Sleeper agents, weak-to-strong generalization, supervision with subversion, problems with unsupervised LLM knowledge, patching activations, explainability benchmark, replication evals, DPO analysis