Fast-Slow Training uses context optimization as fast weights alongside parameter updates as slow weights to achieve up to 3x better sample efficiency, higher performance, and less catastrophic forgetting than standard RL in continual LLM learning.
Under review as a conference paper at ICLR 2026 Clare Lyle, Mark Rowland, and Will Dabney
6 Pith papers cite this work. Polarity classification is still indexing.
6
Pith papers citing it
abstract
The reinforcement learning (RL) problem is rife with sources of non-stationarity, making it a notoriously difficult problem domain for the application of neural networks. We identify a mechanism by which non-stationary prediction targets can prevent learning progress in deep RL agents: \textit{capacity loss}, whereby networks trained on a sequence of target values lose their ability to quickly update their predictions over time. We demonstrate that capacity loss occurs in a range of RL agents and environments, and is particularly damaging to performance in sparse-reward tasks. We then present a simple regularizer, Initial Feature Regularization (InFeR), that mitigates this phenomenon by regressing a subspace of features towards its value at initialization, leading to significant performance improvements in sparse-reward environments such as Montezuma's Revenge. We conclude that preventing capacity loss is crucial to enable agents to maximally benefit from the learning signals they obtain throughout the entire training trajectory.
citation-role summary
background 2
citation-polarity summary
roles
background 2polarities
background 2representative citing papers
TeLAPA preserves behaviorally diverse policy neighborhoods in a shared latent space, improving MiniGrid continual RL transfer, revisit recovery, and retention over single-model preservation.
Different patterns of environmental change (space vs time) require different LLM update behaviors; no single family of methods—prompts, distillation, RL, or compression—handles all regimes.
Dynamical isometry (Jacobian singular values near 1) preserves plasticity in continual learning; an isometry-promoting regularizer and decoupled AdamO optimizer match or beat prior methods on supervised and RL benchmarks.
Smooth-Leaky and Randomized Smooth-Leaky activations mitigate loss of plasticity in continual learning by targeting negative-branch shape and saturation behavior.
Advocates integrating naturalistic paradigms and AI progress into cognitive science to develop generalizable models of natural behavior while retaining experimental control and theoretical insight.
citing papers explorer
-
Learning, Fast and Slow: Towards LLMs That Adapt Continually
Fast-Slow Training uses context optimization as fast weights alongside parameter updates as slow weights to achieve up to 3x better sample efficiency, higher performance, and less catastrophic forgetting than standard RL in continual LLM learning.
-
Beyond Single-Model Optimization: Preserving Plasticity in Continual Reinforcement Learning
TeLAPA preserves behaviorally diverse policy neighborhoods in a shared latent space, improving MiniGrid continual RL transfer, revisit recovery, and retention over single-model preservation.
-
When Does Continual Learning Require Learning
Different patterns of environmental change (space vs time) require different LLM update behaviors; no single family of methods—prompts, distillation, RL, or compression—handles all regimes.
-
Preserving Plasticity in Continual Learning via Dynamical Isometry
Dynamical isometry (Jacobian singular values near 1) preserves plasticity in continual learning; an isometry-promoting regularizer and decoupled AdamO optimizer match or beat prior methods on supervised and RL benchmarks.
-
Activation Function Design Sustains Plasticity in Continual Learning
Smooth-Leaky and Randomized Smooth-Leaky activations mitigate loss of plasticity in continual learning by targeting negative-branch shape and saturation behavior.
-
Naturalistic Computational Cognitive Science: Towards generalizable models and theories that capture the full range of natural behavior
Advocates integrating naturalistic paradigms and AI progress into cognitive science to develop generalizable models of natural behavior while retaining experimental control and theoretical insight.