It next optimizes the LLM's system message via prefix tuning on the synthetic task, then uses the system message on realistic, hard-to-optimize tasks. Across three realistic abstractive summarization ...
Some results have been hidden because they may be inaccessible to you
Show inaccessible results