Suboptimal human inference reflects an efficient and flexible information bottleneck

Human inference is often suboptimal in ways that vary across individuals and tasks. We propose that at least some of this variability reflects information-processing limits. To test this idea, we developed a task-general application of the information-bottleneck framework to quantify how much information individuals use (information capacity) and how effectively they use it (information efficiency). We applied this framework to human choice behavior on two distinct inference tasks and found that individual differences in performance were driven largely by variation in information capacity, while information efficiency remained consistently high. This pattern of variable capacity and high efficiency also applied to a subset of participants who appeared inefficient overall, but were efficient when evaluated relative to a simplified (heuristic) encoding of the observations. Regardless of whether participants based their strategies on full or simplified encodings, they tended to adjust their performance to different task demands by changing their information capacity while maintaining efficiency. We interpret these findings in terms of new analytical results linking optimal inference under information-capacity constraints to evidence-based choice noise, suggesting that a key driver of human behavioral variability is rational adaptation to flexible information-processing limits.