DANA/ADANA thread: momentum windows that grow with training improve scaling
_katieeverett · x · 2026-09-10
Katie Everett et al. discuss DANA/ADANA (Ferbach et al. 2025/2026), optimizers whose momentum windows grow linearly with training steps, improving scaling exponents in some regimes on power-law random features. This post asks whether a momentum schedule is even needed: longer horizons favor longer memory with smaller learning rates, and horizon-specific tuning improves AdamW baselines but doesn't fully explain ADANA's edge.
More from Research
- LightMTP: lightweight latent multi-token prediction boosts LLM planning in pretraining — gdm3000 · 2026-10-27
- EMNLP study: AI literary translation reads fine but varies more, even within one book — tuvllms · 2026-10-27
- Lab monkeys at $30-80K each: a 4-year breeding cycle is the real preclinical drug bottleneck — zebird0 · 2026-10-27
- OpenAI Foundation funds UNC tumor antigen project, hiring senior experimental scientist — iskander · 2026-10-27
- Steering attention heads toward honesty: up moral patienthood, down psychopathy — edelwax · 2026-10-27
- Are transcoders overkill? New paper argues plain neurons suffice for circuit analysis — burny_tech · 2026-10-27