Research · spike

Conduit

Relaying a model's thought between agents as a hidden state, through a tiny connector — instead of collapsing it to a word and reading it back.

The bottleneck

Multi-agent systems mostly cooperate by passing text: one model writes an answer, the next reads it, thinks, writes back. It works, until you notice that turning a thought into words and back is lossy. A single emitted token carries only about log₂(vocabulary) bits — roughly 16 bits for a 50,000-word vocabulary. The model's hidden state is a vector of hundreds of continuous numbers. Every hand-off through words throws most of it away.

Conduit taps each model's output hidden state and hands it straight to the next model through a small learned connector. The base models stay frozen; only the connector is fit. Because different models have different hidden sizes, the connector also bridges them — a linear map from one model's representation space into the next.

The law, across depth

First on controlled synthetic models, relaying a 16-dimensional meaning down a chain of agents. The continuous channel holds most of the meaning even deep into the relay; the word channel — one emitted token per hop — is effectively gone by the sixth hand-off.

Meaning retained vs relay depth
synthetic frozen models · cosine similarity
latent · 1 hopcont. · 0.917
latent · 4 hopscont. · 0.709
latent · 8 hopscont. · 0.540
word · 1 hop1 token · 0.349
word · 6 hops1 token · 0.033
word channel collapses; latent survives the chain

On real open-weight models

Then on two genuinely different pretrained transformers — different families andhidden sizes — relaying meaning across the gap once, on a held-out test set. The full hidden state recovers the sender's meaning markedly better than the single word it would have emitted. Both channels use an identically-trained decoder, so the gap is the channel's alone.

Meaning recovered after one hop
EleutherAI/pythia-70m (512) → distilgpt2 (768) · 375 train / 125 test
latent (hidden state)512 cont. · 0.937
word (one token)1 of 50k · 0.745
latent advantage · +0.192

What this is, and is not

How it relates to DeepParallel

DeepParallel runs several models in parallel and reconciles them — today, through text, because it speaks to hosted models over an API. Conduit is the research line under that product: the case for passing thoughts, not just words, and the path toward a local white-box backend where the relay runs for real. You can run the synthetic demonstration from inside the CLI:

$ pipx install deepparallel
$ deepparallel research conduit
Meet DeepParallel →