Attention networks engage the default mode network related to policy precision under uncertainty

Behavioural modelling of decision-making processes has advanced our understanding of impairments associated with various psychiatric conditions. However, as research increasingly prioritises the development of models that best explain observed behaviour, the question of whether these behaviours stem from biologically plausible brain functions has often been overlooked. To address this gap, we developed a probabilistic two-armed bandit task model based on the active inference framework and compared its performance to established reinforcement learning (RL) models. Our model demonstrated superior explanatory power in capturing individual variability in choice behaviour. A key parameter in our model, policy precision--analogous to the temperature parameter in RL models--is also optimised based on previous outcomes. This optimisation accounts for the balance between model-free (MF) and model-based (MB) decision-making strategies. Notably, incorporating the rate of change in policy precision enhanced the models ability to explain brain network dynamics and their inter-subject correlations. Specifically, we observed a positive correlation with default mode network dominance and a negative correlation with dorsal attention and frontoparietal network-dominant states. These opposing network patterns suggest a cooperative relationship, as evidenced by correlations between state transitions and behavioural parameters. This transition may represent a neural mechanism underlying MB-MF arbitration, which appears to be disrupted by prolonged activation of another state characterised by heightened ventral attention network activity and increased inter-network connectivity. Finally, we found that reduced prior policy precision in loss-related context is associated with suicidal ideation in individuals with major depressive disorders. HighlightsO_LIThe AIF model explains pronounced individual behavioural variability. C_LIO_LINeural signals are better explained by changes in policy precision. C_LIO_LIThe anti-correlation can be explained from the perspective of the MB-MF arbitration. C_LIO_LIThe AIF model better explains the HAM-D score. C_LIO_LIThe AIF model can discriminate suicidal ideation in MDD with a loss task. C_LI