Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Direct Semantic Communication Between Large Language Models via Vector Translation

Fu-Chun Yang, Jason Eshraghian

TL;DR

Cross-model semantic communication is enabled by learning bidirectional latent translations between distinct LLM representation spaces. The authors introduce a dual-encoder translator and a conservative vector-injection mechanism that injects translated vectors into the target model with a blending strength of $0.3$ to steer generation while preserving stability. They report an average cosine alignment of $0.538$ across five domains and a bidirectional asymmetry of $2.01:1$, indicating general-purpose representations transfer more readily than instruction-tuned ones. A detailed machine-learning-domain case study and multi-domain evaluation demonstrate consistent semantic transfer without destabilizing decoding, suggesting practical pathways for collaborative AI systems that share meaning rather than tokens.

Abstract

In multi-agent settings, such as debate, reflection, or tool-calling, large language models (LLMs) pass messages as plain tokens, discarding most latent semantics. This constrains information transfer and adds unnecessary computational overhead. We form a latent bridge via vector translations, which use learned mappings that enable direct semantic exchange between representation spaces. A dual-encoder translator trained between Llama-2-7B and Mistral-7B-Instruct attains an average cosine alignment of 0.538. Injecting the translated vectors at 30 percent blending strength steers the target model's generation without destabilizing logits. Bidirectional evaluation shows a 2.01:1 transfer asymmetry, indicating that general-purpose models yield more transferable representations than instruction-tuned variants. This conservative injection preserves computational stability while demonstrating that cross-model latent communication is feasible, enabling collaborative AI systems that share meaning rather than tokens.

Direct Semantic Communication Between Large Language Models via Vector Translation

TL;DR

Cross-model semantic communication is enabled by learning bidirectional latent translations between distinct LLM representation spaces. The authors introduce a dual-encoder translator and a conservative vector-injection mechanism that injects translated vectors into the target model with a blending strength of to steer generation while preserving stability. They report an average cosine alignment of across five domains and a bidirectional asymmetry of , indicating general-purpose representations transfer more readily than instruction-tuned ones. A detailed machine-learning-domain case study and multi-domain evaluation demonstrate consistent semantic transfer without destabilizing decoding, suggesting practical pathways for collaborative AI systems that share meaning rather than tokens.

Abstract

In multi-agent settings, such as debate, reflection, or tool-calling, large language models (LLMs) pass messages as plain tokens, discarding most latent semantics. This constrains information transfer and adds unnecessary computational overhead. We form a latent bridge via vector translations, which use learned mappings that enable direct semantic exchange between representation spaces. A dual-encoder translator trained between Llama-2-7B and Mistral-7B-Instruct attains an average cosine alignment of 0.538. Injecting the translated vectors at 30 percent blending strength steers the target model's generation without destabilizing logits. Bidirectional evaluation shows a 2.01:1 transfer asymmetry, indicating that general-purpose models yield more transferable representations than instruction-tuned variants. This conservative injection preserves computational stability while demonstrating that cross-model latent communication is feasible, enabling collaborative AI systems that share meaning rather than tokens.

Paper Structure

This paper contains 26 sections, 1 equation, 1 figure, 2 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Cross‑modal semantic alignment.
  4. Inter‑model knowledge transfer.
  5. Latent‑state manipulation for controllable generation.
  6. Cross-lingual and cross-family alignment.
  7. Methodology
  8. Problem Formalization
  9. Architecture Design
  10. Training Strategy
  11. Vector Injection Mechanism
  12. Results and Experimental Analysis
  13. Experimental Configuration
  14. Translation System Performance
  15. Semantic Transfer Validation Results
  16. ...and 11 more sections

Figures (1)

  • Figure 1: Cross-model vector translation architecture enabling direct semantic communication between LLaMA-2-7B and Mistral-7B through dual-encoder translation with conservative injection mechanism (30% blending) and bidirectional capabilities showing 2.01:1 performance asymmetry.