Alignment (auto-discovered)

OpenAI researchers found that reinforcement learning on realistic scenarios targeting beneficial traits like honesty and corrigibility produced broad improvements across dozens of alignment benchmarks…