 benchmarks step 500 | accuracy (%): 4.57 -> 65.03 | Improved |
| use LoRA-RLPO instead of standard LoRA | DeepSeek-R1-Distill-Qwen-1.5B LoRA RLVR · Average of 5 benchmarks step 500 | accuracy (%): 62.40 -> 65.03 | Improved |
| use LoRA-RLMO (orthonormal minor-subspace init) instead of MiLoRA | DeepSeek-R1-Distill-Qwen-1.5B LoRA RLVR · Average of 5 benchmarks step 500 | accuracy (%): 51.71 -> 63.76 | Improved |
| use LoRA-RLMO instead of standard LoRA | DeepSeek-R1-Distill-Qwen-1.5B LoRA RLVR · Average of 5 benchmarks step 500 | accuracy (%): 62.40 -> 63.76 | Improved |
| use LoRA-RLPO instead of standard LoRA | LoRA RLVR · MATH500 (mean@4) step 500 | accuracy (%): 81.93 -> 87.33 | Improved |
| use LoRA-RLMO instead of standard LoRA | LoRA RLVR · AIME22 (mean@32) step 500 | accuracy (%): 43.33 -> 42.22 | No clear effect |
| use LoRA-RLMO instead of standard LoRA | LoRA RLVR · GSM8K (mean@1) step 500 | accuracy (%): 75.64 -> 76.42 | No clear effect |
| add orthonormal data-agnostic init (DCT-LoRA, Wavelet-LoRA) to LoRA (ablation) | LoRA RLVR · DeepSeek-R1-Distill-Qwen-1.5B five math benchmarks | accuracy | Improved |
| add principal-subspace SVD geometry on top of orthonormality (LoRA-RLPO vs DCT/Wavelet-LoRA) (ablation) | LoRA RLVR · average accuracy | accuracy | Improved |
| add minor-subspace SVD geometry on top of orthonormality (LoRA-RLMO vs DCT/Wavelet-LoRA) (ablation) | LoRA RLVR · average accuracy | accuracy | No clear effect |
| decay learning rate for MiLoRA (ablation) | MiLoRA RLVR · training stability | training stability | Improved |
| decay learning rate for PiSSA (ablation) | PiSSA RLVR · training stability | training stability | No clear effect |
Closest papers by meaning
- Hybrid-LoRA: Bridging Full Fine-Tuning and Low-Rank Adaptation for Post-Training (NeurIPS 2026 · similarity 0.92)
- You Only Need Minimal RLVR Training: Extrapolating LLMs via Rank-1 Trajectories (NeurIPS 2026 · similarity 0.89)
- The Unlearnability Phenomenon in RLVR for Language Models (ICML 2026 · similarity 0.89)
- SDS-LoRA: Overcoming Anisotropic Gradient Scaling in Low-Rank Adaptation (NeurIPS 2026 · similarity 0.89)
- Tina: Tiny Reasoning Models via LoRA (ICLR 2026 · similarity 0.89)
- Learning Rate Matters: Vanilla LoRA May Suffice for LLM Fine-tuning (NeurIPS 2026 · similarity 0.89)
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.
