One doc tagged with "weights-analysis"

Analyzing Anthropic's research on Deceptive Alignment, weight-space backdoors, and the mathematical failure of standard RLHF to sanitize poisoned models.