Original title: Rational decisions in multi-step environments with few rollouts
Sixing Chen, Kristopher T. Jensen, Marcelo G Mattar · September 2026
When you make plans, you imagine what might happen next. In real life, one choice leads to endless future possibilities, which should be too much for your brain to process. Researchers built a mathematical theory and ran two behavioral experiments to see how we solve this. They found that the rational choice is to run only a few mental simulations, called rollouts. The first few rollouts give you the most information, so spending extra time on more simulations is not worth the cost. This works because real-world rewards are correlated, meaning what you learn from one path applies to similar paths. The theory shows that having many shallow simulations beats having fewer deep ones when resources are limited, and it explains why people sometimes make short-sighted choices without actually discounting the future. Ultimately, people get higher reward rates with few rollouts and adjust simulation depth to their available resources.
Imagine you're trying to choose the best route to a new restaurant, but you can only explore by walking a few blocks from where you are. You can't walk every possible path because it would take too long. The first few blocks you walk give you a good sense of the area, and further walking adds little new information. In this paper, the brain faces a similar situation when planning a sequence of actions: it only simulates a few steps into the future, because the first few simulations capture most of the useful information about the rewards that lie ahead.
This matters because every day we make decisions with many steps—choosing a route, planning a meal, or navigating a conversation. The brain can't simulate every possibility, so understanding how it efficiently allocates limited mental resources explains why we often make good decisions quickly without overthinking. This insight could improve artificial intelligence by making planning algorithms mimic the brain's efficient shortcut, saving time and energy.
The theory assumes correlated reward structures and may not generalize to independent reward settings. The experimental validation involves two specific behavioral tasks; broader ecological validity remains to be tested. The model treats mental simulations as the primary planning mechanism, ignoring possible interactions with other cognitive strategies. It also does not identify the neural implementation of rollout computation.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Rational decisions in multi-step environments with few rollouts — Sixing Chen, Kristopher T. Jensen, Marcelo G Mattar — PsyArXiv — doi:10.31234/osf.io/gpt39_v2
Abstract licence: CC BY
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.