Many thanks to my PhD supervisor, @shaharnitzan.bsky.social, and to our collaborators Rani Moran, Maayan Pereg and Roy Luria for their invaluable contributions.
Read the preprint here:
osf.io/preprints/ps...

On a practical note, some of what appears to be “random exploration” could be explained by modeling humans associating rewards with random noise in the task.

Do humans automatically assign credit to all task-relevant (but outcome-irrelevant) features?
Does outcome-irrelevant learning persist even when the cost of it goes up?
Do high working memory individuals encode irrelevant values but inhibit them from influencing choices, or ignore them altogether?

Computational modeling shows that outcome-irrelevant learning is quite reliable across sessions, yet not everyone does this equally. Working memory capacity strongly predicts outcome-irrelevant learning. Suggesting working memory is central for maintaining a causal structure guiding learning.

To examine the possibility that participants are not convinced by the instructions, in Experiment 2, we gave 600 trials across three days, allowing them to infer that locations should be neglected. But, we find they keep assigning credit to outcome-irrelevant locations.

So we created a “magical forest” narrative, telling participants that the offered leaves are randomly driven to their locations by the wind. We find participants still show outcome-irrelevant learning, leading them to choose suboptimally and win a smaller money bonus.

Experiment 1 (N=504) was aimed at ensuring people truly understand the causal structure of the task. Previously, it was suggested that such credit assignment is due to participants forming a wrong model of the task, rather than due to an automatic model-free credit assignment.

We asked participants to choose cards to win rewards. Some cards had higher chances of winning than others, but the card locations on the screen were completely irrelevant. No matter how hard we tried, people still assigned value to locations.