Ai2 Builds BenchMIRT to Audit LLM Evaluation Benchmarks
Ai2 tool audits whether benchmark questions test intended model abilities such as safety.
The Allen Institute for AI announced BenchMIRT to audit LLM safety and capability evaluations and determine which model abilities their questions actually test. On the BBQ social-bias benchmark the tool showed questions distinguished models more by reasoning ability than by safety. The institute posted the details in a thread that includes a diagram and links to their blog post on the project.
Combined views
14.5K
3 posts, first seen 1d ago
Reactions
3 postsAI 
Ai2@allen_aiDo LLM safety & capability evals measure what they claim to? We built BenchMIRT to audit them + see which model abilities their Qs actually test. On BBQ, a social-bias eval, it found the Qs distinguished models more by reasoning ability than safety. 🧵
- likes: 38
- replies: 3
- bookmarks: 23
- reposts: 5
AI Ai2@allen_aiDo LLM safety & capability evals measure what they claim to? We built BenchMIRT to audit them + see which model abilities their Qs actually test. On BBQ, a social-bias eval, it found the Qs distinguished models more by reasoning ability than safety. 🧵
- likes: 38
- replies: 3
- bookmarks: 23
- reposts: 5
FB Faeze Brahman@faeze_brhLLM benchmarks do NOT always measure what they intend to? Checkout our recent work with my amazing predoctoral student, Malia Morgan where we developed a method based on Item Response Theory to audit & simultaneously capture competing LLM abilities such as safety, and…
- likes: 40
- replies: 2
- bookmarks: 15
- reposts: 2
