MetaCog-Bench：为大语言模型注入元认知能力的实验框架

MetaCog-Bench is an open-source benchmark framework designed to assess and boost the metacognitive abilities of large language models (LLMs). It focuses on three core mechanisms—intention attribution, self-monitoring, and intentionality anchoring—aiming to address LLMs' lack of self-reflection and cognitive regulation, which are key to human intelligence. This framework marks a shift from performance-focused evaluation to cognitive reliability, supporting the development of more reliable AI systems.

While LLMs excel in knowledge storage and language generation, they lack metacognition—awareness and monitoring of their own thinking processes. This leads to issues like hallucinations and overconfidence, limiting their reliability in high-risk decision-making scenarios. Bridging this gap has become a frontier in AI research.

MetaCog-Bench's evaluation system includes three key mechanisms:

1. Intention Attribution: Assesses the model's ability to infer the intentions behind its own and others' behaviors, enabling more targeted responses.
2. Self-Monitoring: Tests the model's awareness of its cognitive state, such as recognizing knowledge gaps and adjusting confidence.
3. Intentionality Anchoring: Evaluates the model's ability to translate abstract goals into actionable strategies, including task decomposition and plan adjustment.

The framework uses a modular architecture, allowing flexible combination of test modules. Its dataset combines expert-designed cases and LLM-generated adversarial samples. Evaluation metrics go beyond accuracy, including calibration, self-cognitive consistency, and intention alignment to fully capture metacognitive abilities.

Preliminary results show:

• Mainstream LLMs have uneven metacognitive performance.
• Model size doesn't linearly correlate with metacognitive ability (some medium models outperform larger ones).
• Metacognitive skills are domain-dependent, not easily transferable across fields.

MetaCog-Bench provides a standardized tool for researchers to compare models' metacognitive abilities. For developers, it highlights cognitive limitations for safer deployment. Applications include:

• Education: Adjusting teaching strategies based on student understanding.
• Healthcare: Seeking more info when uncertain to avoid misleading advice.
• Scientific research: Identifying knowledge gaps.
• Decision support: Expressing confidence to aid human choices.

MetaCog-Bench represents a shift from 'what LLMs know' to 'how they know'. It lays the groundwork for more reliable AI systems. Future research may lead to AI with true self-reflection—systems that understand their limits and actively improve, becoming intelligent partners rather than just tools.