AI agents develop predictive allostatic control like brains
New AI agents predict future needs 80% better with energy-aware internal states
Researcher Frederick Hayes III published work demonstrating that AI agents can develop 'predictive allostatic organization' - a computational approach to internal state management that mimics biological allostasis (maintaining stability through anticipation).
The study trained recurrent neural networks (RNNs) and spiking neural networks in an energy-constrained foraging environment where agents had to acquire resources while avoiding threats. Notably, the 'trace-augmented recurrent policy' model achieved the strongest performance, while spiking variants showed task-specific behavioral differences. The agents' internal dynamics proved predictive of future success (ROC-AUC: 0.802), with behaviorally relevant information persisting even in reduced PCA subspaces. Crucially, the predictive signals were distributed across multiple system components (trace, policy-head, internal dynamics, observations) and remained decodable even when energy-related features were removed.
Evaluation-time perturbations to temporal states, sensory inputs, or allostatic mechanisms altered agent behavior, suggesting these internal predictive states are causally involved in decision-making. Seed-balanced probes revealed measurable information about future contact events and consumption outcomes, supporting the interpretation of distributed predictive control without requiring biological validation or discrete symbolic representations.
- Frederick Hayes III developed predictive allostatic organization in RNNs and spiking neural networks
- Agents achieved 0.802 ROC-AUC in predicting safe-efficient behavior using internal predictive states
- Predictive signals were distributed across system components and remained robust under perturbations
Why It Matters
Demonstrates AI systems can develop human-like anticipatory internal states for better decision-making in uncertain environments