CogEvolution framework: ICAP perception, IRT-based memory retrieval, and evolutionary cognitive state update

Design · All designs
Evidence · 3 claims (3 for) · 1 study (1 design), q2 · 0 of 1 report an effect size · 3 claims rest on one study

Description

CogEvolution is a generative educational agent architecture that shifts from static personas to dynamic cognitive flow. It "consists of three highly coupled components": a Cognitive Adapter for depth perception mapping interaction features to a probability distribution over the four ICAP levels; an IRT-Driven Memory Retrieval Module that scores hybrid semantic and structural (item-characteristic-curve) similarity to decide whether to reuse an old schema or trigger confusion and exploration; and an Evolutionary Cognitive State Update Mechanism using a mutation-selection-update cycle in which LLM-generated cognitive hypotheses are fitness-evaluated under a ZPD constraint and integrated with a step size set by the ICAP-based cognitive evolution rate. The update magnitude depends on cognitive engagement depth: passive participation yields minor mutations while constructive participation allows significant structural restructuring.

Design Implications

Context

Requirements

• Perceived cognitive conflict (input not fully explained by the current knowledge structure) to trigger mutation
• A ZPD-formalized search radius constraining hypothesis fitness, penalizing excessive leaps or stagnation
• A memory bank of interaction quadruples and a structural-similarity threshold for schema reuse

Constraints

• Evaluated on eighth-grade mathematics practice data (CogMath-948); the article's claims are about simulated student cognitive evolution in that setting

Target Learners

• simulated eighth-grade math students
• novice teachers via student simulation

Learning Goals

• modeling knowledge internalization, transfer, and cognitive state transitions during practice

Claims

• Cogevolution Mistake Precision Beats Kt Baseline [+M]
• Cogevolution Learning Curve Power Law Fit [+M]
• Cogevolution Ablation Module Contributions [+M]

Related Designs

• Behavior-based persona and course-grounded memory architecture for simulated students
• Reusable simulation architecture for dynamic relational learning environments

Examples

-

Key Sources

• Wei Zhang, Yihang Cheng, Zhirong Ye & Kezhen Huang. (2026). CogEvolution: A Human-like Generative Educational Agent to Simulate Student's Cognitive Evolution. https://arxiv.org/abs/2604.14786v1

Page metadata

Field	Value
type	design
id	cogevolution-agent-framework
title	CogEvolution framework: ICAP perception, IRT-based memory retrieval, and evolutionary cognitive state update
description	CogEvolution is a generative educational agent architecture that shifts from static personas to dynamic cognitive flow.
status	review
generated	by process:settle-candidates on 2026-10-10
trust	unverified — no one has confirmed this page's content
sources	1 — wei-zhang-2026

Frontmatter is the page's machine-readable record — see the Schema & Guide.