Experimental setup.

a, Three participants first perform a collaborative space craft control task. In this task, three participants with different perspectives control a single degree of freedom of the spacecraft movement toward earth, coordinate through communication while avoiding obstacles. Afterward, each participant completed the Lottery Choice Task (LCT) individually. b, The temporal structure of a single trial of the LCT. The horizontal axis denotes time in seconds, and analyses were time-locked to stimulus onset. Each trial started with a 1 s fixation. Then, the stimulus was presented, and participants had up to 5 s to make a keep or invest decision (x ≤5). Outcome feedback was then revealed for 2 s following the decision, after which the next trial began with a fixation period. The values and risk probabilities shown in panels b and c are illustrative examples. c, Trials varied in ambiguity level. Under zero ambiguity, risk probability is fully known to participants. Under low (30%) or high (60%) ambiguity, partial risk is unknown when making decisions. d, Based on participants investment decisions across trials, they were categorized as Ideal, Aggressive, or Conservative.

Behavioral and pupil signatures across decision strategies under ambiguity.

a, Investment behavior across ambiguity levels (0%, 30%, 60%) for ideal, aggressive, and conservative investors. Heatmaps show the number and percentage of invest and keep decisions at each ambiguity level (N = 32 per group in panels a–c; total N = 96). b, Participant-level significance breakdown across strategy groups. Stacked bar plot showing the number of participants in each strategy group whose individual test result is statistically significant (P < 0.05, dark blue) versus non-significant (P ≥0.05, light blue) using χ2 test. c, Decision time as a function of ambiguity level and investor type. Bars indicate mean ± standard error of the mean (SEM); dashed lines show fitted regression trends. Horizontal bars above the plots indicate significant group differences assessed using one-way ANOVA with posthoc t-tests and Bonferroni correction (0 Ambiguity: F (2, 93) = 12.71, P < 0.0001; ideal vs. aggressive: T (31) = 3.24, P = 0.0019, aggressive vs. conservative: T (31) = −5.38, P < 0.0001. 30% Ambiguity: F (2, 93) = 4.49, P = 0.0138; aggressive vs. conservative: T (31) = −3.08, P = 0.0031). d, Difference in pupil size between ambiguous and non-ambiguous trials over time for each investor type. Time 0 indicates decision start. Color-shaded regions denote SEM; gray bar on top indicate time intervals with significant effects evaluated using Wilcoxon signed-rank tests on sliding windows with false discovery rate (FDR) correction (P < 0.05). N denotes the number of participants per strategy group. Data in the black box are shown in box plots. The box plot with paired dots show each participant’s mean pupil size change from baseline (%) measured in 0.5-1 s after decision onset under non-zero ambiguity and zero ambiguity conditions. Each dot represents one participant’s averaged pupil response, and error bars indicate the within-participant SEM. Lines connect paired observations across conditions. Statistical significance was assessed using a paired t-test within each strategy group with Bonferroni correction (ideal : T (31) = 2.62, P = 0.0400; aggressive: T (31) = 2.74, P = 0.0304; conservative: T (30) = 2.63, P = 0.0401). Asterisks indicate significance as (*P < 0.05, * * P < 0.01, * * *P < 0.001).

Decision-dependent pupil and EEG signatures across decision strategies.

a, Time-resolved pupil size difference (keep minus invest) under ambiguous conditions for ideal, aggressive, and conservative investor groups. Shaded bands denote SEM across participants. Grey bar on top indicates time windows showing significant differences (P < 0.05), assessed using Wilcoxon signed-rank tests on sliding windows with false discovery rate (FDR) correction. b, Participant-level paired comparison of mean pupil size change from baseline (%) computed within the pre-decision interval highlighted by the black box in the left panels (0.5 s before decision onset). Each dot represents one participant (error bars: within-participant SEM), and lines connect paired observations. Group-level significance was assessed using paired t-tests with Bonferroni correction (ideal : T (30) = 4.82, P = 0.0001; aggressive: T (30) = 3.92, P = 0.0014; conservative: T (30) = 3.96, P = 0.0013). Asterisks denote significance (** P < 0.01, *** P < 0.001) and N = 31 for each strategy group in a and b. c, d, Time–frequency representations of EEG power differences (keep - invest) at frontal (Fz, c) and parietal (Pz, d) electrodes for each strategy group. Black outlines indicate time–frequency regions with significant differences (P < 0.05), evaluated using Wilcoxon signed-rank tests on sliding windows with FDR correction (N = 25 for ideal and aggressive; N = 29 for aggressive). Together, pupil and EEG results reveal distinct decision-dependent arousal and neural dynamics that vary systematically across ideal, aggressive, and conservative decision strategies.

Behavioral and pupil dynamics of divergent internal models under ambiguity.

a, Ratio of keep to invest decisions under ambiguity for ideal, aggressive, and conservative investors. Conservative investors exhibit a significantly higher keep-to-invest ratio than both ideal (U = 257.50, P = 0.0026) and aggressive investors (U = 134.00, P < 0.0001). b, Distribution of the inferred expected highpayoff probability k assigned to the ambiguous portion of the lottery for each investor type. Conservative investors assign lower values of k (ideal U = 659.50, P = 0.1219; aggressive U = 735.00, P = 0.0025), consistent with a pessimistic internal model of ambiguity relative to ideal and aggressive investors. Group differences in panels a and b were assessed using one-way ANOVA with post-hoc Mann-Whitney U test and Bonferroni correction (** P < 0.01, *** P < 0.001). c, Time-resolved pupil response difference between ambiguous and non-ambiguous trials (Ambiguity minus No ambiguity) for each investor type, comparing a fixed ambiguity model (equal split of high or low payoff) versus a subjective ambiguity model (each participant’s inferred high-payoff belief Expected k). Shaded bands indicate SEM; gray bar on top denotes significant time windows of the fixed ambiguity model (Wilcoxon signed-rank test with sliding windows and FDR correction, P < 0.05). d, Strategy-dependent drift diffusion model parameters (mean ± SEM). Baseline drift rate (β0), EEG-related drift modulation, pupil-related drift modulation, starting point bias (x0), and non-decision time (t0) are shown for the ideal, aggressive, and conservative groups. Baseline drift differed across strategies (Kruskal-Wallis, H(79) = 13.46, P = 0.0012). EEG-related drift modulation showed a marginal group effect (H(79) = 5.53, P = 0.0629). Pupilrelated drift modulation and Starting point bias (X0) did not exhibit reliable group differences. Non-decision time differed across groups (H(79) = 7.47, P = 0.0238). N = 25 for ideal and conservative, N = 29 for aggressive. e-f, Leadership rating and team performance from collaborative task for each characteristic of participant. Oneway Analysis of Variance (ANOVA) with post-hoc pairwise comparisons with Tukey’s HSD correction. Statistical significance is indicated as **P < 0.01, ***P < 0.001. N = 32 for each strategy group in a, b, d, and e.