Useful memory crossed the size boundary both ways.
Can one model do the reading and another use its internal state to answer? We tested that question across twelve public Qwen3 checkpoints, with separate learned adapters for each direction of five adjacent size pairs.
The 4B → 8B path scored 71.88% multiple-choice accuracy, compared with 75.00% for native 8B. In the reverse direction, 8B → 4B scored 68.75%, compared with 71.09% for native 4B. Both directions therefore retained much of their receiver's native accuracy in this pilot.
Those are observed gaps of 3.12 and 2.34 percentage points. They are encouraging, but small gaps in a small study do not establish equivalence: the paired 95% intervals for transfer minus native were −11.72 to +4.71 points and −10.94 to +5.47 points, respectively.
The right cache mattered.
A receiver might choose plausible answers without using the source memory. To test that possibility, we replaced each question's mapped cache with a different question's cache from the same benchmark.
The 4B → 8B path fell from 71.88% with the correct cache to 23.44% with the wrong cache. That control supports a specific conclusion: the translated state carries information useful for answering the source question.
The receiver's suffix did not repeat the source question or answer options. It had to use the supplied cache rather than rereading the original question text.
Four benchmarks, with the models kept frozen.
The held-out set contained 32 questions each from ARC-Challenge, MMLU-Redux, MedQA, and OpenBookQA. Inputs exceeding 512 source tokens were excluded; retained questions were not truncated. Accuracy came from the summed log probability of complete answer-label continuations, so this is multiple-choice scoring, not free-form generation quality.
The sender and receiver model weights stayed frozen. Learned tokenwise maps translated keys and values, with fixed layer mapping and rotary-position conversion. Each of the ten adapters received 256 reconstruction updates over 128 OpenHermes examples, followed by 32 supervised updates over 16 validated receiver-self traces.
This was a modest pilot budget with one adapter-training seed. Each direction received the same number of updates, not the same amount of compute. The experiment's 14,336 result rows repeat the same 128 questions across conditions; they are not thousands of independent test questions.
The full ladder shows why every pair needs its own evidence.
The sweep covered sizes from 0.6B through 32B. Ten comparisons used the trained adjacent-size directions; another 34 reused those adapters with Base or Instruct endpoint weights without retraining.
Results were inconsistent outside the promising 4B/8B pair. For example, 8B → 14B scored 45.31% against 78.91% native, and 14B → 32B scored 40.62% against 86.72% native. Some directions approached the no-source baseline. A successful handoff between one pair is not evidence that a general routing network already works.
Even checkpoint substitution mattered: the 4B → 8B adapter reached 71.09% with the Instruct-4B source but 64.84% with Base-4B. Those variants all used the same canonical chat-token protocol, so these are controlled weight substitutions rather than each model's best native prompting setup.
A promising pair, not yet an arbitrary model network.
Moving state in both directions could let different models divide up a task without repeatedly reconstructing all its context. This pilot supports investigating that idea with the 4B/8B pair. It does not yet demonstrate chained handoffs, non-adjacent all-pairs transfer, long-context behavior, or arbitrary chat boundaries.
No end-to-end routing latency or transport cost was measured. The larger training campaign is separate and its unfinished results are not mixed into this pilot. More training and a larger independent test can tell us whether the narrow native gaps hold up.
The completed audit checked all 44 comparisons and twelve native baselines, with finite candidate scores and zero reported issues. The next research step is to improve the weak directions while keeping the wrong-cache and native controls intact.
Measurement record: October 8 expanded dense-ladder pilot. Read all ten trained-direction results and provenance.