Dual-Process Cognition: Fast & Slow Systems

• Dual-process theories of cognition are frameworks that explain intelligent behavior via fast, automatic System 1 and slow, deliberate System 2 processes.
• Formal models, including discrete architectures, mixture policies, and geometric and compression-based methods, rigorously quantify the interactions between the two systems.
• Applications in AI, such as LLM decision routing and adaptive web agents, leverage dual-process architectures to balance computational speed with precision.

Dual-process theories of cognition posit that intelligent behavior emerges from the interaction of two distinct, yet interdependent, cognitive systems. System 1 operates rapidly, automatically, and often unconsciously, using learned associations, heuristics, or implicit world models to make efficient judgments under time or resource constraints. In contrast, System 2 is slower, resource-intensive, and fully conscious, leveraging deliberative, symbolic, or multi-step reasoning for complex problem-solving, carefully structured evaluation, and rule-based inference. This dichotomy, originating in cognitive psychology and formalized by Kahneman and colleagues, has been rigorously extended by recent research to formal models, large-scale neural architectures, geometric frameworks, and applications in both human and artificial intelligence.

1. Formal Models of Dual-Process Cognition

Mathematical formalizations of dual-process theories have been developed at multiple theoretical levels:

• Discrete Architectures: System 1 and System 2 are instantiated as separate modules: fast, pattern-matching routines (e.g., single-pass neural decisions) vs. explicit, multi-step symbolic or chain-of-thought pipelines (Du et al., 17 Aug 2025).
• Mixture Policies: In agent frameworks, the overall policy $\pi_\theta$ can be represented as a convex combination of a System 1 policy $\pi_1$ (stateless, reactive, often learned by imitation) and a System 2 policy $\pi_2$ (stateful, deliberative, trained via reinforcement or behavior cloning), with a context-dependent gating variable $\lambda_t$ (Liu et al., 7 Aug 2025).

$$p_\theta(a_t|s_{1:t}, a_{1:t-1}, g) = \lambda_t \cdot \pi_1(a_t|s_t,g) + (1-\lambda_t) \cdot \pi_2(a_t|s_t, h_t, g)$$

• Geometric Approaches: Dual process phenomena arise as different time-scale regimes in the gradient flow of a scalar cognitive potential $\Phi(x)$ on a Riemannian manifold endowed with an anisotropic metric $G(x)$ (Ale, 13 Dec 2025). Fast dynamics evolve along directions with low computational cost, while slow deliberation aligns with high-drag, high-cost directions:

$$\dot{h} = -\nabla_h \Phi(h,c) \qquad (O(1) \textrm{ time}), \qquad \dot{c} = -\epsilon^2 \nabla_c \Phi(h,c) \qquad (O(\epsilon^2) \textrm{ time})$$

• Compression-Based Theories: Dual-process structure can also be derived from minimum description length (MDL) principles (Moskovitz et al., 2022). The controlled subsystem $\pi$ (System 2) flexibly adapts to maximize expected reward with a KL penalty to the reference policy $\pi_0$ (System 1), which is optimized for simplicity and compressed representation:

$\pi_1$0

• Inter-level Causation and Feedback Control: The Dual-Laws model posits that System 1 (Type 1) processes perform continuous feedback minimization at the neural substrate, while System 2 (Type 2) processes engage in discrete combinatorial selection of symbolic equations or control laws, both operating on feedback error $\pi_1$1 (Ohmura et al., 12 Feb 2026).

2. Operationalization in AI and Cognitive Architectures

LLMs and autonomous agents have adapted dual-process theory in their reasoning pipelines:

• CDR in LLMs: The Cognitive Decision Routing (CDR) framework employs a meta-cognitive router that analyzes features such as correlation strength, semantic boundary crossings, stakeholder multiplicity, and uncertainty to decide between fast (System 1) and slow (System 2) reasoning for each query (Du et al., 17 Aug 2025). This explicit meta-cognition achieves 34% lower computational cost, 23% higher consistency, and 18% better accuracy on expert-level tasks compared to uniform deep reasoning.
• Adaptive Web Agents: In web navigation (WebArena), System 1 is realized by a fast, shallow sub-policy for direct action selection, while System 2 executes chain-of-thought planning using episodic memory. An adaptive gating function $\pi_1$2 decides which system to engage per timestep, yielding strong performance vs. cost trade-offs (e.g., dual-system agents used 75% fewer tokens than deliberation-only baselines without loss of success rate) (Liu et al., 7 Aug 2025).
• Self-Training Dual-Process in LLMs: The CogniDual framework demonstrates that, through self-distillation and targeted fine-tuning, LLMs can internalize System 2 (chain-of-thought, deductive) skills into System 1 (direct, intuitive) responses, accelerating inference without loss of reasoning fidelity (Deng et al., 2024).

3. Extensions and Generalizations

Contemporary work broadens dual-process theory in several significant directions:

• Multi-timescale Models: The System 0/1/2/3 framework introduces “System 0” (embodied, pre-cognitive processes leveraging morphological computation) and “System 3” (collective, supra-individual symbol emergence and social cognition) to create a unified multi-scale timescale model (Taniguchi et al., 8 Mar 2025).
• Unified Geometric Theories: All cognitive processes—including dual-process effects—can emerge from a single geometric principle: Riemannian gradient flow on manifolds with learned metrics and cognitive potentials, enabling fast and slow modes as natural time-scale separations (Ale, 13 Dec 2025).
• Evolutionary Game Dynamics: Evolutionary competition between automatic (System 1) and controlled (System 2) agents under resource constraints can generate dominance, coexistence, bistability, or oscillatory cycles, depending on how environmental feedback affects selection pressures on speed vs. planning (Toupo et al., 2015).

Model/Framework	System 1 (Fast)	System 2 (Slow)
CDR in LLMs (Du et al., 17 Aug 2025)	Single-pass pattern matching	Explicit multi-step reasoning
Unified Theory (Moskovitz et al., 2022)	Compressed, habitual policy ($\pi_1$3)	Flexible, controlled policy ($\pi_1$4)
Geometric (Ale, 13 Dec 2025)	Fast directions in $\pi_1$5	Slow directions in $\pi_1$6
Evolutionary (Toupo et al., 2015)	Automatic type A	Controlled type C
CogniWeb (Liu et al., 7 Aug 2025)	Shallow/one-step policy	Chain-of-thought, episodic planning

4. Empirical and Behavioral Signatures

Meta-analyses and targeted experiments robustly support dual-process architectures:

• Cognitive Performance Gains: Experiments using explicit meta-cognitive routing or context-sensitive gating between System 1 and System 2 consistently demonstrate improved accuracy, lower compute cost, and higher response consistency compared to single-mode baselines (Du et al., 17 Aug 2025, Liu et al., 7 Aug 2025, Manir et al., 10 Sep 2025).
• Bias, Uncertainty, and Meta-Reasoning: Invoking explicit System 2 instructions or cognitive control reduces social and cognitive biases in LLMs (e.g., stereotypical judgments drop by up to 19%) considerably more than chain-of-thought or default prompting (Kamruzzaman et al., 2024).
• Skill Automatization: System 2 training and rehearsal can “compile” complex skills into System 1 responses, paralleling human habit formation (e.g., in math, logic, reasoning), with larger models reaching saturation more rapidly (Deng et al., 2024).
• Behavioral Ecology and Sociality: System 1 is consistently associated with prosocial heuristics, avoidance of harm, and context-driven behavior across cooperation, reciprocity, and moral judgment; System 2 overrides intuition when conflicts or atypical contexts arise (Capraro, 2019).

5. Theoretical and Methodological Implications

Recent advances clarify and extend core tenets of dual-process theory:

• Meta-Cognitive and Contextual Gating: Unlike humans, artificial agents often require explicit, learned gating or meta-cognitive modules to trigger System 2; humans switch implicitly and contextually (Du et al., 17 Aug 2025, Liu et al., 7 Aug 2025).
• Unified Representational Substrate: Geometric and feedback-control theories establish that “fast” and “slow” regimes can arise within a single continuous architecture, without hard modular boundaries (Ale, 13 Dec 2025, Ohmura et al., 12 Feb 2026).
• Description-Length and Compression: MDL-based dual-process models explain adaptive advantages of maintaining both habitual (fast) and flexible (slow) subsystems, predicting patterns of executive control, reinforcement learning, and decision biases in alignment with empirical neural data (Moskovitz et al., 2022).
• Embodiment and Collectivity: The System 0/1/2/3 model frames dual-process theory within a broader hierarchy that spans embodied dynamics, neural/cognitive subsystems, and collective, symbol-level adaptation (Taniguchi et al., 8 Mar 2025).

6. Applications: Creativity, Social Cognition, and Engineering

• Computational Creativity: Dual-process cognitive architectures can decompose creative acts into “exploratory” (divergent, S1), “analytic” (convergent, S2), “tacit” (implicit, S1), and “reflective” (explicit meta-reasoning, S2) modalities; attention/resolution parameters tune the balance between fast associative generation and slow evaluative filtering (Augello et al., 2016, Sowden et al., 2014).
• Theory of Mind and Adaptive Bias: Graph-based and meta-learning agent architectures mirror human dual-process patterns, adapting efficiently to context and replicating canonical reasoning biases via context-gated blending of S1 and S2 outputs (Manir et al., 10 Sep 2025).
• Technical Implementations in LLMs: Explicit dual-process-inspired prompting and dynamic pipeline selection enable both computational efficiency and robustness to bias, with practical relevance to professional QA, policy, and domain-general reasoning tasks (Du et al., 17 Aug 2025, Kamruzzaman et al., 2024).

7. Limitations and Open Problems

Despite substantial progress, key limitations and questions remain:

• Boundary Identification: Automatic meta-cognitive or dynamic “in-flight” control over System 1/2 switching in artificial agents is highly non-trivial; static pre-analysis may misroute cases with emergent complexity (Du et al., 17 Aug 2025).
• Zero-shot Generalization: Current routing and gating functions require domain-specific calibration and may not generalize to entirely novel situations without supervision.
• Representational Transparency: Many architectures lack interpretability, especially after System 2 knowledge is “compiled” into System 1 responses (Deng et al., 2024).
• Cross-Modal and Collective Extensions: Integrating multi-modal inputs (vision, audio, symbolic) and scaling dual-process frameworks to collective and social reasoning remains an open area, addressed in part by quad-process models (Taniguchi et al., 8 Mar 2025).
• Unified Neural Basis: While formal and computational distinctions are clear, neuroimaging suggests overlapping, rather than strictly modular, neural substrates for fast and slow cognition (Ohmura et al., 12 Feb 2026, Sowden et al., 2014).

A plausible implication is that future research will focus on dynamic adaptation and meta-cognitive learning to optimize the fast/slow tradeoff, as well as principled bridging of dual-process theory with geometric, information-theoretic, and embodied frameworks for both human and artificial cognition.