It is envisioned that a high efficacy Metacognitive Layer (MCL) for Artificial Intelligence (AI) Systems (AIS) might be able to reduce AI hallucinations, improve AI construct validity, further explainability, and enhance AI Decision-Making (DM), among other value-added propositions. MCL implementations are on the rise within the AI ecosystem, and promising performance gains have already been reported. In a number of these instances, MCL is deployed atop Deep Neural Networks (DNNs). As part of the Kahneman System 1 (Fast Thinking) and System 2 (Slow Thinking) paradigm, MCL might also leverage Multi-Exit Networks (MENs), which can provide opportunities for early exits for System 1 or continued processing (later exits) for System 2. Oftentimes, these MENs are also built atop DNNs. As to be expected, the promise of the MCL is also accompanied by prospective vulnerabilities at the MEN/DNN. In particular, adversarial actors are increasingly using energy-latency attacks, which are designed to compromise the efficacy of MEN/DNN and MCL by substantially increasing the energy consumption and/or response latency. Although there are mitigation mechanisms available and being researched, revisiting the prospective “Achilles heel” attack surface of the MCL seems prudent. Adaptive Neural Networks (AdNN), the subclass of Input-Adaptive DNN (IA-DNN), and MEN (as a specific type of IA-DNN) are explored, and specific susceptibilities to attack vectors, such as the Sponge Poisoning/Attack (SPA), are exposed.

For certain AI-related combinatorial optimization problems, Hyperheuristics (HH) may have higher Energy Efficiency (EE) and results efficacy than that of a Metacognitive Module (MCM). For these cases, HH tend to excel in the strategy selection solution space (as contrasted to the parameter-tuning solution space). In particular, HH have shown some promise in the Large Reasoning Model (LRM) arena, where MCM have been beset by the paradigm of overthinking and analysis paralysis. This paradigm is not necessarily aligned with the need for heightened EE and effectiveness in the AI arena, and a performance paradox is illuminated for a number of cases (i.e., wherein more tokens actually segues to poorer performance). Given this counterintuitive System 2 dilemma conjoined with less robust contextual reasoning for System 1, there might be a useful intermediary state in the form of a System 1.5 (to serve as mitigation against overthinking) so as to advance the state of LRMs. Widely acknowledged NP-hard and NP-complete Combinatorial Optimization Problems (COPs) were selected, and HH as well as MCM were benchmarked. It turns out that the moniker of System 1.5 may be apropos, as the HH seem to provide some value-added propositions from both System 1 and 2.