06cognitive space: what reasoning changes
success here is the fitted subset (100 games per game length). click any point to load that agent into the model playground.
high-performing agents have memory close to perfect accumulation (\(\beta \approx 0\)), while weak models sit at stubborn or forgetful extremes. performance rises steeply with inference sharpness \(\kappa_f\): across llms it alone explains 63% of the variance in inference loss. sampling loss depends more on memory (27% of its variance) and the sampling gain \(m\) (8%). \(m\) grows for many strong agents but stays well below the optimal agent, which is the sampling gap again.
follow the lines from base to extended reasoning. reasoning usually moves models toward the human region: \(\beta\) moves toward zero, \(\kappa_f\) grows, inertia \(\gamma\) tends to fall and occlusion awareness \(\theta\) often rises. the shift in \(m\) is small, and choice biases are only partly reduced. reasoning improves performance mainly by sharpening evidence integration and the belief-to-choice mapping, and leaves sampling only partly corrected. fine-tuning on human behavior does not recover the human strategy either: the centaur models play near chance and sit in the low-\(\kappa_f\) region, far from humans.