Joint modeling of choices and response times separates working memory, cognitive control, and reinforcement learning—and reveals schizophrenia-related differences

# Why this matters

# What the researchers did They extended the RLWM experimental paradigm (which manipulates WM demands during learning) by jointly modeling both choice outcomes and full RT distributions. Models used hierarchical Bayesian parameter estimation and incorporated decision dynamics—how decisions unfold over time via parameters such as decision boundaries. They compared these joint models to models fit only to choices and tested out-of-sample prediction on a held-out, memory-free test phase.

# Main findings

• Joint modeling improved parameter recovery despite increased model complexity. It produced more accurate estimates of WM and RL contributions than choice-only fits. Choice-only fits tended to inflate RL learning rates.

• The joint model identified a proactive cognitive control strategy: participants widened their decision boundaries under higher WM load, increasing response caution and slowing RTs to accommodate anticipated memory demands. This mechanism was not apparent when fitting choices alone.
• Applying the joint framework to a clinical sample with schizophrenia revealed a pattern of deficits: slowed incremental RL learning and a failure to implement the proactive widening of decision boundaries under WM load. The model also replicated previously established WM deficits in this group. These clinical differences were masked in prior choice-only analyses.

# Why RTs change the conclusions Choices alone often leave a latent ambiguity: the same choice pattern can be produced by fast WM-guided choices or by slower RL-based updates. RT distributions add information about the internal decision process—how quickly evidence accumulates and whether participants adjust caution. Incorporating RTs lets the model attribute variance in choices to distinct mechanisms (WM vs RL vs control adjustments) rather than conflating them.

# Practical takeaways for researchers

• Fit models to both choices and RTs when the goal is to disentangle parallel learning systems and control strategies.
• Use hierarchical Bayesian estimation to stabilize parameter recovery across participants and enable out-of-sample prediction checks.
• Inspect decision-boundary parameters: they can reveal proactive control strategies that choice-only models miss.

# Clinical implications Computational phenotyping that includes decision dynamics can reveal deficits in both learning rates and control adjustments in clinical populations such as schizophrenia. Such deficits may be invisible to models that consider only choices.

# Where the code and data are available