Nested Learning (Hope architecture)
Frames a model and its optimiser as nested optimisation problems updating at different frequencies; Hope, a self-modifying Titans variant, targets continual learning.
Argues architecture and optimizer are one stack of memory systems with different update rates. Introduces a Continuum Memory System and the self-modifying Hope model, reporting better perplexity and long-context recall than Transformers, TTT and Mamba2 at small scale.
- Date
- Friday, 7 November 2025
- Lab
- Google Research
- Kind
- paper
- Access
- paper only
Google Research blog 2025-11-07 by Ali Behrouz and Vahab Mirrokni; the paper 'Nested Learning: The Illusion of Deep Learning Architectures' was at NeurIPS 2025 (venue per press coverage, arXiv id not captured). Small-scale results; not adopted in a frontier model.
Sources
This record was checked against its sources on 6 October 2026. How we check