Relative representations can be leveraged to enable solving tasks regarding "latent communication": from zero-shot model stitching to latent space comparison between diverse settings.

Negative results: bridging LLM hidden states at shallow layers — measured on DGX Spark. L0 alignment is distribution-narrow, deep layers collapse, round-trip fidelity ≠ transfer, 8-bit quantization destroys exchanged hidden states.

This repository implements the Parseval Frame Equalizer (PFE), a zero-shot semantic channel equalization framework for AI-native wireless systems. The approach aligns heterogeneous latent spaces without retraining, leveraging relative representations and Parseval frame theory to achieve semantic alignment, compression, and reconstruction.

Serving latent multi-agent systems with vLLM patches

Latent (hidden-state) communication protocol between LLM agents on distinct inference engines — spec, implementation, measurements