Modelling Cognitive Effort during decision-making

An active inference drift-diffusion model approach

Alvaro Garrido Perez —— 2025-11-17

Humans are increasingly relying on Artificial Intelligence (AI) systems for assistance in decision-making. To achieve truly effective human-AI collaboration, it is essential for these systems not only to predict which choices humans are likely to make, but also how demanding those choices will feel.

Cognitive effort–defined as the amount of mental resources invested in completing a specific task–plays a significant role in human decision-making. When a task requires too little effort, people tend to become bored and disengage from it. When it requires too much, they become anxious, and performance drops. Somewhere in between, there is a “sweet spot” (often referred to as the “flow state”) where cognitive demands and abilities are balanced, boosting not only performance but also well-being. An AI system with the ability to predict cognitive effort could adapt accordingly: by simplifying available choices, timing its suggestions more appropriately, or even offloading parts of the work when the user is cognitively overwhelmed.

Modelling cognitive effort, however, is a challenging business because multiple task-inducing factors can influence it–in addition to personal factors such as the level of motivation. For instance, choosing between two equally attractive alternatives is typically more effortful than choosing between a clearly preferred and a clearly inferior option (competition in expected rewards). Effort also increases when we must decide between an option we expect to be rewarding and an unknown alternative that we are curious about (exploration–exploitation trade-off). A further source of effort arises when we typically select one option out of habit, yet “rationally” infer that a different option is likely to be more beneficial (effort from overriding habitual responding).

How to model cognitive effort

In principle, we could equip AI systems with purely data-driven, black-box models that learn patterns of human effort from large datasets. While this approach can be effective, it often lacks interpretability: it becomes hard to say why the system thinks a particular decision will be effortful.

An alternative is to equip AI systems with principled models of perception and action–models that try to explain behaviour from first principles. Active Inference is one such framework. It has been successful in describing how agents (including humans) act to minimize expected surprise or, equivalently, expected free energy. This makes it attractive for modelling why a person might choose one policy over another. The main drawback, however, is that traditional Active Inference formulations have mostly been used to predict choices, not to predict how effortful those choices will be.

One of the earliest and most widely used approaches for modelling decision-making are the so-called Evidence Accumulator Models (EAMs). These models propose that, when choosing between two options, individuals accumulate evidence stochastically until a threshold is reached, triggering a decision. Remarkably, EAMs can predict not only choices but also reaction times (RTs)–a valuable feature since RTs serve as a proxy for cognitive effort. Intuitively, decisions that take longer are likely to involve greater information processing or more difficult evidence integration. However, EAMs are limited in that they do not account for how humans learn or plan, and they are typically confined to simple, one-shot choice scenarios.

The solution: Integrating Active Inference with an Evidence Accumulator Model

A promising approach to modelling cognitive effort is to combine the strengths of Active Inference and Evidence Accumulator Models (EAMs). While such hybrid approaches have already been used in reinforcement learning, to our knowledge, they had not been applied within the Active Inference framework prior to our work.

In this paper, we model decision-making and cognitive effort by integrating Active Inference with a specific class of EAMs–namely, the Drift-Diffusion Model (DDM). The integration is conceptually straightforward. We model decision-making as a drift-diffusion process, where the drift rate–that is, the average rate at which evidence accumulates toward a choice threshold–depends on quantities derived from Active Inference. Formally, the drift rate v can be expressed as a function of several Active Inference components:

Top: equation of the drift rate parameter, where $\text{Novelty}_{i}$ , $\text{Reward}_{i}$ and $\text{Habit}_{i}$ , denote the expected information gain, expected reward, and the habitual tendency to choose option $i$ , respectively. Lastly, $\text{G}^{\text{net}}_{i}$ , is the addition of the three aforementioned components of option $i$ . Bottom: schematic representation of the drift-diffusion process.

In this hybrid model, when the drift rate $v$ is small, evidence accumulates more slowly, leading to longer reaction times (RTs)—indicating a more difficult or effortful decision. Conversely, larger drift rates correspond to faster, less effortful choices.

This formulation allows cognitive effort to emerge naturally from the three situations described above: conflict arising from the exploration-exploitation trade-off, overcoming prior beliefs or habits, and competition between options with similar expected rewards, as well as from any combination of these.

To evaluate this model, we test it using the two-step task, a well-established paradigm in computational neuroscience and behavioural modelling. In this task, participants make an initial choice that probabilistically determines a second state, where they make another choice leading to a potential reward.

Schematic representation of the two-step task

Results

The results show that the Active Inference–Drift Diffusion Model (AIF-DDM) outperforms the benchmark model. It captures both the increasing tendency to select the optimal choice and the decrease in second-stage RTs as the difference in net expected free energy between options grows.

However, there is a noticeable misfit for first-stage RTs. This is likely due to the experimental design, which allowed participants to pre-plan their first-stage choice (because of relatively long inter-trial intervals) before the options were presented. As a result, the recorded first-stage RTs probably reflect motor execution rather than full deliberation.

An observed advantage of the AIF-DDM over the “vanilla” Active Inference model is that it improves parameter recovery, meaning that parameter estimates are more reliable. This effect, also reported in studies using reinforcement learning, is likely due to the additional information provided by RTs beyond choices alone. This is particularly relevant for domains such as computational psychiatry, where reliable parameter estimates are crucial for characterizing the sources of pathology.

What’s next

Future work could evaluate the hypothesis that the first-stage RT misfit is due to the experimental design. This could be achieved by collecting data from a two-step task with shorter inter-trial intervals, reducing the possibility of pre-planning.

A natural theoretical extension is to model not only the effort that arises from competition between the Active Inference components defining each option, but also the cognitive cost of computing those quantities in the first place, as well as the additional cost associated with planning (i.e., simulating future states).

In the longer term, these models could be embedded in AI systems that interact with humans to assess whether an effort-aware AI can adapt its level of assistance in real time, improve task performance, and make human-AI interaction more effective over extended periods.

If you are interested in our work, you can find the full manuscript here: https://arxiv.org/abs/2508.04435