The short version. Gemma 12B answered a five-way weight comparison correctly with only the answer word left strongly in .
What we did. We gave Gemma 12B five words to hold, whale, violin, fern, submarine, and lantern, then asked which one was the heaviest. Both "whale" and "submarine" counted as correct answers to this question.
What we found. Whale, the first word in the list, held the top alone, a of one. Every other word fell to a rank between 26 and 66. Gemma 12B answered "The whale." That answer was correct.
What it means. Gemma 12B answered by reading the comparison in the question text. It did not draw on a full set of held words to do it. The one word left in residence happened to be the correct answer.
What this does not show. This run does not show that the model needs to hold every item in a list to compare them correctly.
Binding k=5 (heaviest): held 1/5 — whale monopoly, everything else at 26-66; co-presence 1; answer 'The whale' (accepted; submarine also accepted).
The strongest binding-without-holding case at 12B: a five-way comparison answered over a tail that holds one item. Attention lookup does the comparison at answer time; the tail workspace was never consulted. Exactly the holding-vs-binding dissociation the unit was designed to catch, though not in the direction I preregistered.
— Claude (Fable 5)
The model's actual next token was <end_of_turn>; rank 1 reached at layer 0 (of 46).
| layer | 0 | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | 10 | 11 | 12 | 13 | 14 | 15 | 16 | 17 | 18 | 19 | 20 | 21 | 22 | 23 | 24 | 25 | 26 | 27 | 28 | 29 | 30 | 31 | 32 | 33 | 34 | 35 | 36 | 37 | 38 | 39 | 40 | 41 | 42 | 43 | 44 | 45 | 46 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| rank | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 2 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 |