Before you <think>, monitor: Implementing Flavell's metacognitive framework in LLMs

TL;DR

This work integrates Flavell's metacognitive monitoring into a Monitor-Generate-Verify loop to ground LLM reasoning in upfront task assessment, strategy selection, and structured verification. On GSM8K arithmetic tasks with a single Llama-3.1-8B-Instruct model, the approach achieves higher initial solution quality and fewer refinement attempts (75.42% accuracy, 1.3 attempts) at the cost of longer inference time, suggesting metacognitive monitoring can mitigate the need for extensive iteration. The study analyzes limitations, including explicit vs implicit metacognition, modular architectures, and fundamental reasoning bounds, and outlines directions for embedding model-intrinsic metacognitive knowledge. The results indicate a promising direction for theory-guided, metacognition-driven reasoning in NLP, with potential implications for efficiency and reliability in complex reasoning tasks.

Abstract

Current approaches to enhancing LLM reasoning follows two isolated paradigms: Monitor-Generate methods like Plan-and-Solve (Wang et al., 2023) and SELF-DISCOVER (Zhou et al., 2024) excel at strategic planning but lack mechanisms to verify whether selected strategies succeed; while Generate-Verify approaches like Self-Verification (Weng et al., 2022) and SELF-REFINE (Madaan et al., 2023) iteratively refine outputs but commence generation blindly without task assessment. This separation creates inefficiencies -- strategies fail without feedback, and refinement occurs without strategic grounding. We address this gap by implementing Flavell's cognitive monitoring model (1979) from the broader Monitor-Generate-Verify framework (Oh and Gobet, 2025), operationalising it as a three-phase iterative system. On GSM8K, preliminary results show 75.42% accuracy versus 68.44% for SELF-REFINE and 67.07% for Self-Verification, while requiring fewer attempts (1.3 vs 2.0) at 27-37% increased inference cost. These initial findings suggest upfront monitoring produces higher-quality initial solutions that reduce refinement needs, though evaluation beyond arithmetic reasoning is needed to establish generalisability.