Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Position: Truly Self-Improving Agents Require Intrinsic Metacognitive Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Position paper arguing that truly self-improving LLM agents need intrinsic metacognitive learning, proposing a framework of metacognitive knowledge, planning, and evaluation and analyzing existing agents (STAR, Voyager, Generative Agents) through it.
Problem
Current self-improving agents rely on rigid, human-designed meta-level loops that fail to generalize across task domains and do not scale with growing agent capabilities.
Approach
Frames self-improvement through a bi-level metacognitive lens (knowledge, planning, evaluation) and distinguishes intrinsic (agent-controlled) from extrinsic (human-designed) metacognition, arguing intrinsic control is needed for scalable, generalized self-improvement and outlining how to share metacognitive responsibilities.
Key results
- Analysis of case studies (STAR, Voyager, Generative Agents) suggests agents with intrinsic metacognitive functions achieve more sustained progress, though this is a preliminary qualitative observation rather than systematic experiment.
Limitations
- No systematic empirical validation; the claims rest on preliminary case-study observations.
- Acknowledges philosophical tension in defining intrinsic versus extrinsic control when agents are human-bootstrapped.
- Notes open challenges in evaluating and finetuning intrinsic metacognition and in safe oversight.
Experimental setup
Builds on: STAR Voyager Generative Agents Reflexion ReAct
Closest papers by meaning
- B-STaR: Monitoring and Balancing Exploration and Exploitation in Self-Taught Reasoners (ICLR 2025 · similarity 0.86)
- MAGELLAN: Metacognitive predictions of learning progress guide autotelic LLM agents in large goal spaces (ICML 2025 · similarity 0.86)
- Meta$^n$: Recursive Self-Improvement through Emergent Depth (NeurIPS 2026 · similarity 0.86)
- Learning to Continually Learn via Meta-learning Agentic Memory Designs (NeurIPS 2026 · similarity 0.86)
- Adaptive Collaboration with Humans: Metacognitive Policy Optimization for Multi-Agent LLMs with Continual Learning (ICLR 2026 · similarity 0.86)
- Language Models Are Capable of Metacognitive Monitoring and Control of Their Internal Activations (NeurIPS 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.