As an alternative of rewards, we use new kinds of suggestions, akin to demonstrations (within the above instance, human-written summaries), preferences (judgments about which of two summaries is better), corrections (changes to a summary that may make it better), and extra. We hope that BASALT will be used by anyone
