Summary: A new study explains how people solve complex problems by flexibly using strategies such as hierarchical and counterfactual reasoning. In the experiment, participants tried to predict the route of a hidden ball moving through a maze, making split-second decisions based on auditory cues and short-term memory.
Because it is infeasible for the brain to track every possible trajectory in parallel, participants simplified the task by breaking it into steps or by revising earlier choices when evidence suggested an error. Computational models show that which strategy people select depends on the reliability of their memory and the demands of the task.
Key Facts:
- Two principal strategies: Hierarchical reasoning (divide the problem into stages) and counterfactual reasoning (reconsider earlier choices) are the main approaches people use.
- Memory influences strategy: Whether participants switch strategies depends on how reliable they judge their working memory to be.
- Machine-learning confirmation: Recurrent neural networks trained with human-like processing constraints adopted similar strategies when limited in memory and parallel processing.
Source: MIT
Humans excel at breaking complicated tasks into manageable parts, which lets us solve each part with simple, effective rules.
Everyday examples show this clearly: when you go out for coffee, you decompose the activity into leaving the building, walking to the shop, and ordering. If something goes wrong at one step—an elevator is out—you can adjust that step without reworking the whole plan. This modular problem solving relies on heuristics that keep behavior efficient under real-world constraints.
Although prior behavioral evidence demonstrates this ability, it has been difficult to design experiments that precisely reveal the computational strategies people use. In the new study, researchers at MIT created an experimental task that is complex enough to require heuristics but simple enough to measure the computations behind them.
Participants were asked to predict which of four trajectories a ball would take through a maze after it entered out of view. They could not see the ball inside the maze; instead, they heard auditory cues as the ball passed two junctions. The setup makes the task essentially impossible to solve perfectly because solving it optimally would require running four parallel simulations in real time.
“You can’t track four possibilities simultaneously in your mind—it’s like trying to hold four conversations at once,” says Mehrdad Jazayeri, professor of brain and cognitive sciences at MIT and senior author of the study. The task was designed to elicit the kinds of hierarchical and counterfactual strategies that people naturally use when perfect computation is out of reach.
About 150 volunteers participated. Before the maze task, each participant’s ability to estimate short timespans—intervals on the order of hundreds of milliseconds—was measured, since timing accuracy affects how well someone can track the hidden ball.
For each person the team generated computational models that predicted error patterns under different strategies: fully parallel tracking, hierarchical reasoning alone, counterfactual reasoning alone, or hybrids. Comparing models to actual performance revealed a consistent pattern: participants typically used a hierarchical approach—choosing an initial direction at the first junction and committing to it—while sometimes switching back to the alternate path if later cues contradicted that choice.
That switch back represents counterfactual reasoning: revisiting a prior decision and imagining what would have happened had you chosen differently. Importantly, whether participants made that switch depended on how reliable they thought their memory of the earlier cues was. People who had poorer recall tended to avoid costly counterfactual revisions, while those with stronger short-term memory were more likely to revisit and correct earlier choices.
“People use counterfactuals to the extent that doing so is worth the cost,” Jazayeri explains. “If revisiting a decision causes a large performance drop for you, you’ll avoid it; if you can accurately retrieve past information, you’re more likely to use counterfactual reasoning.”
Testing limitations with neural networks
To test whether these behaviors reflect computationally rational adaptations to biological constraints, the researchers trained recurrent neural networks on the same task. Unrestricted networks could track the ball perfectly, but when the team imposed human-like constraints—limited parallel processing and noisy working memory—the networks shifted to hierarchical and counterfactual strategies that matched human behavior.
Altering the degree of memory noise in the models produced graded changes in strategy use, suggesting that humans do not switch strategies at a single threshold but rather adjust gradually as memory reliability changes. The findings indicate that hierarchical, counterfactual, and related postdictive strategies fall along a continuum of rational responses to limited computational resources.
“When we add the same constraints that humans face, networks behave like people,” Jazayeri says. “This supports the view that human decision strategies are rational given the limits under which the brain operates.”
Funding:
The work was supported by a Lisa K. Yang ICoN Fellowship, a Friends of the McGovern Institute Student Fellowship, an NSF Graduate Research Fellowship, the Simons Foundation, the Howard Hughes Medical Institute, and the McGovern Institute.
About this neuroscience research news
Author: Sarah McDonnell
Source: MIT
Contact: Sarah McDonnell – MIT
Image: The image is credited to Neuroscience News
Original Research: Open access.
“Computational basis of hierarchical and counterfactual information processing” by Mehrdad Jazayeri et al. Nature Human Behavior
Abstract
Computational basis of hierarchical and counterfactual information processing
Humans tackle complex, multistage decision problems using a mix of hierarchical and counterfactual strategies. We designed a task that reliably engages these strategies and ran hypothesis-driven experiments to identify the computational limits that produce them. We found three main constraints: a bottleneck in parallel processing that favors hierarchical decomposition, a compensatory but capacity-limited counterfactual process, and working memory noise that degrades counterfactual accuracy. Recurrent neural networks trained under systematically varied constraints reproduced human-like behavior only when all three limits were applied. Further analysis suggests hierarchical, counterfactual, and postdictive strategies lie on a continuum of rational adaptations to these constraints, offering a unified framework for understanding human cognitive flexibility and efficiency.