Firehose

This paper proposes a new method for training large language models in open-ended domains, using evolving contexts as in-training supervision to capture task preferences. Practitioners may care about this approach because it can lead to better performance on open-ended tasks.