Mind Dreamer: Untethering Imagination via Active Causal Intervention on Latent Manifolds

Model-Based Reinforcement Learning yields sample efficiency via latent imagination, yet remains constrained by **Historical Tethering**: imagination is typically initialized from observed states. This creates a learning asymmetry, where the world model's manifold discovery outpaces the policy's sparse-reward optimization. We propose **Mind Dreamer (MD)**, a framework that instantiates **Active Causal Intervention** to transcend Markovian continuity. MD reformulates discovery as the minimization of a global Relay Expected Free Energy. Instead of initializing from historical data, it draws initial states from an adversarial generator $s_0 \sim p_{gen}(\cdot)$, creating non-continuous **latent jumps** to epistemic blind spots that are physically plausible yet cognitively challenging. We derive **Relay Value Function** and **Relay Uncertainty Function** to resolve the credit assignment paradox across these spatial ruptures. Treating synthesized anchors as interventional intermediary states, these potentials propagate pragmatic and epistemic value through Bellman-style backups. Notably, we prove that uncertainty propagation across discontinuities necessitates a quadratic discount $\gamma^2$, establishing a formal epistemic horizon. Theoretically, MD approximates a variance-minimizing importance sampler that expands the manifold's spectral gap, reducing the hitting time to critical bottleneck states. Empirically, MD achieves a **1.67$\times$ average speedup** over DreamerV3 on DeepMind Control Suite, reaching **8.8$\times$** in sparse-reward tasks.

Modern artificial intelligence often does "imagination" through "mental practice" or latent imagination—simulating future actions inside an internal mental model of the world to learn safely and quickly without making costly real-world mistakes. However, current systems suffer from a major limitation: their imagination is a "prisoner of history". They can only start dreaming from places they have already physically visited and saved in their memory logs. This creates a massive bottleneck: while the AI might have already figured out the general physics of its environment, its imagination remains stuck replaying familiar areas, wasting time on aimless random walks when trying to discover something new. To solve this, we introduce Mind Dreamer (MD), a framework that untethers the AI's imagination from its past. Instead of starting its mental practice from past memories, Mind Dreamer uses an adversarial "stress-tester" generator to actively seek out blind spots in the AI's current knowledge. It then allows the AI to perform "mental teleportation" (latent jumps) directly to the edge of its understanding—identifying states that are physically realistic but highly challenging. To help the AI learn across these broken sequences, we design specialized mathematical guides called Relay Potentials, ensuring AI can properly judge the value of jumping to an unfamiliar place without being misled by physically ungrounded fantasies or getting lost in disconnected exploration paths. Our testing shows that Mind Dreamer learns significantly faster than current state-of-the-art methods. On standard robotics and control benchmarks, Mind Dreamer reaches peak performance 1.67 times faster on average, and speeds up learning by up to 8.8 times in difficult tasks where rewards are exceptionally rare. By transforming imagination from a passive replay tool into a proactive explorer, Mind Dreamer opens the door to much more adaptable, efficient, and autonomous AI systems.