Haueis and Colaço (H&C) make a compelling case that metabolic considerations are underutilized in cognitive modeling and should be more systematically integrated into modeling practice. However, the central distinction between evaluative and generative metabolic considerations may not be as theoretically robust as the authors suggest. I explain this through two worries. The first – the collapse worry – concerns the methodological role metabolism plays in the generative examples: I argue that these reduce to evaluative reasoning applied earlier in the modeling process. The second – the parallel tracks worry – asks whether their core generative example illustrates metabolic considerations doing cognitive-modeling work at all, or whether metabolic and cognitive reasoning simply run alongside each other without genuine integration.
The collapse worry
H&C characterize evaluative considerations as constraining the space of possible models by pruning implausible ones and partitioning the remainder by relative plausibility. Generative considerations, by contrast, are said to populate that space with new models. They present this as a genuine difference in the methodological role that metabolic knowledge plays in modeling. However, the examples offered in support of the generative role are better understood as instances of evaluative reasoning applied at an earlier stage of the modeling process.
Consider their core generative example: the efficiency-trade-off hypothesis of Theriault et al. (Reference Theriault, Shaffer, Dienel, Sander, Hooker, Dicker-Son, Barrett and Quigley2023). The central move here is the observation that axons thin enough to maximize information–energy efficiency are also too thin to house mitochondria, making oxidative phosphorylation structurally impossible in these fibers. Aerobic glycolysis is thus identified as the only metabolically feasible mechanism for fueling spike transmission in such axons. This is, at its core, a constraint-based argument: it rules out one computational implementation as metabolically impossible and identifies another as the only metabolically viable alternative. The predictive coding framework is then fitted to what survives this constraint. The metabolic reasoning does not so much originate the model as determine which preexisting computational frameworks are compatible with the metabolic facts. This looks like evaluation applied during model construction, not a distinct generative process.
This suggests that what H&C call generative considerations are evaluative constraint reasoning applied earlier in modeling – during construction rather than assessment. On this reading, the difference is temporal rather than logical or methodological: it concerns when metabolic constraints enter the process, not their methodological role.
The parallel tracks worry
Another problem concerns the relationship between the metabolic reasoning and the cognitive model in the Theriault et al. case. Even setting aside the collapse worry, it is unclear whether the metabolic considerations here are genuinely integrated into the cognitive model or whether they are solving a different puzzle that runs alongside it.
The metabolic puzzle the efficiency-trade-off hypothesis addresses is: why does the brain use aerobic glycolysis – a production-inefficient process – when oxidative phosphorylation is available? The answer follows entirely from structural and metabolic facts about thin axons: they cannot house mitochondria, so aerobic glycolysis is the only viable energy source during peak demand. This puzzle is resolved without any reference to prediction errors or predictive coding. The cognitive model is then appended to the metabolic solution – thin axons happen to have a firing profile compatible with prediction error transmission, so the two are connected. But this connection looks more like an observation of compatibility than genuine explanatory integration. The metabolic considerations solve a metabolic puzzle, and the cognitive model is brought in afterward to interpret what that solution means computationally.
This matters because H&C’s central thesis is that metabolic considerations should be integrated into cognitive modeling. But if metabolic reasoning is doing its main work in a different explanatory register – resolving a puzzle about brain energy use rather than directly informing computational claims – then the example illustrates the co-existence of metabolic and cognitive reasoning rather than their integration. The two are linked by the common substrate of thin axons, but the metabolic considerations do not directly shape the cognitive model itself.
For metabolic considerations to play a genuinely generative role, they would need to enter into the computational story rather than merely constraining the physical implementation of independently motivated computational architectures. The Theriault et al. example stops well short of this, establishing primarily that aerobic glycolysis is compatible with prediction-error transmission, which is a considerably weaker claim.
The same analysis applies to the supporting generative example of Pache and Rossum (Reference Pache and Van Rossum2023). Their exploration of energy-efficient learning strategies proceeds by imposing metabolic constraints on an already-existing neural network architecture trained with backpropagation – again, evaluative constraint satisfaction applied to an existing model space. But the parallel tracks worry applies here too, in a degraded form: unlike Theriault et al., there is no specific metabolic puzzle driving the inquiry. The starting point is the generic assumption that energy efficiency is a biological design principle – precisely the kind of generic metabolic consideration H&C themselves criticize in section 2.1.2 as insufficient for genuine cognitive modeling. Metabolic and cognitive reasoning remain only loosely connected rather than running on truly parallel tracks. Moreover, the model is implemented in artificial neural networks, making its biological relevance more indirect than the authors’ generative framework might ideally require. The generative account, therefore, appears to lack a clear empirical example in which metabolic reasoning demonstrably contributes to the construction of a cognitive model, rather than constraining its implementation.
Conclusion
Neither of these criticisms undermines H&C’s valuable conclusion that metabolism deserves a more prominent role in cognitive modeling. The evaluative account is well-supported and carefully developed. The generative account, however, rests on examples that either collapse into evaluation or fail to demonstrate genuine integration of metabolic considerations into cognitive models. The difficulty is, therefore, not that metabolism fails to constrain cognitive models, but that the paper’s examples do not clearly show metabolism contributing to the construction of those models unless “cognitive modeling” is defined to include metabolic implementation by default. A stronger case for the generative role would require examples in which metabolic reasoning directly contributes to the computational story, rather than solving an adjacent metabolic puzzle to which a cognitive model is subsequently attached. Whether such examples exist remains an open question.
You have
Access
Haueis and Colaço (H&C) make a compelling case that metabolic considerations are underutilized in cognitive modeling and should be more systematically integrated into modeling practice. However, the central distinction between evaluative and generative metabolic considerations may not be as theoretically robust as the authors suggest. I explain this through two worries. The first – the collapse worry – concerns the methodological role metabolism plays in the generative examples: I argue that these reduce to evaluative reasoning applied earlier in the modeling process. The second – the parallel tracks worry – asks whether their core generative example illustrates metabolic considerations doing cognitive-modeling work at all, or whether metabolic and cognitive reasoning simply run alongside each other without genuine integration.
The collapse worry
H&C characterize evaluative considerations as constraining the space of possible models by pruning implausible ones and partitioning the remainder by relative plausibility. Generative considerations, by contrast, are said to populate that space with new models. They present this as a genuine difference in the methodological role that metabolic knowledge plays in modeling. However, the examples offered in support of the generative role are better understood as instances of evaluative reasoning applied at an earlier stage of the modeling process.
Consider their core generative example: the efficiency-trade-off hypothesis of Theriault et al. (Reference Theriault, Shaffer, Dienel, Sander, Hooker, Dicker-Son, Barrett and Quigley2023). The central move here is the observation that axons thin enough to maximize information–energy efficiency are also too thin to house mitochondria, making oxidative phosphorylation structurally impossible in these fibers. Aerobic glycolysis is thus identified as the only metabolically feasible mechanism for fueling spike transmission in such axons. This is, at its core, a constraint-based argument: it rules out one computational implementation as metabolically impossible and identifies another as the only metabolically viable alternative. The predictive coding framework is then fitted to what survives this constraint. The metabolic reasoning does not so much originate the model as determine which preexisting computational frameworks are compatible with the metabolic facts. This looks like evaluation applied during model construction, not a distinct generative process.
This suggests that what H&C call generative considerations are evaluative constraint reasoning applied earlier in modeling – during construction rather than assessment. On this reading, the difference is temporal rather than logical or methodological: it concerns when metabolic constraints enter the process, not their methodological role.
The parallel tracks worry
Another problem concerns the relationship between the metabolic reasoning and the cognitive model in the Theriault et al. case. Even setting aside the collapse worry, it is unclear whether the metabolic considerations here are genuinely integrated into the cognitive model or whether they are solving a different puzzle that runs alongside it.
The metabolic puzzle the efficiency-trade-off hypothesis addresses is: why does the brain use aerobic glycolysis – a production-inefficient process – when oxidative phosphorylation is available? The answer follows entirely from structural and metabolic facts about thin axons: they cannot house mitochondria, so aerobic glycolysis is the only viable energy source during peak demand. This puzzle is resolved without any reference to prediction errors or predictive coding. The cognitive model is then appended to the metabolic solution – thin axons happen to have a firing profile compatible with prediction error transmission, so the two are connected. But this connection looks more like an observation of compatibility than genuine explanatory integration. The metabolic considerations solve a metabolic puzzle, and the cognitive model is brought in afterward to interpret what that solution means computationally.
This matters because H&C’s central thesis is that metabolic considerations should be integrated into cognitive modeling. But if metabolic reasoning is doing its main work in a different explanatory register – resolving a puzzle about brain energy use rather than directly informing computational claims – then the example illustrates the co-existence of metabolic and cognitive reasoning rather than their integration. The two are linked by the common substrate of thin axons, but the metabolic considerations do not directly shape the cognitive model itself.
For metabolic considerations to play a genuinely generative role, they would need to enter into the computational story rather than merely constraining the physical implementation of independently motivated computational architectures. The Theriault et al. example stops well short of this, establishing primarily that aerobic glycolysis is compatible with prediction-error transmission, which is a considerably weaker claim.
The same analysis applies to the supporting generative example of Pache and Rossum (Reference Pache and Van Rossum2023). Their exploration of energy-efficient learning strategies proceeds by imposing metabolic constraints on an already-existing neural network architecture trained with backpropagation – again, evaluative constraint satisfaction applied to an existing model space. But the parallel tracks worry applies here too, in a degraded form: unlike Theriault et al., there is no specific metabolic puzzle driving the inquiry. The starting point is the generic assumption that energy efficiency is a biological design principle – precisely the kind of generic metabolic consideration H&C themselves criticize in section 2.1.2 as insufficient for genuine cognitive modeling. Metabolic and cognitive reasoning remain only loosely connected rather than running on truly parallel tracks. Moreover, the model is implemented in artificial neural networks, making its biological relevance more indirect than the authors’ generative framework might ideally require. The generative account, therefore, appears to lack a clear empirical example in which metabolic reasoning demonstrably contributes to the construction of a cognitive model, rather than constraining its implementation.
Conclusion
Neither of these criticisms undermines H&C’s valuable conclusion that metabolism deserves a more prominent role in cognitive modeling. The evaluative account is well-supported and carefully developed. The generative account, however, rests on examples that either collapse into evaluation or fail to demonstrate genuine integration of metabolic considerations into cognitive models. The difficulty is, therefore, not that metabolism fails to constrain cognitive models, but that the paper’s examples do not clearly show metabolism contributing to the construction of those models unless “cognitive modeling” is defined to include metabolic implementation by default. A stronger case for the generative role would require examples in which metabolic reasoning directly contributes to the computational story, rather than solving an adjacent metabolic puzzle to which a cognitive model is subsequently attached. Whether such examples exist remains an open question.
Acknowledgments
I am grateful to the Alexander von Humboldt Foundation for a postdoctoral fellowship hosted by Prof. Albert Newen at Ruhr University Bochum.
Financial support
None.
Competing interests
None.