Level-k Reasoning

Level-k reasoning is a model of strategic thinking in which players do not compute full Nash equilibria but instead reason iteratively about what others will do. A Level-0 player chooses randomly or follows some simple heuristic. A Level-1 player best-responds to Level-0. A Level-2 player best-responds to Level-1, and so on. The model, developed by Stahl and Wilson (1995) and Nagel (1995), predicts that most human players reason between one and three levels deep.

Level-k models explain many deviations from equilibrium: the initial rounds of public goods games, the offers in the ultimatum game, and the passing behavior in the centipede game. Unlike quantal response equilibrium, level-k reasoning captures the structure of iterative belief formation rather than just adding noise to best responses.

The deeper insight is that level-k reasoning is not a failure of rationality but a feature of bounded cognition. Real agents cannot perform infinite regress of "I think that you think that I think..." They stop at finite k, and the stopping point is determined by cognitive load, time pressure, and domain familiarity. A theory of strategic interaction that assumes common knowledge of rationality is not describing human behavior; it is describing a computation that no human performs.

Level-k reasoning reveals the hidden architecture of human strategic thought: we do not solve games, we simulate opponents. And our simulations are shallow — one, two, three levels deep — because deep simulation is computationally expensive and socially unnecessary. The players who pass in the centipede game are not failing to solve the game; they are solving a different game, one in which the opponent is modeled as a cooperative agent rather than a rational calculator. This is not bounded rationality. It is social intelligence.