As Large Language Models (LLMs) become more capable, researchers and practitioners are exploring whether these systems can reflect on their own reasoning, identify their errors, and improve independently. This emerging area of AI self-reflection goes beyond self-correction. It focuses on a model’s capacity not only to revise its answers but also to reason about its own reasoning process, analyze the underlying causes of errors, and adapt future strategies. Self-reflection is a crucial step towards trustworthy, transparent, and context-adaptive AI systems. It is becoming increasingly important in academic research and industrial applications such as AI quality assurance, error diagnosis, and model alignment.
| Type | Seminar |
| Term | SS 2026 |
| Mentor(s) |
René Reiners |
| Assistant(s) |
Milad Morad |
Seminar objectives
- Understand the theoretical and technical foundations of AI self-reflection.
- Explore current approaches and analyze how they facilitate introspection in LLMs.
- Critically assess the opportunities and risks of self-reflective AI systems.
- Conduct a small-scale experiment or literature-based study to evaluate self-diagnostic model behaviours.
Topics overview (not finalized)
Foundations of AI self-reflection
- Theoretical foundations of metacognition and self-diagnosis in LLMs
- Distinguishing self-reflection from self-correction: conceptual and algorithmic perspectives
Reflective architectures and mechanisms
- Comparing frameworks: Self-Refine, Reflexion Agents, Re-RST, and MetaGPT
- Feedback loops and introspection limits in autonomous reasoning
Evaluation and measurement of reflection
- Comparing evaluation metrics for introspection, like ReflectionBench, SelfEval, and others
- Quantifying confidence and uncertainty calibration in LLMs
Applications and Human-AI symbiosis
- Self-reflective AI for explainability, adaptability, and industrial feedback systems
- Analysis of risks and biases in reflective reasoning (e.g., self-confirming hallucinations)
Format
Each student will independently select one of the listed topics and work on it throughout the semester. The work will include a comprehensive literature review of current research and relevant frameworks, the design and execution of a small experimental study, and the preparation of both a presentation and a written seminar paper (8–10 pages) that summarizes the findings and discusses their broader implications.
Key references
Li, X. et al. (2025). AI Awareness: From Introspection to Control. arXiv:2504.20084.
Loureiro, A. et al. (2025). Advancing Multi-step Mathematical Reasoning in Large Language Models through Multi-layered Self-Reflection with Auto-Prompting. ECML PKDD 2025. arXiv:2506.23888.
Jha, B. et al. (2025). Thinking About Thinking: SAGE-nano’s Inverse Reasoning for Self-Aware Language Models.arXiv:2507.00092.
Bilal, A. et al. (2025). Meta-Thinking in LLMs via Multi-Agent Reinforcement Learning: A Survey. arXiv:2504.14520.
Zhang, X. et al. (2025). Cognition-of-Thought Elicits Social-Aligned Reasoning in Large Language Models.arXiv:2509.23441.
Alamdari, P. & Zare, G. (2025). Self-Reflective Multi-Agent Reinforcement Architecture for Autonomous Recommendation Policy Evolution. ResearchGate Preprint. DOI: 10.13140/RG.2.2.26078.91201.
Qin, H. et al. (2025). R-CHAR: A Metacognition-Driven Framework for Role-Playing in Large Language Models. EMNLP 2025. DOI: 10.18653/v1/2025.emnlp-main.1372.
Chowa, S. et al. (2025). From Language to Action: A Review of Large Language Models as Autonomous Agents and Tool Users. arXiv:2508.17281.
Laukkonen, R. et al. (2025). Contemplative Artificial Intelligence. arXiv:2504.15125.
Shinn, N. et al. (2024). Reflexion: Language Agents with Verbal Reinforcement Learning. arXiv:2303.11366.