Research map / Large language models / Agents, code and math
Evidence for Limited Metacognition in LLMs
Christopher Ackerman
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use more recent / stronger LLMs | Delegate Game introspection · GPQA, SimpleQA, GPSA, SimpleMC | partial correlation with baseline correctness | Improved |
| switch from factual to reasoning questions | Delegate Game · GPQA vs SimpleQA | introspective ability: P=0.32 | No clear effect |
| switch from multiple-choice to short-answer format | Delegate Game · MC vs short answer | introspective ability: P=0.08 | No clear effect |
| use thinking mode instead of non-thinking | Gemini 2.5 models · Delegate Game entropy correlation | partial correlation with entropy | Hurt |
| raise teammate accuracy in context | Delegate Game · all models | delegation rate | Improved |
| use Delegate Game instead of self-reported confidence (ablation) | Delegate Game · four question sets | relation to surface difficulty cues | Improved |
| use Delegate Game instead of Pass Game (ablation) | Delegate Game · four question sets | relation to objective difficulty: AUC difference 0.07 | Improved |
| decision-only Delegate Game variant (ablation) | Delegate Game · select models | introspective access | Mixed |
| tell model its previous answer was wrong | GPT-4.1, GPT-4o, GPT-4o Mini · Second Chance Game (GPQA, SimpleMC) | change rate lift | Improved |
| use thinking mode / smaller Gemini models | Second Chance Game · GPQA, SimpleMC | change rate lift | No clear effect |
| less exhortatory prompt in Second Chance Game (ablation) | Second Chance Game · prompt variation | change rate | Mixed |
Closest papers by meaning
- Decoupling Metacognition from Cognition: A Framework for Quantifying Metacognitive Ability in LLMs (AAAI 2025 · similarity 0.90)
- LLMs Show No Signs Of Individuated Metacognition (NeurIPS 2026 · similarity 0.89)
- Language Models Are Capable of Metacognitive Monitoring and Control of Their Internal Activations (NeurIPS 2025 · similarity 0.88)
- Position: Principles of Animal Cognition to Improve LLM Evaluations (ICML 2025 · similarity 0.87)
- Tuning-Free Accountable Intervention for LLM Deployment – a Metacognitive Approach (AAAI 2025 · similarity 0.87)
- Metacognitive Capabilities of LLMs: An Exploration in Mathematical Problem Solving (NeurIPS 2024 · similarity 0.87)
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.
