# Truly Self-Improving Agents Require Intrinsic Metacognitive Learning

**Authors:** Tennison Liu, Mihaela van der Schaar  
**arXiv:** [2506.05109](https://arxiv.org/abs/2506.05109)  
**Date:** 2025-06-05  
**Summary page:** https://www.summarizepaper.com/en/paper/2506.05109v1/

## Key points

- Continuous acquisition of new capabilities with minimal supervision is a challenging goal in self-improving agents.
- Effective self-improvement relies on intrinsic metacognitive learning, involving active evaluation, reflection, and adaptation of learning processes.
- A formal framework for intrinsic metacognition includes metacognitive knowledge, metacognitive planning, and metacognitive evaluation.
- Existing self-improving agents heavily rely on extrinsic metacognitive mechanisms designed by humans, limiting scalability and adaptability.
- Optimizing the distribution of metacognitive responsibilities between humans and agents is crucial for sustained self-improvement.
- Language model-based agents demonstrate enhanced adaptability across diverse domains compared to traditional rule-based or reinforcement learning approaches.
- Integrating intrinsic metacognitive learning into agent systems can improve performance and adaptability in dynamic environments.

## Summary

In the realm of self-improving agents, the quest for continuous acquisition of new capabilities with minimal supervision is a challenging one. Current approaches often hit roadblocks due to rigid self-improvement processes that struggle to generalize across task domains and scale effectively as agent capabilities increase. To address these limitations, researchers argue that effective self-improvement hinges on intrinsic metacognitive learning - an agent's innate ability to actively evaluate, reflect on, and adapt its own learning processes. Drawing inspiration from human metacognition, a formal framework comprising three key components is introduced: metacognitive knowledge (self-assessment of capabilities, tasks, and learning strategies), metacognitive planning (deciding what and how to learn), and metacognitive evaluation (reflecting on learning experiences to enhance future learning). Upon analyzing existing self-improving agents, it becomes evident that they heavily rely on extrinsic metacognitive mechanisms which are fixed loops designed by humans, ultimately limiting scalability and adaptability. Delving deeper into each component of the proposed framework reveals that many elements necessary for intrinsic metacognition already exist within current systems. The exploration extends to optimizing the distribution of metacognitive responsibilities between humans and agents while robustly evaluating and enhancing intrinsic metacognitive learning - essential challenges that must be tackled to enable sustained, generalized, and aligned self-improvement in autonomous agents. Furthermore, in the preliminary context provided about intelligent agents leveraging language model technology as their core computational engine for autonomous reasoning and action in real-world tasks , it becomes clear that traditional rule-based or reinforcement learning approaches struggled with generalization across environments. In contrast , language model-based agents harness world knowledge acquired through large-scale training to adapt swiftly to novel tasks across diverse domains. This versatility underscores the potential benefits of integrating intrinsic metacognitive learning into agent systems for enhanced performance and adaptability in dynamic environments.