- What changed
- AutoData discovered a data selection algorithm via agentic search that outperforms human-designed curation pipelines in pre-training data optimization.
- Why you should care
- Shows data engineering can be automated using agentic search, expanding AI-driven optimization beyond models and code to data pipelines.
- Your move
- Watch. Results are from a proxy model; transfer to larger scales needs further validation.
- What to watch next
- Replication of AutoData’s results on larger pretraining runs with direct downstream task evaluation beyond CORE metric.
- Event
- research
- Event date
- Sep 18, 2026
- Relevant to
- General AI readers
