Emotional Gödel Machine: A Self-Evolving Architecture for Affective AI
Ryan SangBaek Kim · 2025
We introduce the Emotional Gödel Machine (EGM), a novel AI framework that integrates self-evolving emotional reasoning. EGM unifies four key components: (1) Emotional Self-Evolution, a Darwinian self-modification loop that rewrites both cognitive and affective code; (2) Self-Challenging Task Generation (SCETG), an internal curriculum where the agent alternates between generating and solving novel emotionally-themed tasks; (3) Interpretable Latent Feature Steering, leveraging sparse latent embeddings (via methods like FLUX/ITDA) to directly control and inspect affective outputs; and (4) Prolonged Reinforcement Learning (ProRL), an extended RL training regimen uncovering new strategies over long horizons. These elements form an open-ended learning loop: the agent revises its emotional evaluation modules (a la Gödelian self-modification), invents "emotionally challenging" problems, steers its responses through interpretable affective dimensions, and trains via long-horizon reinforcement learning (RL), an extended training process over thousands of iterations to foster novel strategies. We illustrate the EGM architecture and compare it to prior models (see Table 1),and present conceptual evaluations (Table 2) suggesting that EGM can develop richer, more human-like emotional behaviors than fixed architectures. This work bridges affective computing (Picard, 1997), self-improving AI (Zhang et al., 2025; Zhou et al., 2025), and AI ethics (Crawford, 2021; UNESCO, 2023) into a novel unified framework for next-generation self-evolving emotional AI.