Strong AI performance doesn't just correlate with world models — it mathematically requires them.
Researchers have published selection theorems showing that any agent achieving low regret across a range of tasks must develop specific internal structures: world models, belief-like memory, and — when tasks are mixed — persistent variables that function like emotional states. The proofs cover stochastic policies and partial observability without assuming the agent was designed with these features or given an explicit model. The mechanism works by reducing predictive modeling to binary betting decisions, then showing that regret bounds force the agent to make the right probabilistic distinctions to avoid costly errors.
The significance here is a shift from correlation to necessity. Prior work established that belief states and world models are sufficient for optimal control; this paper proves they are required for strong average-case performance. That closes a meaningful gap in the theoretical foundation for AI alignment and interpretability research — if capable agents must have these structures, then auditors have something concrete to look for.
The emotion-like variables finding will draw attention, though the paper is careful to call them functional primitives, not feelings. It's a useful reminder that "emotion" in AI safety discourse is increasingly a technical term, not an anthropomorphic one — and that the math doesn't much care about the branding.