Researchers teach AI agents to explain their own decisions — and swap strategies between different agents

Scientists developed a method to make reinforcement learning agents' decisions interpretable by identifying discrete concepts that actually drive behavior, then use those concepts to transfer learned strategies between different agents with zero additional training. In practice, this means an AI trained one way could potentially adopt tactics from an AI trained a completely different way, if those tactics were built on similar decision concepts.

For years, reinforcement learning agents have been black boxes — they work, but nobody knows why they chose one action over another, and you can't move what one agent learned to another agent trained differently. This work shows you can reverse-engineer an agent's actual decision logic and use it as a bridge between agents. The catch: it only works in domains where decisions are naturally discrete (like Go), not continuous ones (like Atari games). So the real question is how many real-world problems look like Go versus like Atari.