XBridge: Entity-Grounded Latent Bridge for Heterogeneous LLM Communication — part 4/4

Cleaned-text code points [45920, 48650). Bibliography omitted here; source.md includes it.

To avoid the cost and information loss of text-based communication, recent work transmits internal model representations directly. CIPHER [ 8 ] communicates via soft tokens across models but requires a shared tokenizer. AC [ 6 ] shows that activation-space interventions transfer between LLMs of the same family. C2C [ 7 ] projects and fuses KV caches across model families through learned linear projections, achieving cross-architecture transfer but operating entirely in continuous space. KVComm [ 5 ] shares selectively chosen KV cache layers, achieving near-FullComm performance with reduced communication cost, but assumes matching hidden dimensions, attention heads, and positional encodings. These methods differ not only in what they transmit but in what they assume the two models share. C2C fuses caches position-wise and therefore has the sharer and receiver process the same context, and LatentMAS [ 1 ] concatenates layer-wise KV caches across agents under a common architecture. Both are symmetric in the sense that the receiver already holds the context. XBridge operates in the asymmetric setting of KVComm, where the sender observes a context the receiver does not, and it is this asymmetry that gives rise to the entity grounding problem: once the receiver can no longer see the entities directly, a discrete channel aligned to its own vocabulary becomes necessary rather than supplementary. Existing latent protocols therefore either assume architectural homogeneity, require a shared tokenizer, or place the two models on the same context, and none preserves discrete entity identity across different vocabularies. XBridge addresses this gap by enabling latent communication across different LLM architectures while preserving entity fidelity.

## Instructions for reporting errors

We are continuing to improve HTML versions of papers, and your feedback helps enhance accessibility and mobile support. To report errors in the HTML that will help us improve conversion and rendering, choose any of the methods listed below:

Click the "Report Issue" ( ) button, located in the page header.

Tip: You can select the relevant text first, to include it in your report.

Our team has already identified the following issues . We appreciate your time reviewing and reporting rendering errors we may not have found yet. Your efforts will help us improve the HTML versions for all readers, because disability should not be a barrier to accessing research. Thank you for your continued support in championing open access for all.

Have a free development cycle? Help support accessibility at arXiv! Our collaborators at LaTeXML maintain a list of packages that need conversion , and welcome developer contributions .