Trial history and model-derived reward prediction errors in mouse visual decisions
A reward prediction error (RPE) estimated from an animal's behavior can predict its next choice. Part of this predictive value may come from information about earlier trials that a choice model could also use directly. We asked whether a model-derived RPE improves choice prediction beyond the observable history of stimuli, choices, and rewards. We analyzed 275,292 trials from 107 mice performing the International Brain Laboratory visual decision task and evaluated the main choice models on sessions that were not used for fitting. Recent choices strongly improved prediction. An RPE based on a stimulus-dependent reward expectation added a small gain beyond the preceding choice and reward, and this gain fell to near zero once the model included the preceding stimulus. When the model also summarized choices and rewards over several earlier trials, an RPE trace with the same retention parameter as the history traces did not reliably improve prediction. In simulated agents with known decision rules, a simplified version of this comparison detected an RPE-based learning rule in which choices depended on accumulated prediction errors, and did not improve prediction for an agent that used trial history directly. In mice, the state-dependent RPE and its trace added little predictive information beyond observable trial history, which limits what choice prediction alone can reveal about prediction-error learning.
Authors
- Alireza Khanteymoori (ORCID: https://orcid.org/0000-0001-6811-9196)
- Amitis Mirabedini
- Raya JanatAbadi
Institutions
- Sharif University of Technology (IR)
- University of Freiburg (DE)
Publication Details
- Journal
- bioRxiv (Cold Spring Harbor Laboratory)
- Published
- 2026-11-20
- DOI
- https://doi.org/10.64898/2026.11.19.767881
- Primary Topic
- Neural and Behavioral Psychology Studies
- Type
- preprint