New RLVR method enhances LLM reasoning with positive-negative prompt pairing

Researchers have developed a new method called prompt-efficient RLVR that improves the training of large language models for reasoning tasks. This technique focuses on selecting prompts that provide both positive anchors and signals from rare failures, a departure from previous variance-based methods. By pairing hard-but-solvable and easy-but-brittle prompts, and using a weighted approach to amplify successes and failures, the method enhances sample efficiency and leads to significant performance gains on mathematical reasoning benchmarks. AI

影响 Introduces a more sample-efficient training method for LLMs on reasoning tasks, potentially improving performance on complex problem-solving.

排序理由 This is a research paper detailing a novel method for training large language models. [lever_c_demoted from research: ic=1 ai=1.0]

在 arXiv cs.LG 阅读 →

• AIME 2025
• AMC23
• arXiv
• Qwen2.5-Math-7B
• Qwen2.5-Math-7B-Instruct
• Xin Sheng

• 论文
• 模型发布

AI 生成摘要 · Google Gemini · 来自 1 个来源。 我们如何撰写摘要 →