New ARG method boosts reasoning model training on Qwen3 benchmarks

Researchers have developed a new method called Adaptive Reference Guidance (ARG) to improve the training of reasoning models in reinforcement learning. ARG balances providing a correct reference trajectory with allowing the model to generate its own, using a prefix continuation strategy. This approach aims to construct correct trajectories within a fixed generation budget. Experiments on Qwen3-4B and Qwen3-8B models across five mathematical reasoning benchmarks demonstrated that ARG achieved the highest aggregate pass@12 rate among evaluated methods. AI

IMPACT This research could lead to more efficient and effective training of AI models for complex reasoning tasks.

RANK_REASON The cluster contains an academic paper detailing a new method for training AI models. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• Adaptive Reference Guidance
• arXiv
• Group Relative Policy Optimization
• Hugging Face
• Qwen3-4B
• Qwen3-8B

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →