We find that transferability of adversarial attacks can be substantially improved by using a surrogate model which is not fully trained.

We show that language model based contextualized knowledge graph completion techniques are an effective method for extracting relations and events from noisy user-generated wetlab protocols.

We demonstrate that, contrary to prior belief, a hill climbing approach using an epsilon-greedy strategy is highly effective for forming effective divisions of individuals into groups based on arbitrary fitness scores, outperforming previous approaches in both efficiency and results.