Before You 〈think/〉, Monitor: Implementing Flavell's Metacognitive Framework in LLMs

Nick Oh

Keywords: self-refinement, self-verification, metacognition

TL;DR: Adding an explicit metacognitive monitoring phase before generation improves LLM reasoning accuracy by 7-8% whilst making the reasoning process more interpretable, though at the cost of 30% longer inference time.

Abstract: Current self-refinement approaches for LLM reasoning generate solutions first and verify them afterwards -— a sequence that cannot prevent poor initial attempts from cascading into unrecoverable errors, known as the `prefix dominance trap' \citep{luo2025learning}. We implement Flavell's Monitor-Generate-Verify model from the MGV framework \citep{oh2024system}, which introduces metacognitive monitoring \textit{before} generation, assessing task difficulty and retrieving relevant strategies to guide initial solution attempts. On GSM8K, our implementation achieves 75.42\% accuracy compared to 68.44\% for Self-Refinement and 67.07\% for Self-Verification, whilst requiring fewer solution attempts (1.3 vs 2.0). The monitoring phase makes metacognitive processes interpretable through explicit traces of difficulty assessment, strategy selection, and dimensional verification, though at a computational cost of 27-37\% increased inference time. These preliminary results suggest that explicit metacognitive monitoring can enhance both reasoning performance and explainability, providing a path towards more transparent and cognitively-aligned AI systems.

Paper Published: No

Paper Category: Short Paper

Demography: No, I do not identify with any of these affinity groups

Academic: Others

Submission Number: 34