Murayama and Jach (M&J) suggest that high-level constructs of motivation are often used to explain behavioral findings, but that the use of this abstract terminology jeopardizes our understanding of the actual mechanisms underlying these effects. The authors suggest that we should unpack the black box and look at motivated behavior as the outcome of mental computational processes. We agree that computational models are paramount for our understanding of motivation and other cognitive processes. However, we believe that a focus restricted to the mechanistic aspects is too limited, and that such a narrow focus poses a threat to (theoretical) advancements in the broader field of cognitive (neuro)science. Here, we identify crucial limitations of a purely mechanistic approach and propose ways to reconcile these.

We will discuss the limitations of the approach put forward by M&J by relying on a level-of-analysis approach (e.g., Bechtel & Richardson, Reference Bechtel and Richardson2010; Marr & Poggio, Reference Marr and Poggio1976; Sun, Reference Sun2009). Notably, we will follow Marr's (Reference Marr1982) framework for explaining complex information processing systems that involves three levels of explanation, suggesting that we need all three levels to gain a comprehensive understanding. These levels include: (1) A computational level addressing what the system does (i.e., what is the goal of the system), (2) the algorithmic level, relating to how the system achieves this, and (3) the implementation level relates to the physical realization. Although Marr formulated this framework in the context of visual processing, these levels have been applied over and above the originally intended domain. For example, the theory of reinforcement learning (RL) has been termed the “poster child” of Marr's framework as it spans all three levels from the computational goal of reward maximization to multiple algorithmic solutions and robust neural implementations in the brain (e.g., Niv & Langdon, Reference Niv and Langdon2016).

Following the rationale of a level-of-analysis approach, M&J criticize that current research on motivation is overly concerned with high-level computational accounts, while neglecting the algorithmic realization (i.e., “mental computational processes”). In line with Marr, we want to highlight the importance of describing a complex information system on all three levels to attain a comprehensive and complete understanding of how such a system works. Naturally, this also includes the implementation level.

One missing link in the proposed approach by M&J is the biological plausibility of the proposed mechanisms. Although claims of biological plausibility are often (rightly) labeled as “empty” and “inconsistent” (e.g., Love, Reference Love2021), we use the term to emphasize the need to test any proposed mechanisms against reality using appropriate behavioral and neural measures. Instead of limiting ourselves to just one level, a multi-level and multi-measure approach is required to provide a full perspective in order to not send us astray. We agree that if your aim is to mimic problem solving or mimic behavior – for example, as a computer scientist or robotics engineer – a purely algorithmic approach may be fruitful, as implementation can be achieved in various ways. However, as cognitive neuroscientists, we are actually strongly concerned with the exact implementation and neural mechanisms underlying cognitive, affective, and behavioral functions. Therefore, we must be careful in selecting computational models (Mars, Shea, Kolling, & Rushworth, Reference Mars, Shea, Kolling and Rushworth2012; Nassar & Frank, Reference Nassar and Frank2016; Nassar & Gold, Reference Nassar and Gold2013).

Noteworthily, an important method in our toolkit to support claims about cognitive functions is the falsification of computational models by testing specific predictions and identifying evidence that contradicts them (Palminteri, Wyart, & Koechlin, Reference Palminteri, Wyart and Koechlin2017). With algorithmic accounts of high-level concepts on the rise (e.g., Brielmann & Dayan, Reference Brielmann and Dayan2022; Gershman & Cikara, Reference Gershman and Cikara2021; Shenhav et al., Reference Shenhav, Musslick, Lieder, Kool, Griffiths, Cohen and Botvinick2017) we propose such a falsification approach on the implementation to be extremely important. Again, the theory of RL provides a poignant example of where a purely mechanistic approach revealed its limitations. First, the representation of value is a central aspect in most RL models. Based on a plethora of model-driven neuroimaging studies, it was concluded that specific neurons or brain regions implement value-based RL algorithms. However, more recent and scrutinizing investigations suggest that behavioral and neural patterns are better explained by so-called policy-based RL algorithms (Hayden & Niv, Reference Hayden and Niv2021). The difference between these algorithms may seem subtle but has strong implications for theoretical accounts. Second, recent work even challenges the implementation of RL in the brain as such and proposes a model that can explain dopaminergic activity more readily (Jeong et al., Reference Jeong, Taylor, Floeder, Lohmann, Mihalas, Wu and Namboodiri2022). This demonstrates that in order to make progress in the field of cognitive neuroscience, there is a necessity to link mechanistic algorithms to neural substrates, or else we may be explaining behavior, while failing to understand how the brain implements the solutions.

M&J acknowledge that other levels of explanation have merit, but they do not provide a framework on how to bridge and integrate these levels. We propose that for a mechanistic approach to be valuable, our models need be tested against empirical data. High-level constructs can act as tools to shape our thinking, to communicate our ideas to others, and define relevant input and output measures for our algorithms. Consistent with the suggestion of the authors, behavioral data are a first important step from high-level concepts to low-level mechanisms. However, we should not stop there and continue to evaluate algorithms in light of neural measures. To reconcile the limitations of a purely mechanistic approach, we propose a multi-level, multi-measure approach. Lower-level information can help us identify “biologically plausible” models, and higher-level constructs can help us formulate measurable behavioral and neural outcomes when constructing computational models.