Automated Scaling Law Discovery: A First Step Towards AI for AI Research
We frame scaling law discovery as a benchmarked "scientific agent" task by curating thousands of experiments across diverse settings. We introduce SLDAgent, an evolution-based agent that jointly searches scaling-law functional forms and fits parameters. Across tasks, the discovered laws extrapolate more accurately than established human-derived laws, validating usefulness in both pretraining and finetuning scenarios.
Comparison of the best discovered scaling laws vs. human-derived laws across all benchmark tasks.
Note: The 4 laws shown above are examples. To view all discovered laws, please click "View" in the leaderboard below.
Loading leaderboard data... Please wait a moment if you don't see data yet.
| Rank | Agent | Model | Mean R² | Max R² | Solution |
|---|
@article{lin2025sld,
title={Can Language Models Discover Scaling Laws?},
author={Lin, Haowei and Ye, Haotian and Feng, Wenzheng and Huang, Quzhe and Li, Yujun and Lim, Hubert and Li, Zhengrui and Wang, Xiangyu and Ma, Jianzhu and Liang, Yitao and Zou, James},
journal={arXiv preprint arXiv:2507.21184},
year={2025}
}
We welcome contributions to SLDBench! Here's how you can get involved:
Contribute new scaling law discovery tasks to expand the benchmark.
Make a PR to pkuHaowei/sldbench →Upload your SLD agent's results to the leaderboard.
Make a PR to pkuHaowei/scaling_law_discovery_results →Questions? Contact: Haowei Lin (linhaowei@pku.edu.cn)