High WM load shifts learners from model-based toward model-free reinforcement learning strategies¶

Claim · All claims
Evidence · 1 study (2 entries) · causal r2 · q3 · i3 large
Currency · newest study 2025 · not yet checked against its sources · not reviewed by a person

Subclaims¶

q3 i3 The mixing weight, representing the degree of model-based control, was significantly reduced under high WM load (d = 0.98). → Chengyan Yang 2025 q3 i? GLMMs corroborated this: under low load previous rewards affected stay probability without state-similarity interactions (model-based pattern), whereas under high load a significant reward × state-similarity interaction emerged (model-free pattern). → Chengyan Yang 2025 (2)

Evidence¶

Chengyan Yang 2025¶

Chengyan Yang, Tongran Liu, Mengxin Wen & Xun Liu. (2025). The distinct functions of working memory and intelligence in model-based and model-free reinforcement learning. npj Science of Learning 10:72. https://doi.org/10.1038/s41539-025-00363-w

q3 · i3 · causal · r2

Paired t-tests on dual-system RL model parameters from 50 participants. The article reports the mixing weight "was also significantly reduced in the high WM load condition", with t(49) = 6.95, d = 0.98, and a Wilcoxon test r = 0.79.

"the mixing weight, which represented the degree of model-based control, was also significantly reduced in the high WM load condition compared to that observed in the low WM load condition (t(49) = 6.95, p < 0.001, d = 0.98; W = 1140, p < 0.001, r = 0.79)"

Chengyan Yang 2025 (2)¶

Chengyan Yang, Tongran Liu, Mengxin Wen & Xun Liu. (2025). The distinct functions of working memory and intelligence in model-based and model-free reinforcement learning. npj Science of Learning 10:72. https://doi.org/10.1038/s41539-025-00363-w

q3 · i? · causal · r2

Separate GLMMs on trial-level stay probabilities for each WM load condition. Under high load the optimal model showed a significant prevrew × same interaction (β = 0.357, p < 0.001), a model-free signature; under low load previous rewards had a significant positive effect (β = 1.330, p < 0.001) with no significant interaction.

"The optimal model indi- cated signi ficant main effects for both previous rewards ( β = 0.572, p < 0.001) and state similarity ( β = 0.844, p < 0.001), along with a sig- nificant interaction effect between them (β = 0.357, p < 0.001), indicating that the in fluence of previous reward was ampli fied when participants encountered identical states"

Discussion¶

Related Claims¶

• Exploitation (inverse temperature) and information utilization (learning rate) correlate with reward acquisition under both WM loads — related
• Model-based strategy use and intelligence positively predict RL learning performance, while WM load negatively predicts it — related
• Model-based strategy use partially mediates the effect of WM load on RL learning performance — related
• Participants earned higher corrected reward rates under low WM load than high WM load in a two-stage RL task — related
• Intelligence correlates with RL performance under low WM load but not under high WM load — related
• A 50% shock reinforcement rate yielded the most learners compared with 35% and 60% rates — related