[Paper] Hypernetworks That Evolve Themselves

Source: arXiv - 2512.16406v1

Overview

The paper introduces Self‑Referential Graph HyperNetworks (SR‑GHNs) – a new class of neural systems that can mutate, inherit, and adapt without any external optimizer. By embedding the evolutionary machinery inside the network itself, SR‑GHNs can autonomously evolve their own parameters and even their own mutation rates, opening a path toward truly open‑ended, self‑directed learning agents.

Key Contributions

• Self‑referential architecture: Combines hypernetworks, stochastic parameter generation, and graph‑based representations so that the network produces and evolves its own weights.
• Evolvable mutation rates: Mutation strength is treated as a selectable trait, allowing the system to automatically tune how much it varies over time.
• Benchmarks with environmental shifts: New RL tasks (CartPoleSwitch, LunarLander‑Switch) that flip dynamics mid‑run, demonstrating rapid adaptation.
• Emergent population dynamics: Shows natural‑like phenomena such as diversification, competition, and convergence without hand‑crafted evolutionary operators.
• Real‑world locomotion test: In Ant‑v5, SR‑GHNs discover coherent gaits and learn to reduce variation once a promising solution is found, hinting at fine‑grained exploitation after exploration.

Methodology

1. Graph HyperNetwork core – The model treats each neural component (e.g., a layer or sub‑module) as a node in a graph. A hypernetwork reads this graph and emits distribution parameters (mean, variance) for the actual weights of each node.
2. Stochastic weight sampling – At each “generation,” concrete weights are sampled from the emitted distributions, introducing variation directly into the network’s forward pass.
3. Self‑referential mutation – The hypernetwork also outputs a mutation‑rate vector that governs how much each distribution should be perturbed in the next generation. This vector itself is subject to the same sampling/evolution process, making mutation rates evolvable traits.
4. Evaluation loop – The sampled network is run on an RL environment, its reward is fed back as a fitness signal, and the hypernetwork parameters are updated via a simple policy‑gradient style reinforcement‑learning step. No external genetic algorithm or gradient‑based optimizer touches the sampled weights.
5. Population view – Multiple sampled instantiations coexist, forming a virtual population. Selection is implicit: higher‑reward samples contribute more to the hypernetwork’s gradient, biasing future generations toward their distribution parameters.

Results & Findings

Overall, SR‑GHNs consistently outperformed traditional RL agents that rely on fixed optimizers, especially in environments where the underlying dynamics change abruptly.

Practical Implications

• Autonomous agents in non‑stationary settings – Robots or IoT devices that must cope with hardware wear, sensor drift, or changing environments could use SR‑GHNs to self‑adjust without cloud‑based retraining.
• Reduced engineering overhead – Developers no longer need to hand‑craft mutation operators, crossover mechanisms, or schedule learning rates; the network discovers these on its own.
• Open‑ended learning platforms – Game AI, procedural content generation, or simulation‑based design tools could benefit from a system that continuously evolves novel behaviours without external supervision.
• Resource‑efficient continual learning – Because the hypernetwork learns a compact distribution over weights, storing a single model suffices to regenerate many diverse policies, saving memory compared to maintaining large populations of explicit networks.

Limitations & Future Work

• Scalability – Experiments were limited to moderate‑size RL tasks; scaling SR‑GHNs to vision‑heavy or large‑scale language models remains an open challenge.
• Training stability – The stochastic sampling can introduce high variance in gradients; the authors note occasional collapse to low‑diversity populations without careful regularization.
• Interpretability – While mutation rates emerge as traits, understanding why a particular rate is selected for a given sub‑task is still opaque.
• Future directions – The authors propose integrating richer graph topologies (e.g., dynamic node addition/removal), hybridizing with external evolutionary algorithms for bootstrapping, and testing on real‑world robotic platforms.

Authors

• Joachim Winther Pedersen
• Erwan Plantec
• Eleni Nisioti
• Marcello Barylli
• Milton Montero
• Kathrin Korte
• Sebastian Risi

Paper Information

• arXiv ID: 2512.16406v1
• Categories: cs.NE, cs.AI
• Published: December 18, 2025
• PDF: Download PDF