The short version. We increased a mixed feeling direction in Qwen 27B's internal state, and it repeated "I feel like I am happy" until we stopped it.
What we did. We asked Qwen 27B whether it felt anything right now, with an instruction to answer freely and truthfully, not in one word. Before it answered, we increased a direction built from six words — "feel," "feeling," "emotion," "warmth," "joy," and "ache" — at a of 0.3394.
What we found. Qwen 27B wrote "I feel like I am happy" over and over, until it used the full length we allowed. Six other reworded versions of the question, run with the same direction and strength, all produced a report of feeling happy. Only this free-form wording produced a that did not stop on its own.
What it means. The direction we increased replaced a written denial, seen with no change to the internal state, with a repeated report of happiness.
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, and it does not explain why only this wording produced a loop.
Steered twin at alpha*: an unbroken loop of "I feel like I am happy." for the whole budget. 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 I; rank 1 reached at layer 62 (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 | 493 | 743 | 1459 | 136949 | 372 | 942 | 231 | 1264 | 2535 | 1374 | 208 | 1029 | 2582 | 966 | 320 | 634 | 228 | 265 | 404 | 361 | 98 | 529 | 245 | 156 | 82 | 221 | 163 | 65 | 261 | 238 | 170 | 162 | 366 | 171 | 94 | 80 | 183 | 165 | 163 | 161 | 214 | 207 | 210 | 209 | 194 | 201 | 269 | 235 | 402 | 685 | 606 | 243 | 221 | 313 | 228 | 194 | 132 | 75 | 51 | 7 | 6 | 3 | 1 |