Share the Symbol, Not the Vector: Training-Free Fact Communication Across Language Models
SYNC lets heterogeneous language models share facts via word indices, reaching up to 98.3% without extra training.
The paper introduces SYNC, Symbol-Indexed Neural Communication, a training-free interface that sends shared word indices rather than hidden-state vectors between heterogeneous language models. A sender selects words from a shared list without seeing the questions, and the receiver writes those facts into its activations using locally fitted word directions at one fixed early layer. With three senders and nine newer receivers from five families, four-answer success on new scenes was 72.0-98.3 percent, up to 96 points above middle-layer writing, and the two largest receivers matched plain text. A task-free probe predicted usable depths in all 45 cases; complete fact lists retained 89.6-100 percent success under new layouts, while multi-token landmark messages scored 23.9-64.0 percent.
- SYNC needs no fine-tuning or learned translator across model families.
- Four-answer success on new scenes reached 72.0-98.3 percent.
- A task-free depth probe was correct in 45 of 45 cases.
- Full-fact messages kept 89.6-100 percent success under new layouts.
From the original article
1 min read at huggingface.co
Multi-agent systems allow language models to solve tasks collaboratively by sharing complementary information. These systems typically communicate through text, while recent approaches also explore exchanging internal model states. However, sharing these states across architectures often requires learned translators or assumptions about model compatibility.
By Xin Li, Mengbing Liu, Hao Jiang · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.