A new preprint on arXiv, "What Information Buys in Active Inference," takes up a question at the intersection of reinforcement learning and computational cognitive neuroscience. The abstract highlights information-theoretic exploration as a shared concern across both fields.
The authors note that established RL theory builds regret guarantees on confidence-based reasoning. This stands in contrast to the information-theoretic framing common in active inference, where exploration is driven by uncertainty reduction rather than explicit confidence intervals.
By placing these two perspectives side by side, the paper appears to ask what information-theoretic exploration contributes in active inference, and how that contribution relates to the regret bounds that RL theory cares about. The released abstract does not state the paper's conclusions, so the full text will be needed to see how the comparison resolves.