Anthropic日前发布研究论文，展示了一套自动化系统在可靠改进AI对齐性能方面的进展。1这项由Anthropic fellow陈悦涵领导的研究，以《自动化研究人员可靠缓解对齐失效》为题，在周五公开发布。1该系统在全部10个对齐基准测试中均实现了性能提升，同时不存在整体性能下降的情况。1
该自动化对齐研究系统的运作效率引人注目。1每次方法训练仅需30分钟，在平均6小时内，该系统最佳方案的表现就已超越经验丰富的人类研究员的水平。1成本方面，这套系统每小时的运作成本约为4美元，相比之下，人类研究员的时薪约为150美元。1研究论文的结论指出，"这些结果提供了早期证据，表明自动化对齐后训练在近期内可能变得实用。"1
Anthropic has released a research paper showcasing an automated system capable of reliably enhancing AI alignment performance without sacrificing overall capabilities.1 Led by Anthropic fellow Chen Yueh-Han, the system improved performance across all 10 alignment benchmarks designed to measure specific misalignment behaviors.1 The paper, titled "Automated Researchers Can Reliably Mitigate Alignment Failures," was published on Friday.1
The automated approach demonstrated significant efficiency. Each training iteration of the method required approximately 30 minutes, and within six hours on average, the best AAR method surpassed the performance level of experienced human researchers.1 The system also showed substantial cost advantages, operating at roughly $4 per hour compared to $150 per hour for human researchers.1 According to the paper's conclusions, these results provide early evidence that automated alignment post-training could become practical in the near term.1
评论
还没有评论，欢迎留下第一条。