2026-10-10T07:57:02Z https://api.figshare.com/v2/oai

oai:figshare.com:article/34036265 2026-09-30T17:47:31Z category_15 category_734 category_931 category_106 portal_5 item_type_1 month_year_09_2026

RLWM task, behavior and model. Krishn Bera (25145477) Alexander Fengler (24611274) Megan A. Boudewyn (25145480) Cameron S. Carter (5421878) Molly A. Erickson (5954450) James M. Gold (25145483) Steven J. Luck (25145486) J. Daniel Ragland (19372260) Andrew P. Yonelinas (25145489) Angus W. MacDonald III (25145492) Deanna M. Barch (6759263) Michael J. Frank (8703909) Neuroscience Biological Sciences not elsewhere classified Information Systems not elsewhere classified Science Policy support robust long novel neurocognitive process limited working memory increase response caution flexibly adapt behavior despite added complexity proactive control mechanisms oriented computational phenotyping computational advances support manipulating wm demands increased wm load slowed incremental rl cognitive control processes rl learning rates disentangle separable learning cognitive control computational modeling wm vs yielded accurate test phase term retention schizophrenia revealed sample test sample prediction rlwm paradigm previous choice people rely often equally given choice decision making (A) RLWM task design. Participants perform an instrumental learning task where they learn a fixed number of stimulus-response pairs per block (set size). Each stimulus is shown 10 times within a block. On every trial, participants choose between a set of three responses. The paradigm contains a number of blocks, with set sizes ranging from 2 to 5. The WM load in the paradigm is manipulated parametrically with between-block set size manipulation and within-block delay (the number of elapsed trials since the last exposure to a given stimulus). Three example blocks are shown alongside a trial sequence. (B) Typical behavioral phenomena in the RLWM paradigm. Learning is plotted as a function of iterations per stimulus in terms of accuracy and RT. Higher set sizes exhibit less accurate and slower performance as compared to lower set sizes. Error bars show SEM. (C) Schematic overview of the PADDL (joint) model – the learning component is combined with the decision process. The learning component has parallel contributions from two modules – reinforcement learning and working memory. The RL module learns stimulus-response associations via slow, incremental learning from reward prediction errors. RL has persistent memory with no capacity constraints. The WM module allows fast learning via one-shot updating but is capacity-limited and prone to forgetting/interference. The RL and WM policies are derived independently from the respective Q-values, and the policies are combined via a weighted mixture to derive the final trial policy that also incorporates undirected noise. The decision process is a linear ballistic accumulator with collapsing bounds (LBA Angle) parameterized by bias (starting point variability), rate of bound collapse, and a decision bound that is a linear function of set size. The final trial policy informs the drifts of each response accumulator. (D) Model predictions using posterior predictive checks (for HC group). Learning curves for each set size plotted in terms of accuracy and RT for the model simulations. The model simulations were generated by sampling the posterior and using the post hoc absolute fit method. Shaded regions show 94% HDI ranges of the group mean. (<a href="http://www.ploscompbiol.org/article/info:doi/10.1371/journal.pcbi.1014796#pcbi.1014796.g001" target="_blank">Fig 1A</a> is adapted from [<a href="http://www.ploscompbiol.org/article/info:doi/10.1371/journal.pcbi.1014796#pcbi.1014796.ref044" target="_blank">44</a>], © 2025 The Authors, under CC BY 4.0.). 2026-09-30T17:47:28Z Image Figure 10.1371/journal.pcbi.1014796.g001 https://figshare.com/articles/figure/_p_RLWM_task_behavior_and_model_p_/34036265 CC BY 4.0