aiqbench analysis
This research addresses the instability of gradient descent in large models over long timescales by proposing a meta-learning framework where agents optimize their own weights and biases. The approach combines high-dimensional inner-loop optimization with low-dimensional outer-loop updates, enabling the use of standard zeroth-order methods to accelerate learning processes without relying on traditional optimizers that fail at arbitrary time horizons.
Greedy dynamical meta-learning
arXiv:2607.23925v1 Announce Type: new Abstract: Gradient descent scales well to large models, but becomes unstable over long time horizons. Gradient-free optimizers can scale to arbitrary timespans, but are hobbled by high dimensions. Since learning occurs in large models over long timescales, neither of these approaches is likely to produce traits which can accelerate the learning process. Instead, we propose a meta-learning algorithm in which the agent learns to modify its own weights and biases. Our algorithm consists of an inner loop, wherein the agent performs some high-dimensional optimization upon itself, and an outer loop, wherein we perform some low-dimensional optimization upon the inner loop. Since the outer loop handles very few parameters, standard zeroth-order methods may be used.
Abstract:Gradient descent scales well to large models, but becomes unstable over long time horizons. Gradient-free optimizers can scale to arbitrary timespans, but are hobbled by high dimensions. Since learning occurs in large models over long timescales, neither of these approaches is likely to produce traits which can accelerate the learning process. Instead, we propose a meta-learning algorithm in which the agent learns to modify its own weights and biases. Our algorithm consists of an inner loop, wherein the agent performs some high-dimensional optimization upon itself, and an outer loop, wherein we perform some low-dimensional optimization upon the inner loop. Since the outer loop handles very few parameters, standard zeroth-order methods may be used.
Machine Learning (cs.LG)
arXiv:2607.23925 [cs.LG]
(or arXiv:2607.23925v1 [cs.LG] for this version)
https://doi.org/10.48550/arXiv.2607.23925 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 207 words
Related coverage
Memory-Efficient Meta-Reinforcement Learning for Adaptive Safety-Critical Control in Adversarial Spacecraft Proximity Operations
arXiv · Jun 17, 2026
Don't Gamble, GAMBLe: An Analytical Framework for AI-Driven Research Systems
arXiv · Jun 3, 2026
[AINews] not much happened today
Latent Space · Oct 3, 2026
PAC-Bayesian Meta-Learning for Few-Shot Identification of Linear Dynamical Systems
arXiv · Sep 22, 2026