Share the Symbol, Not the Vector: Training-Free Fact Communication Across Language Models

Share the Symbol, Not the Vector. Training-Free Fact Communication Across Language Models. By Xin Li and seven others, from Nanyang Technological University, Yale University, and one other institution. Published on the fourteenth of October. Its code repository is worked out: about ninety-five source files, with a readme and a licence.

Multi-agent systems let language models solve tasks together by sharing information that each one holds and the others lack. Usually that sharing happens in text. More recent approaches have tried something more direct: exchanging internal model states. The trouble arrives when the two models are built differently. Sharing states across architectures often needs a learned translator sitting between them, or else assumptions about how compatible the two models are. Either way, connecting a new pair of models costs something before a single fact can pass between them.

SYNC, short for Symbol-Indexed Neural Communication, swaps the vector for a symbol. It is a word-indexed interface that connects heterogeneous models without fine-tuning and without a learned translator. The two sides agree only on a shared list of words. The sender picks words from that list and transmits their indices. The receiver turns those indices back into something its own hidden states understand, using directions that belong to the receiver alone. Nothing about the sender's internals ever has to be mapped onto the receiver's.

On the sending side, the sender works from descriptions that only it can see. For each object, it selects words from the shared list and sends their indices, and it does this without seeing the questions that will later be asked. On the receiving side, the receiver writes the described facts into its hidden states using its own word directions: vectors fitted once from its own activations. Where to write matters, so the authors choose an early write layer using four development receivers, and then hold that choice fixed.

The test is built to demand selective use of facts rather than recall of a blob. One attribute is swapped between two objects, and the receiver must answer correctly about both the changed and the unchanged attributes, before and after the swap. That is four answers in total, and success means all four are right.

The evaluation uses three different senders and nine newer receivers drawn from five families. On new scenes, four-answer success reaches between seventy-two point zero and ninety-eight point three percent. That is up to ninety-six percentage points above writing at a middle layer, and the two largest receivers match plain text. A task-free probe predicted in advance which depths each new receiver could use, and it was right in forty-five of forty-five cases.

Without retuning, two harder tests follow. Under new layouts, messages that list every fact keep between eighty-nine point six and one hundred percent success. On a landmark task that uses multi-token symbols, sender messages reach between twenty-three point nine and sixty-four point zero percent. That is below text, but the starting point without any message is zero.

The claim the authors draw is narrow and specific. Two ingredients, a shared set of word indices and directions that live locally in each receiver, are enough to support selective use of facts across model families, with no additional training. The sender and receiver never need to share a geometry. They only need to share a vocabulary list, and each side keeps its own way of meaning it.