LLMs can invent their own compression - Rajan Agarwal

Co-training a summarizer and generator via RL lets an LLM learn its own text compression scheme, achieving ~9.5% context size with minimal prediction loss.