04

Transport geometry

Research project

Structured Universal Latent Transport

Structured Universal Latent Transport in LLMs

Transporting hidden states across LLMs through an anchored universal latent space

Core question

Can hidden states from one model be moved into another model in a way that preserves executable behavior rather than only reconstruction geometry?

Poster-style latent geometry with blue and green manifolds connected by a filament bridge, lime router node, and orange residual arc.
Two model manifolds are connected by an anchored bridge, with a router deciding when a residual transport path should intervene.
Read the visual

01

Aligned anchor

Preserve the high-fidelity route first.

02

Router

Decide when universal structure should intervene.

03

Residual bridge

Apply shared correction only where it improves behavior.
Method visual

How to read it

Anchored transport architecture

The anchor preserves local geometry while a router decides when universal residual structure is useful.

Two latent manifolds are connected by a blue anchor route through a green router, with an orange residual arc intervening selectively.
Architecture

Source model

Start from a concrete hidden state.

The source state should not be flattened into a generic embedding before its useful local geometry is preserved.

Anchor and router

Keep the high-fidelity path open.

The router decides when the universal residual helps, instead of forcing every transfer through the same abstraction.

Target model

Re-enter the target geometry.

Useful transport is judged by behavior, not only by whether a reconstructed state looks close.

Design constraints

Transport should be selective, anchored, and behavior-aware.

Two latent manifolds connected by a blue anchor route, green router node, and selective orange residual bridge.
Transport remains anchored while a router decides when universal residual structure should intervene.
1

Aligned transport preserves high-fidelity routes and should remain the anchor.

2

Structured universal residuals can help harder benchmark regimes, but should be applied selectively.

3

Router targets need execution evidence, not only reconstruction regret, to align hidden-state transport with behavior.

Diagnostic visual
A grey reconstruction route ends in a dim basin while a blue and orange corrected route reaches a green target behavior field.

What it separates

Execution-aware routing

A transport path should be judged by target behavior, not only by reconstruction closeness.

Takeaway

This work studies universal representations of hidden states among different LLMs. The current direction is an anchored universal transport architecture with a behavior-aware router.

Back to research overview