2026
Bloomdow
Autonomous x-risk evaluation framework for large language models
Built Bloomdow to run end-to-end existential-risk behavioural evaluations for LLMs. We solved the mode collapse issue present in Anthropic's December 2025 SOTA workflow by generating synthetic data from multiple seeds, enforcing diversity with cosine dissimilarity between seeds, and then applying the reward model and synthetic data generation stages as before for stronger coverage and more reliable evaluation outcomes.
