Fig. 5: Humans also infer task rules using low-dimensional strategies.
From: Abstract rule learning promotes cognitive flexibility in complex environments across species
a Healthy adults solved an adapted version of the multidimensional rule-learning task during MEG recordings. Left: order of task rules. Right: trial structure. ITI/inter-trial interval. b Higher strategy ratio for low-dimensional strategies in humans performing the random rule vs. a random agent (two-sided Mann–Whitney test). In contrast, the ratio for high-dimensional strategies was not higher than chance (two-sided Mann–Whitney test). The ratio was higher for low- vs. high-dimensional strategies in humans (1.58 × 10−6, two-sided Wilcoxon signed rank test). c Significant Pearson correlation between the onset of the first correct strategy sequence and the performance change point mirrors rat findings. d Significant Pearson correlation between the onset of the first correct strategy sequence and a decrease in reaction time. e Replication of RL model findings in humans. Dashed lines connect values (circles) of the same subject; filled circles represent the best (i.e., highest cross-validated likelihood) model of each subject. Model comparisons between ACL and all other models are Benjamini–Hochberg corrected two-sided Wilcoxon matched pairs test with **p < 0.01, ****p < 10−4, *****p < 10−5. f–h Representative example of simulations using model parameters estimated from experimental data of one human. Color-coded sigmoidal learning curves show transitions instead of gradual learning. Sudden performance changes could only be reproduced if choice was modulated by strategy-specific attention (ACL in f), but not AL (g) or 4D-RL models (h). i Trial-based decoding of current attention focus across time (similar to rats; Fig. 4e) is shown with respect to cue onset (vertical dashed line), 33% corresponds to decoding at chance level (horizontal dashed line). j Decoding increased following cue onset (cluster-based permutation test69, comparison against pre-stimulus baseline). Shaded areas correspond to time points for which the H0 hypothesis of no difference between decoding was rejected. k Source-level topography of attention decoding during task (0–1000 ms) vs. baseline (−500 to 0 ms). Note increased attention decoding following cue onset in task-related sensory areas and in PFC (two-sided, paired t-test; t-values converted to Cohen’s d). Box plots showing median, 25%–75% percentile, whiskers: 1.5 IQR, and outliers. Created in BioRender. Böhme, N. (2025) https://BioRender.com/w17o467. Source data are provided as a Source Data file.
