MIT's OPAL: an AI agent watches training and rewrites the optimizer's own code live
MIT_CSAIL · x · 2026-11-03
MIT CSAIL researchers (Xingjian Bai, Jia-Hao Liang, Jayden Teoh, Vincent Sitzmann) unveiled OPAL (Reactive Optimizers): placing an AI agent at the heart of the optimizer that watches training dynamics and continuously rewrites the optimizer's own source code during training.
Key points:
- Conventional optimizers fix their strategy (adaptive steps, preconditioning, schedules) before training starts; OPAL reacts to surprises like loss spikes, gradient oscillations, and stalls
- With no task-tuned recipe, OPAL matches or beats the best off-the-shelf recipe across vision and language tasks
- The agent spontaneously invents sophisticated schedules and strategies to prevent loss spikes and escape overfitting
- Open-source OPAL package available; paper at arXiv:2610.33826
More from Research
- Apple's iPrefBench tests personal agents on 130 underspecified iOS tasks: infer or ask? — hhsun1 · 2026-11-03
- NeurIPS Paper Maps Where Temporal Context Belongs in Low-Resource Video Adaptation — HildeKuehne · 2026-11-03
- CRoS: rewarding causality on scrambled CoTs boosts faithfulness 13.8% on Gemma-3-4B — mohitban47 · 2026-11-03
- Google's 'sleep' paradigm lets LLMs consolidate memories and self-modify for continual learning — teortaxesTex · 2026-11-03
- Yale lab open-sources a Claude skill that makes publication-quality schematic figures — CatAstro_Piyush · 2026-11-03
- EMNLP 2026 interpretability tutorial slides released, praised by attendees — PMinervini · 2026-11-03