Haueis and Colaço make a compelling case that metabolic considerations remain systematically underused in cognitive modeling, functioning as background assumptions rather than active constraints or generative tools. We complement their treatment of mental effort in section 4.1 with considerations from the cognitive fatigue literature that illustrate both the evaluative and generative roles their framework identifies.

A persistent methodological obstacle in this literature is that standard fatigue paradigms confound fatigue mechanisms with motivation, sleepiness, boredom, and practice effects. Most studies use sustained effortful performance to induce fatigue in participants, making it difficult to isolate the contribution of local metabolic alterations from other behavioral and computational factors. This confound has probably contributed to the stalemate between resource and computational accounts, mentioned in the review.

Ioannucci et al. (Reference Ioannucci, Chirokoff, Dilharreguy, Ozenne, Chanraud and Zénon2023) addressed this issue by inducing fatigue passively through prolonged repetitive visual stimulation of a single quadrant of the visual field. Performance on a subsequent texture discrimination task was impaired specifically for stimuli in the stimulated quadrant, and multivariate pattern analysis of fMRI data revealed degraded neural representations of that region in the visual cortex. Adjacent, unstimulated regions were unaffected. This spatial specificity instantiates the evaluative role of metabolic considerations in the sense of guidelines E4 and E5: it prunes the space of possible models by ruling out any account based on global resource depletion or prefrontal-specific mechanisms, and it constrains the locus of the relevant metabolic process to the activation history of a circumscribed neural assembly. The result does not identify a mechanism but narrows the space of possible hypotheses considerably.

Despite this constraint, there is currently no consensus on the local metabolic mechanism underlying such representational degradation. Recent data suggest glutamate accumulation could be a prominent candidate. Wiehler et al. (Reference Wiehler, Branzoli, Adanyeguh, Mochel and Pessiglione2022) reported elevated glutamate concentrations in the lateral prefrontal cortex following a day of cognitively demanding work. They proposed that glutamate accumulation signals metabolic saturation and downregulates neural processing. However, the reported effect was driven primarily by concentration decreases in the resting control group rather than increases in the fatigued group, which is the opposite of what an accumulation account would predict. Moreover, it was restricted to frontal regions, providing no generalizable mechanism for assembly-level representational failure. Therefore, this metabolic account does not satisfy guideline E2 (specifying the relevant process with sufficient precision) in this case.

We propose that the glycolytic account of Theriault et al. (Reference Theriault, Shaffer, Dienel, Sander, Hooker, Dickerson, Barrett and Quigley2023), the core generative example the authors discuss in the review, offers an alternative starting point. Thin, information-energy-efficient axons, present in large proportion in the visual cortex, do not contain mitochondria and thus depend on aerobic glycolysis. Since sustained capacity of aerobic glycolysis is constrained by the renewal rate of NAD⁺ and lactate clearance (Dienel, Reference Dienel2019), prolonged or repeated recruitment of the same neuronal assembly could progressively deplete glycolytic throughput below the level required to maintain precise signaling. This mechanism is not prefrontal-specific: it applies in principle to any assembly whose axonal composition favors thin fibers, making it consistent with the spatial specificity of the Ioannucci et al. findings. This constitutes a generative metabolic consideration in the sense of guidelines G2 and G4: it populates the space of possible models with a candidate mechanism grounded in known enzymatic constraints. More importantly, it generates testable, quantitative predictions: the degree of representational degradation (indexed by MVPA classifier accuracy) should scale with the local proportion of thin axons or with the proportion of aerobic glycolysis in the stimulated region. Closing this loop, by combining passive induction designs with spectroscopic measurement of local lactate dynamics as a proxy for glycolytic activity, would instantiate the virtuous cycle of generative and evaluative modeling that Haueis and Colaço depict in their Figure 7.