# Reimagining MDPs: The Power of Minimum Action Distance > Source: > Published: 2026-07-10 19:18:25+00:00 # Reimagining MDPs: The Power of Minimum Action Distance A new approach to Markov decision processes focuses on minimum action distance, reshaping how agents understand environments without rewards or actions. Markov decision processes (MDPs) are set to undergo a transformation. This evolution hinges on a concept called minimum action distance (MAD), which redefines how agents perceive their environments. By focusing exclusively on state trajectories, this new framework allows agents to learn without reward signals or knowing the actions taken. ## The MAD Metric At its core, MAD measures the minimum number of actions needed to transition between states. It effectively captures the underlying structure of an environment. This isn't just a fancy metric. it's a foundational rethinking of how we measure progress in agentic systems. The AI-AI Venn diagram is getting thicker as MAD aligns closely with tasks like goal-conditioned [reinforcement learning](/glossary/reinforcement-learning) and reward shaping, by offering a dense metric that feels inherently geometric. This [self-supervised learning](/glossary/self-supervised-learning) approach to MDPs constructs an [embedding](/glossary/embedding) space reflective of MAD. Distances between embedded pairs mirror their MAD, supporting both symmetric and asymmetric approximations. The [compute](/glossary/compute) layer is adapting, offering new insights into agent autonomy. ## Empirical Triumphs Testing this framework across diverse environments, both deterministic and stochastic, proved its mettle. Whether dealing with discrete or continuous state spaces, or even when faced with noisy observations, MAD consistently delivered accurate representations. Moreover, it outperformed existing state representation methods. A notable achievement, indeed. But why does this matter? If agents have wallets, who holds the keys? In a world where decision-making is increasingly agent-driven, understanding the minimal steps required for state transitions could revolutionize machine autonomy. It suggests a future where machines no longer need explicit rewards or action histories to make informed decisions. ## The Future of Agentic Systems Why should readers care about yet another metric AI? Simply put, MAD offers a glimpse into the future of autonomous systems. As machines become more independent, equipping them with the ability to infer environmental structures without direct instructions could change the game. Isn't it time we rethink how agents are trained? By removing the reliance on rewards and action histories, we're building the financial plumbing for machines in a new way. The convergence of agentic and AI models isn't just a trend, it's the infrastructure for tomorrow's intelligent systems. Get AI news in your inbox Daily digest of what matters in AI. ## Key Terms Explained [Compute](/glossary/compute) The processing power needed to train and run AI models. [Embedding](/glossary/embedding) A dense numerical representation of data (words, images, etc. [Reinforcement Learning](/glossary/reinforcement-learning) A learning approach where an agent learns by interacting with an environment and receiving rewards or penalties. [Self-Supervised Learning](/glossary/self-supervised-learning) A training approach where the model creates its own labels from the data itself.