Researchers introduce LETHE, a self-referential GAN-inspired architecture

A preprint published on arXiv introduces LETHE (Latent-parameter Evolution with Temporal Hierarchical quasi-Equilibrium), a self-referential sonic system implemented in SuperCollider. The framework utilizes concepts from Generative Adversarial Networks (GANs) within a closed configuration, requiring no external datasets or supervision after its initial setup.

Audio is processed through a 3x3 mixing matrix built around two delay lines. Its nine coefficients and two delay times evolve via the interaction between a five-feature linear discriminator and a random-perturbation optimizer analogous to single-sample REINFORCE. The discriminator evaluates current energy behavior against an archive of the initial state to drive parameter updates.

The system supports independent mixing of circular, fixed, and live audio sources. Ablation tests conducted across multiple sessions confirmed that the active generator is essential for parametric evolution.