Understanding dual process cognition via the minimum description length principle

Moskovitz, Ted; Miller, Kevin J.; Sahani, Maneesh; Botvinick, Matthew · 2024 · OpenAlex-citations

DOI: 10.1371/journal.pcbi.1012383

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper addresses the theoretical question of why neural information processing is organized into dual-process systems, a structure prominent in executive control, reward-based learning, and judgment and decision making. While dual-process theories distinguish between fast, automatic mechanisms and slow, effortful ones, existing computational models are typically domain-specific. The authors propose that these diverse phenomena arise from a single underlying computational principle: the need to minimize the description length of an agent’s behavior to enhance generalization. By applying the Minimum Description Length (MDL) principle, the study seeks a normative explanation for dual-process cognition, deriving the division between effortful and automatic behavior from first-principles reasoning about compression and adaptive behavior rather than from empirical fitting. The authors developed a model called Minimum Description Length Control (MDL-C), implemented using recurrent neural networks. The model consists of two policies: a control policy ($\pi$) that maximizes cumulative reward and a default policy ($\pi_0$) that compresses behavioral patterns. The objective function balances reward maximization with the minimization of the complexity of the default policy and the deviation of the control policy from it, quantified using Kullback-Leibler divergence. Variational dropout was used to enforce sparsity in the default policy. The authors tested this model through computer simulations across three domains: executive control (spatial navigation and the Stroop task), reward-based learning, and judgment and decision making. The model was trained to solve tasks and then queried to see if its internal dynamics reproduced canonical dual-process phenomena without being explicitly optimized for those specific behaviors. The simulations demonstrated that MDL-C successfully reproduces key dual-process findings. In executive control tasks, the default policy learned to ignore context cues and rely on frequent stimulus-response associations, mimicking automatic habits. When the control policy was disabled, the agent defaulted to these frequent actions, mirroring behavioral deficits seen in patients with prefrontal cortex damage. In the Stroop task, the model replicated delayed reaction times on incongruent trials, explaining this delay as the computational cost of the control policy overriding the conflicting default policy. The model also showed that policy compression enhances generalization, allowing the agent to learn new tasks faster than unregularized baselines. These results indicate that the bipartite structure of MDL-C naturally gives rise to behaviors characteristic of System 1 (fast, heuristic) and System 2 (slow, deliberative) processing. The significance of this work lies in providing a unified, normative framework for dual-process cognition. By linking the dual-process architecture to the fundamental problem of behavioral generalization via compression, the authors argue that seemingly disparate phenomena across psychology and neuroscience are domain-specific consequences of a single computational principle. This approach bridges gaps between distinct subfields and offers a mechanistic explanation for why the brain might organize information processing into two concurrent systems, balancing the mastery of routine tasks with the flexibility to adapt to new circumstances.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• dual task multitasking