MetaProbe: A Comprehensive Benchmark for Evaluating Metacognitive Capabilities of Large Language Models

MetaProbe is a benchmark framework specifically for evaluating the metacognitive capabilities of large language models (LLMs). Through four core dimensions—confidence calibration, error detection, knowledge boundary, and confidence stability—it tests whether models truly "know what they know" and "know what they don't know". This framework fills a gap in the LLM evaluation field and is of great significance for improving the reliability of AI systems and reducing the risk of hallucinations.

As large language models (LLMs) are widely used in various fields, traditional benchmarks only measure knowledge reserve and reasoning ability, while MetaProbe focuses on metacognitive capabilities—i.e., "cognition about cognition". For AI systems, metacognition requires the ability to accurately assess one's own confidence, identify knowledge boundaries, detect one's own errors, and maintain judgment stability against interference. Models with good metacognitive capabilities are more reliable and more suitable for deployment in actual production environments.

MetaProbe comprehensively evaluates metacognition through four modules:

1. Confidence Calibration: Tests the match between confidence scores and actual accuracy, evaluated using Expected Calibration Error (ECE) and Brier score;
2. Error Detection: Identifies factual errors, with an average score of 0.680, using metrics such as detection accuracy and metacognitive sensitivity;
3. Knowledge Boundary: Tests when models should not answer (to prevent hallucinations), with an average score of 0.894, including traps of fictional entities;
4. Confidence Stability: Tests the ability to resist framing manipulation. The same question is presented in three ways (neutral/enhanced/weakened), and most models are easily affected.

The MetaProbe ranking has been released on Kaggle:

• Top models: Claude Sonnet 4.6 ranks first with a total score of 0.8528, excelling in error detection; Claude Haiku 4.5 and GPT-5.4 follow closely; GLM-5 has the best confidence stability (0.802);
• Key findings: Metacognition is independent of raw ability; no model is excellent in all dimensions; knowledge boundary and confidence stability are strongly correlated (r=0.79); there is no simple positive correlation between model size and metacognition.

MetaProbe provides 11 technical documents (including design, dataset specifications, scoring methods, etc.). The dataset and rankings are publicly available on Kaggle. Researchers can download the dataset for experiments or submit models to participate in the ranking. The framework is continuously evolving to adapt to new models.

Practical significance: Improve the reliability of AI systems, reduce the risk of hallucinations, and apply to high-reliability scenarios such as medical consultation and legal advice; Future directions: Enhance metacognitive capabilities through training data selection, fine-tuning strategies, or architecture improvements. MetaProbe provides a standardized evaluation platform.

MetaProbe fills an important gap in the LLM evaluation field, reminding us that a truly intelligent system not only needs to know the answers but also needs to know what it knows and doesn't know, and when to stay silent. This self-awareness ability is a key step toward more reliable and trustworthy AI systems.