💬 NLP

How Psychological Learning Paradigms Shaped and Constrained Artificial Intelligence

This paper argues that artificial intelligence's historical limitations stem from inheriting the structural constraints of the psychological learning paradigms that inspired it, and proposes the ReSynth trimodular framework—integrating a cross-cultural perspective on rote learning—to architecturally separate reasoning, purpose, and knowledge as a necessary condition for achieving artificial general intelligence.

Original authors: Alex Anvi Eponon, Ildar Batyrshin, Christian E. Maldonado-Sifuentes, Grigori Sidorov

Published 2026-03-20

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Alex Anvi Eponon, Ildar Batyrshin, Christian E. Maldonado-Sifuentes, Grigori Sidorov

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Idea: Why AI Gets Stuck

Imagine you are teaching a child how to build with LEGO.

Current AI is like a child who has memorized exactly how to build a specific castle because they saw a picture of it once. If you ask them to build a different castle using the same bricks, they get confused. They can't figure out the rules; they just remember the picture.
This paper argues that AI is stuck because we built it using old ideas from psychology that were already flawed. We need to rebuild the "brain" of the AI so it can actually learn to learn, rather than just memorizing answers.

Part 1: The Three "Parents" of AI (And Their Flaws)

The paper traces the history of AI back to three major schools of thought in psychology. Think of these as the "parents" who gave AI its DNA, but they also passed down some bad habits.

1. Behaviorism → Reinforcement Learning (The "Treat and Scold" Parent)

The Old Idea: In psychology, this was the idea that animals (and humans) learn only through rewards (treats) and punishments (scolding). You don't care what's happening inside their head; you just care if they press the lever when they get a cookie.
The AI Version: This became Reinforcement Learning. An AI plays a game, gets points for winning, and loses points for losing. It adjusts its moves to get more points.
The Flaw: The AI learns what to do to get a reward, but it doesn't understand why. It's like a dog that knows pressing a button gives a treat, but doesn't understand electricity or buttons. If the rules of the game change slightly, the dog (AI) is lost because it never learned the underlying logic.

2. Cognitivism → Deep Learning (The "Black Box" Parent)

The Old Idea: This school said, "Wait, the inside of the mind matters!" It focused on how we store memories, pay attention, and organize thoughts.
The AI Version: This became Deep Learning (the tech behind Chatbots and image recognition). These are massive neural networks that try to mimic how our brains organize information.
The Flaw: While these AI systems are smart, they are opaque. Imagine a giant library where all the books are melted together into a single, giant block of wax. You can see the shape of the block, but you can't read a single book inside it. When the AI makes a mistake, we can't fix a specific part of its "brain" without breaking everything else. It's a "black box."

3. Constructivism → Modular AI (The "Builder" Parent)

The Old Idea: This theory says we build new knowledge by stacking it on top of old knowledge. We don't just absorb facts; we construct understanding.
The AI Version: This led to ideas like "Curriculum Learning," where AI learns easy things first, then hard things.
The Flaw: We know that we should build knowledge, but we don't have a clear blueprint for how to do it mechanically. It's like telling a builder, "Build a house," but not giving them a hammer or a blueprint. The AI accumulates data but can't reliably combine it to solve new problems.

Part 2: The Missing Piece (The "Rote Learning" Misunderstanding)

The paper points out a cultural misunderstanding about Rote Learning (memorization).

The Western View: In the West, we often think memorization is "bad" or "mindless." We think you either memorize facts (bad) or you understand concepts (good). We treat memorization like a hard drive: just a pile of disconnected files.
The Eastern View: In many Eastern traditions (like Chinese education), memorization is seen as a necessary first step. You memorize the pieces of a puzzle first. You don't understand the whole picture yet, but you have the pieces organized and ready. Then, you use those pieces to build understanding.
The AI Lesson: Current AI skips the "organized pieces" stage. It tries to jump straight to understanding without having a structured library of facts to work with. The paper suggests we need to teach AI to "memorize" in a structured way first, so it has the raw materials to build understanding later.

Part 3: The Solution – "ReSynth" (The New Blueprint)

The authors propose a new framework called ReSynth. They argue that current AI fails because it mixes three different jobs into one giant, messy brain.

Imagine a Construction Company.

Current AI: The CEO, the Architect, and the Bricklayer are all the same person, and they are all shouting over each other in the same room. If the CEO changes their mind, the Bricklayer forgets how to lay bricks.
ReSynth: They separate these roles into three distinct, independent departments.

1. The Intellect (The Architect)

What it does: This is the pure reasoning engine. It doesn't care about what it's building; it just knows how to break things down and put them back together.
The Metaphor: It's like a universal toolset (a hammer, a saw, a screwdriver). It doesn't have a personality; it just knows how to use tools. Because it's separate, it can be used to build a house, a bridge, or a boat without needing to be retrained.

2. The Identity (The CEO)

What it does: This provides the purpose. It decides why we are building and what to focus on.
The Metaphor: This is the boss who says, "We are building a hospital, not a prison." It guides the Architect. If you change the Identity, the AI changes its goals, but it doesn't lose its ability to reason.

3. The Memory (The Warehouse)

What it does: This stores the facts, but not as a messy pile. It stores them as structured signatures (like organized puzzle pieces).
The Metaphor: Instead of a chaotic junk drawer, this is a library where every book is on a specific shelf with a clear label. When the Architect needs a brick, the Warehouse hands it over instantly without the Architect having to dig through the whole building.

Why This Matters

By separating these three parts, the AI can:

Learn new things without forgetting old things (solving "Catastrophic Forgetting").
Adapt to new situations because the "Architect" can apply its tools to new problems immediately.
Be transparent because we can look at the "Warehouse" to see exactly what facts it knows, rather than guessing inside a black box.

The Bottom Line

The paper concludes that to create truly smart, adaptable AI (Artificial General Intelligence), we can't just make the current systems bigger or faster. We have to change the architecture. We need to stop mixing up "thinking," "wanting," and "remembering" into one blob. Instead, we need to build a system where these three functions work together like a well-organized team, each doing its own job perfectly.

Based on the provided paper, here is a detailed technical summary of "How Psychological Learning Paradigms Shaped and Constrained Artificial Intelligence."

1. Problem Statement

The paper identifies a fundamental structural failure in contemporary Artificial Intelligence (AI): the inability of current systems to adapt to genuinely novel situations without external intervention. While AI models can be retrained or fine-tuned, they lack the architectural capacity for genuine generalization and systematic reasoning.

The authors argue that this is not an engineering oversight but a structural consequence of the psychological learning theories that inspired AI paradigms. Current AI architectures entangle three distinct cognitive functions—Reasoning (Intellect), Purpose (Identity), and Knowledge (Memory)—into shared parameter spaces. This entanglement leads to:

Catastrophic Forgetting: Inability to learn new tasks without overwriting old knowledge.
Opacity: Representations are statistical compressions rather than interpretable, structured knowledge.
Lack of Systematicity: Models fail at compositional reasoning (e.g., chaining simple problems) because systematic behavior is an accidental property of the training data rather than an architectural necessity.
Cultural Bias: The Western reduction of "rote learning" to passive storage has deprived AI of a structured pre-learning phase essential for building understanding.

2. Methodology: Genealogical Analysis and Architectural Critique

The paper employs a multi-faceted methodological approach:

Genealogical Tracing: The authors map the direct lineage from psychological learning theories to specific AI methodologies, analyzing the "structural commitments" inherited by each.
- Behaviorism $\rightarrow$ Reinforcement Learning (RL): Inherited the exclusion of internal representational structure.
- Cognitivism $\rightarrow$ Deep Learning: Inherited the opacity of internal representations and the inability to perform principled updates.
- Constructivism $\rightarrow$ Integrative/Compositional AI: Inherited the lack of formal operators for how knowledge is constructed.
Cross-Cultural Reappraisal: The paper challenges the Western pedagogical view of rote learning (memorization without understanding). It contrasts this with the Eastern conception (specifically Chinese/East Asian), where memorization is a structured, multi-phase precursor to understanding (Repetition $\rightarrow$ Memorization $\rightarrow$ Understanding $\rightarrow$ Practice $\rightarrow$ Reviewing).
Theoretical Critique: Drawing on the Systematicity Debate (Fodor & Pylyshyn) and Aizawa's critique, the paper argues that neither connectionism (neural networks) nor classicism (symbolic systems) can make systematic behavior a necessary consequence of their architecture without additional, non-entailed hypotheses.
Proposed Framework: Based on these diagnoses, the authors propose a new architectural framework called ReSynth.

3. Key Contributions

The paper's primary contribution is the ReSynth (Reasoning Synthesis) framework, a trimodular architecture that decouples cognitive functions to ensure systematic behavior is a design necessity.

A. The Trimodular Architecture

ReSynth separates the system into three architecturally independent components:

The Intellect (Reasoning):
- A domain-agnostic engine containing universal operators: Decomposition (revealing constraints) and Recomposition (assembling configurations).
- It carries no fixed identity or purpose, allowing it to be applied across different domains and knowledge bases without modification.
The Identity (Purpose):
- Provides direction, impulses, and attention mechanisms.
- It determines why the Intellect reasons and what constraints it prioritizes.
- Unlike current alignment methods (e.g., RLHF) which correct outputs post-hoc, Identity operates upstream, shaping the knowledge structure during generation.
The Memory (Knowledge):
- Stores structured constraint-to-operator signatures rather than raw data or compressed statistical correlations.
- Distinguishes between Active Memory (real-time workspace) and Passive Memory (expansive repository of prior signatures).
- Integrates the Eastern conception of rote learning: Passive memory acts as a structured repository of "fragments" (acquired via repetition) that the Intellect later decomposes and recomposes into understanding.

B. The Separation Thesis

The core theoretical claim is that adaptability requires the physical separation of Reasoning, Purpose, and Knowledge.

In current monolithic models, updating one function (e.g., learning new knowledge) inevitably degrades others (catastrophic forgetting).
In ReSynth, components can be updated independently: Memory can expand additively, Identity can be reconfigured, and the Intellect remains universal.

4. Results and Evidence

While the paper is primarily theoretical, it synthesizes existing empirical evidence to support its diagnosis and proposed solution:

Catastrophic Forgetting: Cites Kirkpatrick et al. [27] showing standard neural networks fail at sequential tasks due to entangled parameters. Conversely, modular architectures (Ellefsen et al. [28]) mitigate this by separating functionality.
Compositional Failure: Cites Mirzadeh et al. [30] and Li et al. [31] demonstrating that even advanced Large Language Models (LLMs) fail at simple compositional math tasks due to missing intermediate memory and activation drift.
Systematicity Failure: References the FORGE evaluation framework [38], showing that LLM accuracy drops dramatically on formally verified novel combinations, proving that benchmark success often reflects memorization rather than genuine generalization.
Modular Success: Cites ICLR 2025 reviews [39] and Progressive Neural Networks [40] showing that modular networks achieve superior sample efficiency and compositional learning compared to monolithic counterparts.

5. Significance

Paradigm Shift: The paper argues that the path to Artificial General Intelligence (AGI) requires moving beyond "better engineering" of current paradigms toward a fundamental architectural reorganization.
Cultural Insight: It highlights a critical blind spot in AI research: the dismissal of Eastern pedagogical traditions regarding rote learning. By reinterpreting memorization as a structured, rule-governed phase, the paper offers a novel mechanism for how AI can acquire the "raw material" for understanding.
Systematicity as a Necessity: By separating Intellect, Identity, and Memory, ReSynth aims to make systematic behavior a necessary consequence of the architecture, rather than an accidental byproduct of training data distribution.
Future Direction: The paper sets the stage for future empirical validation of ReSynth, specifically in syntactic parsing and other domains requiring compositional reasoning, suggesting that the "psychology-to-AI" pipeline must be re-evaluated to include cultural diversity and structural modularity.