Standard decision models output probabilities over a set of options, which is useful for classification but not for users who want a natural-language answer. The new arXiv preprint introduces System X, an LLM decision model that adds a text-reply option to the usual probability output.
The paper's title says System X replies from its own KV cache — the key-value cache that stores attention states during generation. That suggests the model uses its internal representations to produce the verbal reply rather than relying on a separate generation pipeline.
The abstract in the arXiv listing is truncated, so the full method and any experimental results are not available from the announcement. The contribution, as described, is the addition of a text-reply mode to a decision model.