Use Meta-Tokens for Context Compression

If part of the problem with context compression is that you lose some of the implicit knowledge given by a context with exactly those tokens in exactly that order, and its possible to train LLMs with meta-tokens that represent entire spans of a context window in a single embedding (https://arxiv.org/abs/2509.16278), you should train the summarizer model with a tool that allows it to verbatim insert meta-tokens into the summary (directly bringing over the residual stream values). Now, you can also integrate this into part of the RL process--a model doing long horizon tasks during RL posttraining will learn to experience these enriched compression steps where we're both doing RL on the "reasoner" to get good at working with only summaries in context (could be an independent thing to try without the meta-tokens--though I'd be shocked if this isn't already done) and on the summarizer to get better at moving meta-tokens from the old to the new context.