Reverse engineering the self

What is it about?

The paper is about how people solve problems from their own first-person point of view, e.g., knowing who they are, where they are, when they are, and what they can see and do from that spot. We call this "centered problem solving" and argue that standard computational models of rational agents (POMDPs, widely used in AI and cognitive science) leave this perspectival, self-locating aspect out. The proposal has two parts. First, we extend the standard model into "ePOMDPs," which explicitly separate the agent from its environment and represent the world from the agent's particular vantage point. Second, we add a layer on top — "meta-ePOMDPs" — that models how a person decides *which* version of themselves-in-a-situation to treat as the current problem, and when to switch (for example, realizing you're not where you thought you were on a map and re-orienting). We test the framework with two sets of video-game-style "centering games," compare it against alternative models, and find it better matches how humans actually behave, particularly the difficulty people have when they need to switch perspectives. We end by suggesting that this machinery of locating and re-locating yourself in a problem could be a starting point for a computational account of what it is to "have a self."

Featured Image

Photo by GeoJango Maps on Unsplash

Photo by GeoJango Maps on Unsplash

Why is it important?

Cognitive science has spent decades formalizing rational agency with POMDPs and Bayesian inverse-planning models, but those models typically describe behavior from an outside view rather than from inside the agent's own situated perspective. Philosophy has a parallel tradition on "self-locating" beliefs, the kind expressed by "I am here, now," which has remained largely outside computational reach. This paper operationalizes this philosophical work inside the standard cognitive-science toolkit. The current wave of agentic systems (LLM-based agents, robotic policies, world-model architectures) need to act from a specific perspective in a specific environment, often switching contexts or correcting wrong assumptions about where they are or what they're doing. Engineers are running into the same gap: existing decision-theoretic scaffolding doesn't cleanly represent who the agent is, where it is, or which problem it's currently solving. The meta-ePOMDP idea, treating the choice of which problem to solve as itself a decision problem, is directly relevant to agents that have to manage shifting tasks or notice they've misidentified their situation.