What question did this study set out to answer?

This work aims to explore the contradictions within large language models related to context length and performance.

April 17, 2026Open Access

Mooyeon's Paradox

Key Points

This work aims to explore the contradictions within large language models related to context length and performance.
Defined Mooyeon's Paradox in relation to context expansion in LLMs
Analyzed the impact of context on cognitive performance and economic costs
Considered the implications of existing API solutions on the paradox
Increasing context windows lead to attention dilution and cognitive degradation
Economic costs escalate due to O(N^2) complexity
Current infrastructural changes do not resolve the fundamental issues presented by the paradox

Abstract

The Large Language Model (LLM) industry is currently focused on expanding context windows to facilitate long-term user narratives. However, this approach overlooks the inherent cognitive and economic constraints of the Transformer architecture. This paper defines "Mooyeon’s Paradox": a structural contradiction where increasing context leads to intelligence degradation through attention dilution while simultaneously causing costs to increase as a cumulative series — each new turn bearing the full cost of all prior turns in maintenance costs. We analyze this collapse through two dimensions: Cognitive Collapse, characterized by the "Lost in the Middle" phenomenon, and Economic Collapse, driven by the O (N²) complexity and the stateless nature of APIs. We further demonstrate that recent infrastructure-level responses, such as OpenAI's Responses API (2025), merely shift the computational burden rather than resolve the underlying paradox.

Mooyeon's Paradox

Key Points

Abstract

Cite This Study