Researchers introduced a reactive optimizer that uses an agent to watch training dynamics. The agent rewrites the optimizer’s source code during training. It matches or beats hand-tuned recipes across image and language tasks. It prevents loss spikes and escapes overfitting without prior tuning. The approach is zero-shot and uses no data from repeated runs. It outperforms static recipes on modded-nanogpt, Goldfish, LightningDiT-XL/1 and RAEv2-B. The agent writes and updates two programs: one for analysis and one for updates. It dynamically adjusts learning rates, schedules and optimizer types per parameter group. The code is available in the OPAL package for public use.
What changed
Static optimizers ran unchanged through training, adapting only by pre-written rules.
What it unlocks
An optimizer can now rewrite its own code during training to react to loss spikes, stalls, or oscillation.
- 83.48 to 81.11 bits per sentence on Amharic
- 80.89 to 80.73 bits per sentence on Amharic (swept baseline)
- FID-50K 1.83 vs 2.18 (LightningDiT-XL/1, with guidance)
- FID-50K 4.33 vs 5.15 (LightningDiT-XL/1, without guidance)
What you need to act on it
- Claude Opus 5.5 or equivalent LLM to power the agent
- OPAL package integration
Sources