Draft-KV: Learning Useful Latent Communication Between Language Models

Draft-KV. Learning Useful Latent Communication Between Language Models. By Linquan Wu and nine others, from the City University of Hong Kong, the University of Science and Technology of China, and four other institutions. Published on the twenty-eighth of September. Its code repository is worked out: about forty-five source files, with a readme and a licence.

Latent communication passes internal states between language models instead of decoded text. The catch is that a more accurate receiver does not show the receiver used what the message said. The authors test this directly. Across five pairings of method and dataset, replacing each message with one from an unrelated question changes accuracy by at most zero point six points, even where communication adds fifteen point four four points over the receiver alone. The interface can supply the gain while making the sharing model dispensable.

Draft-KV sends something tied to the question at hand: the key-value states the sharer forms while drafting an answer to the current question. If the message is going to help, it should be because of what it contains.

Linear projections place those states in a side memory, which the receiver reads through a gated attention branch. Training is progressive. It moves from reconstructing the message toward supervising the answer, under a guard against harm from mismatched messages. Both models stay frozen. The interface trains one point zero five million parameters, three hundred and forty-eight times fewer than C2C.

With Qwen3-8B as the sharer, a frozen Qwen2.5-0.5B-Instruct receiver reaches seventy-eight point zero four percent on MMLU-Redux. On its own it scores thirty-seven point four five percent, and with reassigned messages thirty-six point four. That last figure is the control that matters: hand the receiver the wrong message, and it falls back to roughly its solo level.

Holding the interface size fixed, scaling the sharer from zero point six billion to eight billion parameters raises accuracy from forty-six point one one percent to seventy-eight point zero four. Communication also transfers to held-out tasks, and can exceed both models when each holds different evidence.

A latent channel deserves the name only if the content moving through it matters. Draft-KV makes its case with two tests pointing the same way: swap the message and the gain disappears, strengthen the sender and the gain grows.