GPT-5.2 outperforms GPT-4o in team assessment for cybersecurity exercises
Clustering and LLMs automate grading of team problem-solving in tabletop exercises.
Deep Dive
Researchers compared clustering and LLMs (GPT-4o, GPT-5.2) for assessing 81 participants across two countries in tabletop exercises. Clustering proved valid and reliable with low computational cost, grouping teams by similar approaches. GPT-5.2 showed considerably lower error than GPT-4o against instructor rubrics. Both methods are now integrated into the open-source INJECT platform.
Key Points
- Clustering method groups teams by similar task approaches, enabling fast, targeted feedback with low computational overhead.
- GPT-5.2 achieved considerably lower error than GPT-4o when scoring team communication against instructor rubrics.
- Both methods are now integrated into the open-source INJECT platform for scalable, automated team assessment.
Why It Matters
Automated team assessment in cybersecurity exercises saves instructors time and provides faster, more scalable feedback to learners.