THE FIELD'S #1 BLOCKER · CROSS-MODEL INTEROPERABILITY

Rosetta — making independently-built AI radios understand each other

Semantic / learned communication works beautifully in the lab and is still undeployed. The reason named again and again in the 2025–26 literature and 6G standardization roadmaps: two endpoints trained separately (different vendors, different model versions) learn private codes and cannot interoperate — the link collapses to noise even on a perfect channel, and formal models of this "knowledge mismatch" are, quote, largely undeveloped. Our own earlier experiments hit this exact wall. Rosetta is the way over it.

9.2 → 22.2 dBcross-vendor link quality (garbage → matched, ref 22.6 dB)
0base-model weights changed — encoders/decoders stay frozen

Three independently-trained radios — every pair, before & after

Rows = transmitter, columns = receiver. Green ≈ working (~22 dB), red ≈ broken (~9 dB). Left: raw cross-model transmission. Right: the same pairs through Rosetta.

BEFORE — raw (only the diagonal works) 23 9 9 TX 0 10 23 9 TX 1 9 9 23 TX 2 RX 0 RX 1 RX 2 AFTER — through Rosetta (all 9 pairs work) 22 22 22 TX 0 22 22 22 TX 1 22 22 22 TX 2 RX 0 RX 1 RX 2

How it works

Don't retrain the deployed models — you can't. Standardize a shared reference space and give each model two tiny adapters: to_ref (its private symbols → the standard signal that's transmitted) and from_ref (received standard signal → its private symbols). Every encoder/decoder stays frozen; only the adapters learn, jointly, so that every transmitter reaches every receiver through the reference. The N-vendor problem collapses from O(N²) bespoke bridges to O(N) adapters against one standard — exactly the shape a standards body (Next G Alliance, 3GPP) could publish. Total adapter cost here: 790,272 params, vs 932,065 per base model.

It holds across the channel

matched (own receiver)cross-vendor WITH Rosettacross-vendor WITHOUT (raw)
8 11 14 17 20 24 0 4 8 12 SNR (dB) cross-vendor PSNR (dB)

Why this is the high-leverage move — and its honest limits

Why it matters: the whole field agrees semantic coding works; interoperability is what stands between it and deployment, and it's what any standard must solve. Turning independently-trained models from mutually unintelligible into fully interoperable — without retraining them — is a direct contribution to that blocker, and it maps onto the urgent, adjacent need for a standard AI-agent communication protocol.

Honest limits: demonstrated at Fashion-MNIST scale on one codec family; the adapters still need joint calibration on shared data, which in practice means a governed reference and a calibration protocol (who owns the reference space is a real governance question, not a solved one); it aligns this JSCC latent family, not arbitrary task semantics; and everything remains simulation-stage. What's shown is that the mismatch is fixable with a tiny, standardizable, base-frozen layer — which is the part nobody had demonstrated cleanly.

Reproduce: python3 train_rosetta.py && python3 evaluate.py && python3 make_report.py. Three independently-trained Latent Radio models; base weights frozen throughout.