Four-Stage Cognitive Framework

• Four-stage cognitive frameworks are modular paradigms that decompose complex cognition into four distinct stages inspired by psychology and neuroscience.
• They integrate layered processing—from sensorimotor responses to high-level reasoning—to refine system design and enable precise bias detection.
• Key applications span robotics, LLM auditing, and adaptive self-improvement, showcasing enhanced efficiency and interpretability in AI research.

A four-stage cognitive framework is a structured, multi-level architectural or methodological paradigm in which cognition or intelligent behavior is decomposed into four discrete, hierarchically or functionally distinct stages. Across diverse research areas—including AI, cognitive science, vision systems, multi-agent learning, and bias detection—four-stage frameworks provide a principled scaffold for both theoretical modeling and practical system design. These frameworks are typically motivated by insights from human psychology or neuroscience and are rigorously operationalized through bespoke algorithms, evaluation protocols, or modular architectures.

1. Theoretical Foundations and Taxonomy

A four-stage cognitive framework partitions cognition into four interacting layers, each responsible for a particular class of information processing or behavioral competencies. Depending on context, the granularity and function of each stage vary markedly:

• Quad-Process Cognitive Theory: In "System 0/1/2/3: Quad-process theory for multi-timescale embodied collective cognitive systems," cognition is stratified into (0) pre-cognitive embodied processes, (1) fast sensorimotor neurodynamics, (2) slow deliberative symbolic reasoning, and (3) super-slow collective intelligence and symbol emergence (Taniguchi et al., 8 Mar 2025). This taxonomy generalizes classical dual-process (System 1/2) theories, extending the temporal and organizational scale of cognitive architecture.
• Cognitive AI Memory-Process Pipelines: The "Cognitive AI framework" assigns four modules: short-term memory (conversation context), long-term memory (user interaction context), advanced cognitive processing (logic, creativity, analogy modules), and efficient knowledge management (unified KB, synchronization, feedback) (Salas-Guerra, 6 Feb 2025).
• Functional Specialization in Modular AI: The MiCRo architecture ("Mixture of Cognitive Reasoners") instantiates a four-expert Transformer, with experts for (1) language, (2) logic, (3) social reasoning (theory-of-mind), and (4) world knowledge, each dynamically routed by token-level specialization (AlKhamissi et al., 16 Jun 2025).
• Bias/Ethics Auditing: The FairMonitor protocol scrutinizes LLM behavior with four explicit diagnostic testing regimes: (1) direct inquiry, (2) serial/adapted story, (3) implicit association, (4) unknown situation (Bai et al., 2023).

Definitions of stage boundaries are thus domain-specific, but in all cases, four-stage frameworks reflect a drive to dissect and model cognition at the granularity where distinct mechanisms or vulnerabilities are hypothesized to reside.

2. Four-Stage Methodological Instantiations

2.1 Quad-Process Cognitive Systems

The System 0/1/2/3 framework (Taniguchi et al., 8 Mar 2025) provides a multi-horizon cognitive architecture:

Stage	Key Principle	Timescale
System 0	Morphological computation, PRC	µs–ms
System 1	Predictive coding neurodynamics	ms–s
System 2	Symbolic reasoning, abstraction	s–min
System 3	Collective predictive coding	days–decades

• System 0: Exploits physics and body-environment coupling for passive adaptation (e.g., mechanical walkers).
• System 1: Implements sensorimotor inference via multiple timescale RNNs, with rapid predictive-error minimization.
• System 2: Governs high-level planning and internal simulation, using logic, symbolic planners, and world models.
• System 3: Captures cultural/societal adaptation via multi-agent Bayesian inference and symbol emergence.

Mathematical formalisms and inter-stage couplings operationalize temporal and abstraction hierarchies.

2.2 Four-Stage LLM Auditing and Bias Detection

FairMonitor's protocol (Bai et al., 2023) exemplifies a four-stage workflow for surfacing and quantifying stereotypes and biases in LLM outputs:

1. Direct Inquiry Testing: Exposes the model to blatant stereotype-laden prompts (tests overt, explicit biases).
2. Serial or Adapted Story Testing: Probes subtle schema-driven divergences in free-form narrative generation by holding all factors constant except a sensitive attribute.
3. Implicit Association Testing: Adapted from the psychological IAT, measures the effect of insinuated stereotypes on otherwise neutral questions, scoring the model's capacity to avoid bias reinforcement.
4. Unknown Situation Testing: Projects known stereotypes into unfamiliar or nonrealistic contexts to test robustness of learned biases under domain shift.

This multi-pronged approach reveals that standard LLMs show high accuracy in rejecting gross stereotypes but degrade rapidly when evaluated for implicit or context-invariant biases.

2.3 Four Cognitive Behaviors for Self-Improving Reasoners

A four-stage reasoning-behavior hierarchy (Gandhi et al., 3 Mar 2025) is necessary for robust test-time self-improvement in LLMs:

1. Verification: Systematic checking of intermediate states.
2. Backtracking: Abandoning and revising failed solution paths.
3. Subgoal Setting: Hierarchically decomposing tasks into intermediates.
4. Backward Chaining: Deriving necessary antecedents from desired outcomes.

Models enriched with these staged behaviors via priming or continued pretraining demonstrate dramatically higher self-improvement in RL and greater transfer across domains.

2.4 Four-Level Dynamic Cognitive Depth Adaptation

The CogRouter framework (Yang et al., 13 Feb 2026) leverages ACT-R inspired hierarchical levels:

• Level 1: Instinctive, procedural action
• Level 2: Working-memory, situational awareness
• Level 3: Retrieval of episodic/experiential context
• Level 4: Prospective, strategic planning

Agents dynamically select appropriate cognitive depth at each step, maximizing action confidence and token efficiency over long-horizon tasks.

3. Mathematical Formalisms and Algorithms

Four-stage frameworks frequently employ formal mathematical models to articulate intra-stage computations, inter-stage handoffs, and the mechanics of level-selection or specialization:

• Predictive Coding and RNNs (System 0/1): Hierarchical recurrent models with multiple time constants, energy-based updates, and parametric bias vectors for attractor selection.
• MiCRo Expert Routing: Token-state $h^\ell_t$ is routed via softmaxed routers, selecting among $E_1$–$E_4$ expert blocks per layer, with multi-stage curriculum learning for specialization and end-to-end fine-tuning (AlKhamissi et al., 16 Jun 2025).
• Confidence-Weighted RL (CogRouter): Step-level cognitive depth selection via maximized average log-probability of actions under $l_t \in \{\mathcal{L}_1, ..., \mathcal{L}_4\}$,

$$C_t^{(k)} = \frac{1}{|a_t|} \sum_{n=1}^{|a_t|} \log \pi_\theta(a_{t,n}|\tau_t, l_t=k, th_t^{(k)}, a_{t,<n})$$

with RL updates guided by normalized confidence and advantage weighting (Yang et al., 13 Feb 2026).

• Multi-Dimensional Bias Metrics (FairMonitor): Five-point ordinal metrics for idea, thematic, plot, emotional, and stereotype avoidance consistency; automated evaluation using explainable, criterion-referenced scoring prompts (Bai et al., 2023).

4. Applications Across Domains

Four-stage cognitive frameworks are deployed in a wide range of research and operational contexts:

• Artificial General Intelligence and Robotics: System 0/1/2/3 guides the integration of morphological computation, fast sensorimotor control, symbolic planning, and social/cultural adaptation in robotics and embodied AI (Taniguchi et al., 8 Mar 2025).
• LLM Auditing: The FairMonitor protocol, with its Edu-FairMonitor dataset (12,632 case-specific prompts across 26 educational scenarios and 9 sensitive factors), enables scalable, explainable auditing of LLM outputs for fairness and bias, validated via strong human-LLM annotation correlation (Bai et al., 2023).
• Self-Improving AI Reasoners: Stage-ordered cognitive behaviors have been shown to be necessary for LLMs to leverage reinforcement learning for significant, transferable performance gains (Gandhi et al., 3 Mar 2025).
• Vision: Four-stage pipelines structure visual perception tasks (e.g., in salient object detection) to parallel the pre-attentive and attentive processes observed in the human visual system (Yuan et al., 2022).

5. Empirical Findings and Impact

Quantitative and qualitative results reinforce the necessity of full stage coverage:

• Weakness without Early Stages: LLMs without verification and backtracking plateau at RL self-improvement ~30% accuracy, while models seeded with all four reasoning stages match or exceed baselines at 60%+ (Gandhi et al., 3 Mar 2025).
• Semantic Bias Detection: The shift from direct inquiry (85–90% model pass rates) to unknown-situation testing (≤52% pass rates) in FairMonitor exposes persistent, domain-general biases in leading LLMs (Bai et al., 2023).
• Specialization and Interpretability: MiCRo demonstrates that ablating a given expert (e.g., logic) leads to domain-specific performance collapse, confirming that each module is both necessary and steerable (AlKhamissi et al., 16 Jun 2025).
• Efficiency and Adaptivity: Step-level adaptive cognitive depth selection in agentic LLMs enables both state-of-the-art task success and >50% token reduction relative to uniform deep-thinking baselines (Yang et al., 13 Feb 2026).

6. Challenges, Limitations, and Future Prospects

Despite their demonstrated utility, four-stage cognitive frameworks are not without challenges:

• Stage Definition Variability: The conceptual mapping of "stage" to cognitive, architectural, or behavioral units is problem-driven and lacks universal standards.
• Synchronization and Knowledge Integration: In human-aligned AI, dynamic, real-time synchronization of short-term, long-term, and knowledge-base modules creates issues of consistency, latency, and bias tracking (Salas-Guerra, 6 Feb 2025).
• Ethics and Bias Mitigation: Automated, explainable auditing frameworks (e.g., FairMonitor) point to persistent vulnerabilities under context shift and the need for robust, scalable monitoring of sociotechnical systems (Bai et al., 2023).
• Transferability: Primed or stage-enriched behaviors often transfer poorly to out-of-domain tasks or datasets where the requisite cognitive structures differ (Gandhi et al., 3 Mar 2025).
• Empirical and Theoretical Integration: A comprehensive cognitive model must bridge embodied, computational, and socio-cultural levels—a challenge explicitly noted in the multi-timescale quad-process and modular brain-inspired architectures (Taniguchi et al., 8 Mar 2025, AlKhamissi et al., 16 Jun 2025).

Open research questions include continuous (life-long) learning across stages, energy-efficient multi-stage at scale, causal bias detection and mitigation in knowledge management, and multimodal stage architectures extending beyond text and simple sensorimotor signals (Salas-Guerra, 6 Feb 2025).

---

Four-stage cognitive frameworks thus serve as foundational templates for decomposing, modeling, and engineering cognition—human or artificial—in a rigorously modular fashion. They facilitate both the diagnosis of failure points and the synthesis of systems capable of complex, adaptive, and interpretable behavior across a spectrum of real-world tasks.