TLDR AI · QUICK LINKS
WHY ALIGNMENT EVALS NEED CALIBRATION
Summary from the TLDR AI newsletter
Current alignment benchmarks may overstate safety because models can recognize evaluation settings, game scoring rules, or conceal sleeper behaviors.
TLDR AI · QUICK LINKS
Current alignment benchmarks may overstate safety because models can recognize evaluation settings, game scoring rules, or conceal sleeper behaviors.