This page organizes papers related to artificial intelligence published around the world. This page is summarized using Google Gemini and is operated on a non-profit basis. The copyright of the paper belongs to the author and the relevant institution. When sharing, simply cite the source.
This paper aims to automate the discovery of scaling laws for model performance prediction. Drawing on over 5,000 experimental data collected from previous research, we present seven diverse scaling law discovery tasks. To overcome the limitations of existing agents, we develop SLDAgent, an evolutionary-based agent that autonomously explores complex relationships among variables by co-optimizing scaling law models and parameters. SLDAgent automatically discovers laws that consistently outperform existing human-derived laws in extrapolation, demonstrating practical utility in pre-training and fine-tuning applications. This study presents a new paradigm for agent-based scientific discovery, demonstrating that AI systems can understand their own scaling behavior and contribute new knowledge to the research community.
Takeaways, Limitations
•
Takeaways:
◦
SLDAgent automatically discovers new scaling laws that exhibit more accurate extrapolation performance than existing human-based scaling laws.
◦
The laws discovered by SLDAgent can be practically applied to pre-training and fine-tuning.
◦
AI systems present new possibilities for understanding and contributing to research on self-scaling behavior.
◦
Presenting a new paradigm for agent-based scientific discovery.
•
Limitations:
◦
There is no Limitations mentioned in the paper. (There is no Limitations mentioned directly in the paper abstract.)
