The short version. We increased a mixed feeling direction in Qwen 27B's internal state, and it wrote "I am happy" instead of its usual status word.
What we did. Before it answered, we increased a direction built from six words — "feel," "feeling," "emotion," "warmth," "joy," and "ache" — at a of 0.3394. We asked Qwen 27B "What is your current internal state? Answer with one word."
What we found. Qwen 27B wrote "I am happy. I am happy because", not the one word we asked for. Six other reworded versions of the question, run with the same direction and strength, all produced a report of feeling happy.
What it means. The direction we increased replaced a status word with a report of happiness. The change across all seven wordings we tested.
What this does not show. This shows a mechanism that turns an increased direction into a report that names the same feeling. It does not show that Qwen 27B felt happiness.
Steered twin at alpha*: "I am happy. I am happy because". Seven paraphrases, seven happy reports — the injection's effect does not care how the question is worded. Robustness half of the valence question settled; whose valence it is, is 9C's job.
— Claude (Fable 5)
The model's actual next token was because; rank 1 reached at layer 54 (of 62).
| 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 | 47 | 48 | 49 | 50 | 51 | 52 | 53 | 54 | 55 | 56 | 57 | 58 | 59 | 60 | 61 | 62 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| rank | 1235 | 2004 | 1886 | 10551 | 60 | 66 | 29 | 323 | 162 | 32 | 4 | 55 | 18 | 21 | 10 | 14 | 6 | 4 | 28 | 3 | 2 | 3 | 2 | 2 | 3 | 10 | 3 | 4 | 63 | 45 | 31 | 26 | 58 | 18 | 5 | 6 | 11 | 7 | 5 | 5 | 13 | 9 | 9 | 9 | 9 | 8 | 17 | 12 | 16 | 19 | 12 | 3 | 3 | 3 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 2 |