safety & efficiency evals: 44 curated resources
safety & efficiency evals
44 resources · page 2 of 2
Announcing Inspect Evals (UK AISI)
UK AISI announcement of Inspect Evals, an open-source collection of LLM evaluations.
ChinAI #315: Abandoned? Checking in on Three Key AI Safety Benchmarks
ChinAI newsletter issue checking whether three Chinese AI safety benchmarks are still maintained.
ChinAI #351: CAICT launches 2026 AI Safety Evaluations
ChinAI newsletter issue 351 on CAICT's 2026 AI safety evaluations and lessons from its 2025 assessments.
FLI AI Safety Index Summer 2026 (PDF)
Future of Life Institute AI Safety Index Summer 2026 full report (PDF) with company grades by domain.
How Some of China's Top AI Thinkers Built Their Own AI Safety Institute (Carnegie)
by Carnegie Endowment for International Peace
Carnegie Endowment analysis of the China AI Safety and Development Association and its role in Chinese AI governance.
Wikipedia: AI safety institutes
by Wikipedia (Wikimedia Foundation)
Wikipedia article on AI safety institutes: state-backed bodies that evaluate AI safety, listed by country.
Accuracy incentives and hallucinations
by Adam Tauman Kalai et al. (OpenAI)
Nature paper (2026) showing that accuracy-only evaluation of LLMs incentivises hallucination.
Inspect India Evals
by Abhishek Kumar Singh et al.
arXiv paper (2026) introducing Inspect India Evals, an open framework of Indian benchmarks with a fairness index.
Korea AISI 5th AI Safety Forecast Report
by Korea AISI
Korea AI Safety Institute 5th AI Safety Forecast Report covering May–July 2026.
LabSafety Bench
by Yujun Zhou et al. (University of Notre Dame)
Nature Machine Intelligence paper introducing LabSafety Bench for LLM safety risks in scientific labs.
Safetywashing
by Richard Ren et al. (Center for AI Safety)
arXiv paper (2024) "Safetywashing", testing whether AI safety benchmarks measure safety or general capability.
The Leaderboard Illusion
by Shivalika Singh et al. (Cohere)
arXiv paper (2025) "The Leaderboard Illusion", auditing how private testing and data access distort Chatbot Arena rankings.