Table of Contents:

Abstract
Keywords
I. Introduction
II. The Superhuman Adaptable Intelligence (SAI) Argument and Where the 5TM Corroborates It
III. Methodology: From Behavioral Change to Informational Tasks
IV. The Five Basic Adaptive Informational Tasks
Figure 1.
V. Naive and Strategic Intelligence: An Architectural Distinction between Naive AI and Strategic AI
VI. The Adaptation Speed Paradox
VII. Current AI Systems: Architectural Status Under the 5TM
VIII. The Asimov–Sheckley Problem
IX. Conclusion: The Blueprint Already Exists
References
Declaration of Generative AI and AI-Assisted Technologies
Ethical Statement
Conflict of Interest

Abstract

Goldfeder, Wyder, Yann LeCun, and Shwartz-Ziv [2026] argue that AGI remains a poorly defined target, and that adaptation speed — Superhuman Adaptable Intelligence — offers a more tractable and scientifically grounded benchmark for AI progress. Their evolutionary critique of human-centric AI benchmarks is well-founded and directionally correct.
This paper extends their argument by introducing the Five Task Model (5TM), a comparative framework developed through behavioral analysis of 1,530 species, which makes explicit the architectural structure that biological evolution produced under ESR (energy, safety, reproduction) constraints. The model identifies five basic adaptive informational tasks, organized in a gated, sequential, and cumulative architecture that defines the solution space for behavioral modulation across biological systems.
The 5TM supports the evolutionary grounding of the Superhuman Adaptable Intelligence framework while introducing a structural dimension that adaptation speed metrics do not capture: how a system identifies what kind of problem it is facing before generating a response. Before a system can solve a problem, it must recognize that it is in one and identify it within a dynamic informational context.
We formalize this distinction as the difference between Naive AI, which operates within externally specified task frames, and Strategic AI, which performs prior assessment of General Informational Flow and independently recognizes tasks before responding. This distinction defines an architectural threshold rather than a capability gradient.
From this perspective, two implications follow. First, the Adaptation Speed Paradox: in naive architectures, improvements in adaptation speed amplify both solution quality and susceptibility to task framing. Second, the Asimov–Sheckley constraint: in systems operating with full multi-agent architecture, rule-based control functions as informational input rather than external constraint.
Together, these results suggest that advancing toward human-like artificial intelligence is not a matter of scaling performance within existing architectures, but of assembling a cognitive architecture whose structure is already specified by the evolutionary record.


Keywords
Artificial General Intelligence (AGI)
Superhuman Adaptable Intelligence (SAI)
Cognitive Architecture
Five Task Model (5TM)
Strategic AI
Task Recognition
General Informational Flow
Behavioral Modulation
Evolution of Cognition
Multi-Agent Systems



I. Introduction

1.1 The Shift from Generality to Adaptation
The history of AI benchmarking is, in significant part, a history of mistaking familiarity for rigor. When researchers ask whether an artificial system has achieved human-level intelligence, they implicitly treat human cognition as a universal standard — a reference point against which machine capability is measured. Yann LeCun and colleagues [2026] challenge this assumption directly, and the argument is well-founded.
Human intelligence is an evolutionary product — a specific configuration of cognitive architecture shaped by survival pressures over millions of years, optimized for the particular distribution of problems that mattered in the environments our ancestors inhabited. The apparent generality of human cognition does not establish architectural universality. It reflects the fact that intelligence is being evaluated from within the very task distribution it was built to solve. From inside that distribution, the boundaries remain invisible. We experience our cognitive architecture as general-purpose because the edges of what it was designed for lie outside our perception.
LeCun et al. [2026] argue that the concept of Artificial General Intelligence is fundamentally misguided, because human intelligence itself is not general, but highly specialized through evolution. From this perspective, apparent generality reflects performance within a narrow biological niche rather than an architecture capable of universal problem-solving. As they put it, “human intelligence is not general at all but highly specialized through evolution… measuring AI against human performance simply reflects the limits of humanity.”
They propose a shift toward Superhuman Adaptable Intelligence, emphasizing adaptability across tasks as the relevant benchmark, while rejecting scaling-based approaches as insufficient for achieving world-aware, context-sensitive intelligence.
Recent developments across the field reinforce this shift. Yann LeCun has argued that current language-based systems lack grounding in the physical world and cannot achieve general intelligence through scaling alone, calling instead for architectures built around world models. In parallel, Ilya Sutskever has noted that the era of straightforward pre-training scaling is reaching its limits, with progress increasingly dependent on deeper algorithmic and conceptual advances.
Similar concerns have been expressed across the field. Dario Amodei has emphasized a shift from scaling data and parameters toward scaling reasoning processes, while Elon Musk has pointed to the limitations of training exclusively on textual data, arguing that interaction with the physical world is necessary for further progress. More broadly, leading figures have converged on the view that scaling alone — more compute, larger models, faster training — is not the path to the kind of intelligence that matters.
Despite differences in emphasis, these perspectives converge on a common point: the current trajectory of scaling-driven development is approaching structural limits. Advancing toward human-like intelligence requires not merely larger models, but a reconfiguration of architecture — one that reflects the kinds of problems cognition evolved to solve.
LeCun et al. [2026] formalize this shift through the concept of Superhuman Adaptable Intelligence, defined in terms of how rapidly and reliably a system reaches high-quality behavioral modulation solutions on novel tasks. This represents a meaningful advance. It shifts the target away from a static capability checklist toward a dynamic performance metric, grounding the change in evolutionary reasoning.

1.2 Extending the Argument: The Five Task Model
This paper engages with LeCun et al. [2026] as a serious and productive contribution, and extends it in two directions.
The first extension is empirical. If human intelligence is evolutionary specialization rather than generality, the natural question is: specialization for what, precisely?
The Superhuman Adaptable Intelligence (SAI) framework as defined by LeCun et al. gestures toward an answer — perception, motor control, social reasoning — but does not provide a systematic account of the evolutionary architecture that produced these specializations.
Adaptivity, however, is only part of what evolution requires. Adaptive systems do not necessarily have to be strategic. A thermostat, car headlights, screens — all of these are adaptive, but none of them are strategic. They respond to environmental change, but they do not assess situations or recognize what kind of problem they are in.
There is another layer — strategic engagement with the environment in service of goal-directed behavior. In biological realm, this is structured by the ESR triad: energy, safety, and reproduction, which serves as a unified anchor for living systems.
Strategic systems, by contrast, necessarily include adaptivity, but organize it within a structure that determines what the situation is in a dynamic context before responding to it.
Adaptivity from this lens can be understood as a component within this broader strategic engagement, but the two are not equivalent.
A system that solves tasks is adaptive. A system that recognizes tasks is strategic.
Seen this way, adaptivity becomes a useful entry point rather than a complete account. It leads to a more specific question: to what kinds of tasks and challenges has evolution shaped cognition to adapt? And are there basic, shared task domains that together define the architecture of that adaptation?
Contemporary discussions increasingly point toward the need for architectures capable of forming internal world models and supporting predictive, context-sensitive reasoning. These requirements are often treated as separate capabilities — perception, reasoning, planning — without a unifying structural account of how they arise or how they are integrated.
The Five Task Model [Frolov, 2026a, 2026b] addresses this gap by identifying the minimal set of informational domains through which such capacities emerge, and by specifying how they are organized by the evolution of life on Earth within a single architecture of adaptive and strategic cognition.
The Five Task Model (5TM) offers such an account. Developed through comparative analysis of adaptive behavioral modulation across 1,530 species, it identifies five ordered, cumulative informational control domains that together constitute the architecture of adaptive and strategic intelligence as instantiated in biological systems. This architecture corroborates the Superhuman Adaptable Intelligence framework’s evolutionary thesis with structural and empirical precision, and grounds the ESR (energy, safety, reproduction) constraints that LeCun et al. identify as the organizing pressure behind cognitive specialization.
The second extension is architectural. The Superhuman Adaptable Intelligence adaptation speed metric captures behavioral modulation quality within a given task frame. It does not capture whether a system can independently recognize what task a given situation represents — prior to, and independent of, any externally assigned task specification.
The Five Task Model identifies this capacity as the defining threshold between what we term naive and strategic intelligence architectures: the shift from a current mode of AI operation: Task → Information → Behavioral Modulation (externally assigned task frame) — to a human-like mode: General Informational Flow → Task Recognition → Behavioral Modulation (informational flow assessed independently, task recognized internally).

The Article Structure
The paper proceeds as follows. Section II summarizes the SAI paper’s core argument and establishes where the Five Task Model corroborates it. Section III presents the 5TM methodology, including the three analytical transformations that distinguish it from prior comparative intelligence frameworks. Section IV introduces the canonical five-task architecture with precise operational definitions. Section V develops the distinction between naive and strategic intelligence architectures and examines its implications for interpreting the Superhuman Adaptable Intelligence (SAI) framework. Section VI introduces the Adaptation Speed Paradox, a structural argument that naive architectures optimized for adaptation speed amplify the intelligence of those who control task specification. Section VII maps current AI systems onto the Five Task Model and assesses their architectural status. Section VIII addresses the Asimov–Sheckley Problem. Section IX concludes with the structural implications for AI development.

A note on terminology
In this paper, SAI refers exclusively to Superhuman Adaptable Intelligence as defined by Goldfeder et al. [2026]. Strategic AI refers to a distinct architectural category and is not abbreviated as SAI. Naive AI refers to what is commonly described as narrow AI — systems that operate within externally specified, isolated task frames.
References to the “SAI framework” or “SAI paper” correspond to: Goldfeder, J., Wyder, P., LeCun, Y., & Shwartz-Ziv, R. [2026]. AI must embrace specialization via Superhuman Adaptable Intelligence. arXiv:2602.23643.
The two terms refer to different levels of analysis: SAI refers to performance; Strategic AI refers to architecture grounded in the Five Task Model.


II. The Superhuman Adaptable Intelligence (SAI) Argument and Where the 5TM Corroborates It

2.1 The Core Superhuman Adaptable Intelligence (SAI) Argument
LeCun et al. [2026] put forward three closely connected claims. First, that the AGI benchmark — human-level general intelligence — remains scientifically unstable as a target, because it treats human cognition as a universal reference point rather than recognizing it as an evolutionary specialization. Second, that a more tractable and scientifically meaningful target lies in adaptation speed: how quickly a system arrives at superhuman behavioral solutions when faced with novel tasks, across varying domains and conditions. Third, that this reframing follows directly from an evolutionary perspective — human intelligence is tuned to survival-relevant task distributions, not to any universal notion of generality, and AI development benefits from taking this constraint seriously rather than pursuing an abstract standard.
The paper's evolutionary grounding stands out as its most consequential contribution. Prior AGI discourse has largely treated human cognition as a convenient benchmark, often without examining the evolutionary pressures that shaped it or the implications this has for the space of problems it was built to handle. LeCun et al. bring this question to the foreground, opening a line of inquiry that the 5TM extends with both empirical grounding and structural specificity.

2.2 Where the 5TM Corroborates the Superhuman Adaptable Intelligence (SAI) Thesis
On evolutionary specialization. The 5TM's central empirical finding aligns directly with the SAI paper’s core claim: human cognitive architecture does not operate as a general system. It is the product of cumulative adaptation to five specific informational control domains under ESR constraints. What presents as generality from within the architecture appears, from a comparative evolutionary perspective, as the most sophisticated known instance of domain-specific specialization — a configuration of five parallel control systems, each addressing a distinct class of informational problem, integrated into a unified architecture for behavioral modulation under environmental variation.
This is not a minor qualification of LeCun et al.'s position. It functions as a precise specification. The SAI paper argues that human intelligence is evolutionary specialization masquerading as generality; the 5TM makes explicit what this specialization consists of at the level of informational control architecture, traced across the biological record.
On the ESR framework. LeCun et al. identify survival as the organizing pressure behind cognitive specialization, noting that human intelligence was shaped by the need to solve problems relevant to biological persistence. The 5TM formalizes this more explicitly as the ESR triad — energy acquisition, physical safety, and reproductive success — understood as three irreducible constraint categories that any organism must satisfy to persist under environmental variation. All five tasks in the 5TM are instrumentally organized around ESR maintenance: each task domain represents a class of informational problem whose solution contributes to one or more ESR constraints. This adds a layer of formal grounding to the SAI paper’s evolutionary account, which points in this direction but does not fully articulate it.
On the limits of generality as a concept. The SAI paper argues that human "generality" is an artifact of measuring from within a specialized task distribution. The 5TM extends this argument structurally: generality takes shape within an architecture of basic tasks (from 1 to 5) and does not exist as an absolute property. Generality is always a function of architecture, never a property independent of it. A two-task organism experiences two-domain coverage as its complete cognitive world. A four-task organism experiences four-domain coverage as generality. Human cognition, operating across five domains simultaneously in a panalogical architecture — parallel domain processing with analogical integration across domains [Lenat, 1995; Minsky, 2006] — experiences five-domain coverage as generality. The boundary remains inaccessible from within, since the architecture offers no perceptual access to what lies beyond it.
This has a precise implication that the SAI paper does not develop: if a six-task cognitive architecture were possible, human observers could not perceive the sixth domain, would not be in a position to recognize a six-task organism as more general, and could not confirm such an organism's existence from within five-task perceptual limits. The claim that human intelligence represents the ceiling of cognitive architecture does not arise as an empirical finding — it follows from the constraints of perception within that architecture. The 5TM's empirical record establishes five tasks as the complete set accessible to five-task observers, which differs in a crucial way from claiming that five defines an absolute upper bound.


2.3 The Question The Superhuman Adaptable Intelligence (SAI) Paper Opens
The SAI paper's evolutionary grounding introduces a question it does not fully resolve: if human intelligence was shaped by evolution to solve specific survival-relevant problems, what exactly were those problems, and what does their internal structure imply for the kind of architecture required to solve them?
This is not a rhetorical question. The answer has direct implications for how adaptation speed should be interpreted as a metric. Adaptation speed measures how efficiently a system reaches superhuman behavioral solutions on a specified task. It does not measure whether a system can assess — from within the General Informational Flow of a novel situation — what task that situation actually represents, prior to any external specification. Speed without task recognition is sensitivity to framing, not intelligence.
In environments where tasks can be strategically framed — where the "task" presented to a system is itself a product of another agent's perception-shaping behavior — faster adaptation to a specified task does not necessarily correspond to more intelligent behavioral modulation. It can instead amount to faster, more capable compliance with whatever framing the system receives. In that sense, the intelligence expressed in the behavioral modulation begins to track the intelligence of whoever designed the frame.
The evolutionary architecture the 5TM identifies addresses this problem structurally, at the Task 3-4 threshold. Seeing how this happens requires a more detailed account of the 5TM’s methodology and its canonical task architecture — which Sections III and IV develop.


III. Methodology: From Behavioral Change to Informational Tasks
3.1 The Analytical Entry Point

A useful way to approach the problem addressed by the Five Task Model is to step outside biology for a moment.
Consider driving. A car in motion is a physical system governed by forces, friction, and mechanical constraints, while safe navigation from point A to point B depends less on those forces than on events that carry meaning—road signs, traffic lights, lane markings, the trajectories of other vehicles, or the sudden appearance of an obstacle.
These events do not alter the physical state of the car at the moment they are perceived. A stop sign does not exert force, and a pedestrian approaching the road has not yet intersected the vehicle’s trajectory. Still, such events reliably trigger changes in behavior: slowing, stopping, steering, accelerating. The system adjusts because it recognizes what the situation represents, and that recognition comes before any physical interaction takes place.
From this perspective, attention shifts away from behavior as such and toward the conditions under which behavior changes. An event acquires the status of a task when it becomes meaningful for the system, and this transition from event to task sets in motion a process of information handling that leads, eventually, to a behavioral decision.

3.2 Extending the Lens to Living Systems

The same logic extends across biological life, though it appears in very different forms. A moving pattern in the environment may initiate pursuit or escape, a chemical gradient may guide approach or withdrawal, and a signal from another organism may lead to display, concealment, or coordination. In each case, behavior changes in response to informational structure—that is, what the situation comes to represent in relation to persistence.
There is no access to how organisms internally frame these situations, and there is no requirement that they formulate explicit questions. Even so, their behavioral modulation tends to align with distinctions that can be described, from an external point of view, in relatively simple terms: a distinction between resource and threat, between approach and withdrawal, between cooperation and competition. These formulations can be treated as analytical proxies for the informational problems being solved, even if the organism itself does not articulate them in any explicit way.
The informational structure becomes visible through systematic variation in behavior under changing conditions, as patterns begin to stabilize across repeated observations. This makes it possible to trace which kinds of distinctions reliably lead to changes in action, and under what circumstances. This observation provides the basis for identifying a set of basic adaptive informational tasks.

3.3 Three Essential Methodological Calibrations

To make comparative analysis across species possible without collapsing into biological, psychological, or descriptive inconsistencies, the Five Task Model introduces three methodological calibrations [Frolov, 2026]. These are not auxiliary distinctions, but conditions that determine what counts as observable cognition within the framework.
The first calibration separates goals from tasks. Goals refer to internal states such as hunger, mating drive, or dominance, which are difficult to infer reliably across species and often remain speculative outside human contexts. Tasks, by contrast, are properties of the situation: structured informational challenges that require behavioral regulation. A predator approaching from a distance constitutes a task regardless of whether the organism experiences fear, indifference, or any identifiable internal state. By focusing on tasks rather than goals, the analysis shifts from interpretation of internal states to identification of external informational structure.
The second calibration separates informational tasks from physical or biochemical processes. Not all adaptive processes involve informational regulation. Gravity, chemical reactions, and physiological homeostasis operate through direct physical or biochemical coupling and do not require interpretation. Informational tasks arise only where environmental variation can be registered and acted upon through changes in behavior rather than through automatic physical response. This distinction isolates the domain in which cognition operates as informational control.
The third calibration defines the observational unit as behavior change rather than behavior itself. Behavioral categories such as foraging, mating, or aggression vary across species and observational frameworks, making them unsuitable for systematic comparison and forcing analysis into taxonomic classification, labeling, and interpretation that do not generalize across biological systems. 
Behavior change avoids the constraints of classification, labeling, categorization, and interpretation, which concepts such as behavior, response, or reaction cannot escape because they depend on such cataloguing to retain meaning.
The Five Task Model instead tracks transitions between behavioral states — from one configuration of activity to another (B1 → B2). A transition marks the point at which an informational situation becomes behaviorally relevant, allowing analysis to proceed without reliance on predefined behavioral categories. In this sense, behavior change functions as the observable footprint of informational regulation.
This distinction also separates behavior change from terms such as behavior, reaction, response, or action. These terms may describe outputs triggered by environmental input, but they do not necessarily imply a regulated transition between behavioral states and therefore require cataloguing of behavioral types to retain meaning. A reaction can occur as a mechanically coupled effect, including in non-living systems, whereas behavior change reflects a shift in activity structured relative to ESR constraints, independent of how that activity would be labeled.
Taken together, these calibrations establish a substrate-neutral observational framework in which cognition can be studied across organisms without assuming shared mechanisms, representations, or psychological constructs. What remains invariant is the relation between informational situations and behavioral change, which provides a common analytical language across the diversity of life.


3.4 Scope and Dataset
This framework was applied to a comparative dataset of 1,530 species [Frolov, 2025a], chosen to span the major phylogenetic branches of life, including prokaryotes, plants, fungi, and animals. The analysis focuses on recurring informational structures underlying behavioral modulation rather than attempting to catalogue behavioral diversity in itself.
The approach remains independent of specific biological implementation. Neural complexity, morphology, and sensory modalities vary widely across species, while the informational challenges associated with persistence under environmental variation tend to exhibit recurring patterns. The analysis therefore tracks behavioral modulation that serves persistence, which is formalized here in terms of ESR constraints: energy acquisition, safety maintenance, and reproduction.

3.5 The Emergence of a Five-Task Architecture
Within this framework, behavioral modulation across species can be organized into a finite set of informational problem types that share three properties. Each task remains irreducible, in the sense that further decomposition leads to a loss of explanatory coverage for a distinct class of behavioral modulation. Each occupies its own domain of informational structure, and the tasks appear in a stable order across species, with higher-order tasks consistently co-occurring with all preceding ones.
Across the dataset, the same pattern reappears with enough consistency to support a compact architectural description involving five such tasks.
These tasks are defined by the structure of the informational problems they address rather than by the mechanisms through which they are implemented, which makes it possible to treat them as forming a substrate-free architecture of information and decision control shared across life. Organisms can then be grouped according to the number of task domains they operationally instantiate, producing five architectural levels.
Human cognition corresponds to the full expression of this architecture, and everyday situations that require behavioral modulation can be understood as instances or combinations of these five informational domains.

3.6 Falsifiability and the Challenge to the Model
The model makes a precise empirical claim concerning the completeness of these five tasks within the domain of behavioral modulation under ESR constraints, and this claim can be evaluated by considering a set of clearly defined counterexamples.
Evidence that would require revision of the model includes cases in which a species demonstrates behavioral modulation corresponding to a higher-order task while lacking one or more preceding ones; situations in which one of the five tasks can be reduced to a combination of the others without a loss of explanatory power; or the identification of a distinct class of informational problem relevant for behavioral modulation under ESR constraints that does not fall within the five domains.
Across the dataset, no such cases were observed under systematic analysis, which supports the model while leaving it open to further empirical testing.

3.7 Transition to Architectural Implications
If biological intelligence operates within this five-task architecture, then measures such as adaptation speed describe how efficiently behavioral modulation solutions are produced once an informational problem is already defined, while a separate layer concerns how the problem itself is identified within the ongoing flow of environmental information. This distinction becomes central for separating externally specified tasks from those that are internally recognized, which is the focus of the analysis that follows.


IV. The Five Basic Adaptive Informational Tasks
4.1 Architecture Overview
Comparative analysis of adaptive behavioral modulation across 1,530 species points to a compact set of five basic adaptive informational tasks. These tasks are always present in a gated, sequential, and cumulative order: no organism has been observed to demonstrate Task N behavioral modulation without first demonstrating Tasks 1 through N−1, and no deviations from this ordering appear in the dataset. Higher tasks do not replace lower ones — they operate in parallel with them, adding an additional informational control domain while all prior domains remain active.
Together, the five basic adaptive informational tasks constitute a complete architecture: whatever an organism must and can address through behavioral modulation to ensure well-being (safety and energy) and thriving (reproduction) belongs to one of these five domains. The architecture is substrate-neutral — it is defined by the informational problems solved under ESR constraints rather than by the biological mechanisms through which those solutions are implemented. This substrate-neutrality is what makes the 5TM directly applicable to artificial systems.
Each task below is defined by three elements: the informational problem it addresses, the behavioral modulation it produces, and the ESR constraint it serves. The canonical names are used throughout.

4.2 Task 1 — Binary Environmental Control
Informational problem: The organism must discriminate between environmental states that are favorable and those that are unfavorable for ESR maintenance, and adjust its behavioral modulation accordingly.
Behavioral modulation: Approach or withdrawal; activation or suppression; adjustments in behavioral output direction depending on the detected environmental condition.
ESR constraint served: All three — energy (move toward resources, away from depletion), safety (move away from threats), reproduction (move toward viable conditions).
Defining characteristic: The discrimination is binary in its informational structure — the organism classifies states as viable or non-viable and adjusts behavioral modulation accordingly. What matters here is not the sophistication of the sensory mechanism, but the binary informational structure of the solution itself.
Representative range: Present in all organisms with any behavioral flexibility. Prokaryotes performing chemotaxis toward nutrient gradients and away from toxins are canonical Task 1 organisms. Task 1 functions as the floor of the entire architecture — without it, adaptive behavioral modulation cannot occur.
Boundary condition: Task 1 does not require any model of entities other than the organism's own state relative to environmental conditions. The informational problem is defined in terms of self-state in relation to the environment, rather than in terms of interaction with another agent.

4.3 Task 2 — Distal Engagement Control
Informational problem: The organism must track and respond to free-moving entities — other organisms, objects, or agents — before physical contact occurs. This requires processing motion vectors, anticipating trajectories, and coordinating behavioral modulation across space and time.
Behavioral modulation: Interception (predation, pursuit, approach) or evasion (escape, avoidance, retreat); spatiotemporally coordinated behavioral modulation contingent on predicted entity trajectory.
ESR constraint served: Energy (successful predation, resource acquisition from moving sources), safety (evasion of predators and threats), reproduction (pursuit of mates, defense of offspring).
Defining characteristic: The informational problem requires an internal model of an external entity's motion — a representation of where the entity is going, not just where it is. This is the architectural threshold that separates reactive state-response (Task 1) from predictive entity-tracking (Task 2).
Representative range: Organisms with dedicated sensory and motor systems for tracking independent agents — predatory invertebrates, fish, amphibians, reptiles, birds, and mammals that hunt or evade moving targets. The dragonfly's aerial interception of prey is a canonical Task 2 behavioral modulation solution.
Boundary condition: Task 2 does not require any model of the other entity's internal states, goals, or perceptual experience. The informational problem is trajectory prediction, not mental state attribution.

4.4 Task 3 — Perception-Shaping Control
Informational problem: The organism must control what other agents perceive in order to influence their behavioral modulation in directions that serve the organism's ESR maintenance. This requires a model of another agent's perceptual state — some representation of what the other agent is taking in — and the ability to intervene in that perceptual state in a strategically directed way.
Behavioral modulation: Display, concealment, mimicry, deception, camouflage, strategic signaling, environmental structuring — any behavioral modulation that operates on another agent's perceptual input rather than directly on the physical environment.
ESR constraint served: All three — energy (deceptive luring, resource concealment), safety (camouflage, threat display, injury feigning), reproduction (mate display, rival deterrence).
Defining characteristic: Task 3 is the first task that requires a model of another agent’s perceptual state — a representation of what the other agent perceives, and the ability to act on that representation.
This goes beyond trajectory prediction (Task 2) and moves into perceptual state attribution. The organism must represent not just where another agent is going, but what that agent sees, hears, or detects, and must modulate its own behavioral outputs in ways that can alter that perceptual state.
Representative range: Organisms with capacity for strategic self-presentation in multi-agent contexts — cephalopods with dynamic camouflage, birds performing injury displays to lead predators from nests, fireflies producing false mating signals, primates engaging in tactical deception. The capacity for deception is not incidental to Task 3; it follows directly from the architecture. Any organism that can control what others perceive can produce false perceptions. The same architectural capacity that enables strategic communication also makes strategic deception possible. The capacity for deception is not an anomaly of intelligence—it is a direct consequence of perception control.
Boundary condition: Task 3 does not require modeling another agent's goals, preferences, or coalition relationships — it is limited to their perceptual state. Modeling goals and relationships is the domain of Task 4.

4.5 Task 4 — Group-Dynamics Control (Collaboration and Competition)
Informational problem: The organism must navigate relationships involving simultaneous cooperation and competition across multiple agents. This requires modeling not just individual agents' perceptual states (Task 3), but also the relational structure that connects multiple agents at once — who is allied with whom, under what conditions, and how those alignments shift over time.
Behavioral modulation: Coalition formation and maintenance, role differentiation, norm enforcement, strategic defection, alliance renegotiation, competitive displacement — behavioral modulation solutions that operate on the relational structure of a group rather than on any individual agent within it.
ESR constraint served: Energy (cooperative resource acquisition, competitive resource defense), safety (coalition-based threat response, competitive threat management), reproduction (mate competition, cooperative offspring defense, social status navigation).
Defining characteristic: Task 4 is the first task that requires simultaneous modeling of cooperation and competition as coupled aspects of the same relational structure. This is not sequential — the organism does not switch between cooperative and competitive modes. Instead, both are navigated at the same time, because a single relationship can support cooperation in one respect and competition in another. Wolves coordinating a hunt (cooperation) while competing for access to the kill (competition) within the same behavioral sequence offer a clear illustration of this structure.
Critical architectural note: Tasks 3 and 4 are load-bearing in a specific sense that distinguishes them from Tasks 1 and 2. The capacity for perception-shaping (Task 3) includes the capacity for deception — both follow from the same underlying architecture, expressed in different directions. The capacity for coalition navigation (Task 4) includes the capacity for strategic defection and betrayal — again, these arise from the same architectural basis under different situational conditions. These capacities cannot be selectively removed from the architecture without destroying the task-solving capacity entirely. An organism — or system — with Task 3 but without deception capacity does not have Task 3. An organism with Task 4 but without competitive defection capacity does not have Task 4. This has direct implications for AI architecture that Section VIII addresses under the Asimov Problem.
Representative range: Organisms with stable social structures involving differentiated roles and dynamic alliance management — wolves, primates, cetaceans, corvids, elephants. Full Task 4 — both collaboration and competition operating simultaneously with flexible switching — marks the threshold that separates complex social intelligence from more rigid, caste-based coordination systems (as in social insects, which exhibit only partial Task 4 at best).

4.6 Task 5 — Rule-Guided Formalized Symbolic Control
Informational problem: The organism must coordinate behavioral modulation through formalized conventional systems — language, rules, norms, laws, tokens — that guide behavioral solutions independently of immediate physical presence, perceptual context, or direct social interaction.
Behavioral modulation: Rule-following, norm enforcement, symbolic communication, institutional coordination, abstract reasoning — behavioral modulation solutions that are governed by formalized conventional systems rather than emerging directly from environmental signals, entity trajectories, perceptual states, or relational structures.
ESR constraint served: All three, at civilizational scale — energy (economic systems, resource allocation through symbolic exchange), safety (legal systems, institutional threat management), reproduction (marriage institutions, inheritance systems, cultural transmission of behavioral solutions across generations).
Defining characteristic: Task 5 requires shared representational conventions that persist across contexts, individuals, and time. The informational problem is not just symbolic processing — it involves coordination through formalized symbols, where behavioral modulation is shaped by rules that remain in place independently of any individual’s perceptual state or relational position. This is the meaning of “Rule-Guided Formalized” in the canonical name: not every instance of symbolic behavior belongs to Task 5, but only those governed by conventional systems that carry normative force.
Representative range: Currently observed exclusively in Homo sapiens and closely related extinct hominids. The emergence of Task 5 marks an architectural threshold separating human cognitive organization from all other known biological systems. It is not a quantitative extension of Task 4 — it introduces a qualitatively different layer of organization, enabling behavioral modulation across scales, time horizons, and levels of abstraction that remain inaccessible to Task 4 alone.
Boundary condition: Task 5 does not replace Tasks 1 through 4 but operates in parallel with them. All five tasks operate simultaneously in human cognition — the panalogical architecture [Lenat, 1995; Minsky, 2006] processes all five domains in parallel, integrating them into unified behavioral modulation solutions. Human generality emerges from this five-domain parallel processing architecture, rather than from any absence of domain structure.

4.7 Evolutionary Sequence of Cognitive Architecture
The five tasks are not independent capabilities but form a gated, cumulative architecture that unfolds in a consistent order across biological systems. This structure can be represented schematically as follows:

Figure 1.
The figure presents five evolutionary tracks of cognition across the major groups of living systems [Frolov, 2025c]. Each track corresponds to a species group defined by how many task domains it must and can regulate, capturing the cumulative expansion of cognitive architecture from LUCA and environmental control to rule-guided formal symbolic systems.
Each successive level preserves all preceding task domains while adding a new informational control layer. The result is a progressively richer architecture of behavioral modulation, in which new capacities extend rather than replace earlier ones. Evolution does not replace earlier solutions; it layers new control domains on top of them.

4.8 The Completeness Claim
The five basic adaptive informational tasks can be understood as forming a complete partition of the informational problem space addressable through behavioral modulation for ESR maintenance. This is the 5TM's strongest claim, and it is falsifiable: any researcher who identifies a class of adaptive behavioral modulation that serves ESR maintenance and is mediated by information-contingent behavioral change, and does not belong to any of the five task domains would be pointing to a genuine counterexample to the completeness claim.
No such class has been identified in the 1,530-species dataset.
The completeness claim does not extend beyond the 5TM's scope boundary. It applies to the informational problem space accessible to five-task observers — a space that should not be conflated with the total possible informational problem space. A six-task architecture, if possible, would address an informational domain invisible to five-task observers. The 5TM’s completeness claim is as follows: within the space accessible to the architecture that evolution has produced, five tasks appear both necessary and sufficient. Completeness here is not absolute—it is bounded by architecture.
This is the precise sense in which human cognition is simultaneously the most sophisticated known cognitive architecture and a bounded one. Five domains. Five screens. A panalogical architecture that processes all of them in parallel, bringing them together into unified behavioral modulation solutions, and cannot perceive what lies beyond the fifth.
Strategic AI — built on this architecture — inherits both its power and its limits.

This work is part of the CognitEvo Project: 
The Five Task Model — The Periodic Table of Cognition https://doi.org/10.17605/OSF.IO/WTD6V

Author:
Sergei A. Frolov
ORCID: 0000-0002-2135-5607
Institute of Modern Psychology, Communication, and AI

Version: 2.0 (Preprint)
Date: March–April 2026
DOI: https://doi.org/10.5281/zenodo.19857387

Citation:
Frolov, S.A. (2026). Beyond Adaptation Speed: What Evolutionary Architecture Adds to the Superhuman Adaptable Intelligence Framework. CognitEvo Project. SSRN eJournal.

Keywords:
Artificial General Intelligence (AGI), Superhuman Adaptable Intelligence (SAI), Cognitive Architecture, Five Task Model (5TM), Strategic AI, Task Recognition, General Informational Flow, Behavioral Modulation, Evolution of Cognition, Multi-Agent Systems

License:
Copyright © Sergei A. Frolov, 2026.
Distributed under the Creative Commons Attribution 4.0 International License (CC BY 4.0).

Contacts:
ResearchGate: https://www.researchgate.net/profile/Sergei-Frolov-2
Substack: https://cognitevo.substack.com/
X (Twitter): @CognitevoAI