Benchmark RadarAI BENCHMARK PROFILE
StabilityBench
StabilityBench is a benchmark operator that converts single-turn benchmark queries into multi-turn interaction histories with injected user simulations, evaluating LLM performance stability across demographic proxies and sycophantic baits on existing benchmarks.
- Released
- 2026-07-17
- Readiness
- Paper only
- Primary field
- General AI
Why it matters
Static benchmarks may not capture real-world conversational variability; StabilityBench highlights performance instability under realistic multi-turn conditions, motivating more realistic evaluation settings.
Motivation
AI Assistants are increasingly deployed in high-stakes settings, such as healthcare or government services.
Primary resources
Benchmark Radar records only publicly supported details and links back to primary sources for verification.