Decision-theoretic testing is used to explore whether students can improve their realism, i.e. congruence between reported and true probabilities. Randomized sets of math problems were presented at computer terminals to 49 seventh graders from two classes (high/low achievers) over a period of three weeks. The subject assigned values to each of four alternatives for each problem. Two groups worked individually, one with realism feedback. A third group, also with feedback, worked in teams of size three. Only the individual-study feedback groups in the low class improved on Posttest I. On Posttest II, with harder items and peer interaction, nearly all subjects overvalued information. Deviation from realism consistently increased with task difficulty and the direction of deviation was nearly always toward overconfidence. (Author)