Research map / Large language models / Instruction and fine-tuning
Evaluating Parameter Efficient Methods for RLVR
Qingyu Yin, Yulun Wu, Zhennan Shen, Sunbowen Lee, Zhilin Wang, Yanshu Li, Chak Tou Leong, Jiale Kang, Jinjin Gu
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 DoRA instead of LoRA | DeepSeek-R1-Distill-Qwen-1.5B RLVR · Math suite avg | average accuracy: 42.5 -> 46.6 | Improved |
| use AdaLoRA instead of LoRA | DeepSeek-R1-Distill-Qwen-1.5B RLVR · Math suite avg | average accuracy: 42.5 -> 44.2 | Improved |
| use MiSS instead of LoRA | DeepSeek-R1-Distill-Qwen-1.5B RLVR · Math suite avg | average accuracy: 42.5 -> 43.4 | Improved |
| use LoRA+ learning-rate ratio | LoRA · Math suite avg | average accuracy: 42.5 -> 43.9 | Improved |
| use PiSSA principal-component initialization | LoRA · Math suite avg | average accuracy: 42.5 -> 0.2 | Hurt |
| use MiLoRA minor-component initialization | LoRA · Math suite avg | average accuracy: 42.5 -> 18.0 | Hurt |
| freeze matrix A (LoRA-FA) | LoRA · Math suite avg | average accuracy: 42.5 -> 43.0 | No clear effect |
| use VeRA frozen-projection scaling vectors | LoRA · Math suite avg | average accuracy: 42.5 -> 40.7 | Hurt |
| use IA3 activation scaling | LoRA · Math suite avg | average accuracy: 42.5 -> 22.3 | Hurt |
| reduce LoRA rank to 1 (ablation) | LoRA · Math suite avg | average accuracy: 42.5 -> 40.5 | Hurt |
| raise LoRA rank to 16 (ablation) | LoRA · Math suite avg | average accuracy: 42.5 -> 43.9 | Improved |
| reduce batch size from 128 to 32 (ablation) | LoRA · Math suite avg | average accuracy: 42.5 -> 43.0 | Improved |
Closest papers by meaning
- Does Reinforcement Learning Really Incentivize Reasoning Capacity in LLMs Beyond the Base Model? (NeurIPS 2025 · similarity 0.91)
- The Reasoning Boundary Paradox: How Reinforcement Learning Constrains Language Models (NeurIPS 2026 · similarity 0.91)
- The Unlearnability Phenomenon in RLVR for Language Models (ICML 2026 · similarity 0.91)
- On the Direction of RLVR Updates for LLM Reasoning: Identification and Exploitation (ICLR 2026 · similarity 0.91)
- Provable Benefits of RLVR over SFT for Reasoning Models: Learning to Backtrack Efficiently (ICML 2026 · similarity 0.91)
- Reinforcement Learning with Verifiable Rewards Implicitly Incentivizes Correct Reasoning in Base LLMs (ICLR 2026 · 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.
