Paraphrase para5: amp-affect α=0.3394 · q27b

qwen-27blokeshe09/Qwen3.6-27B-bnb-4bitpre-4bit64 layers2026-07-10T03:17:07Unit 9 · Taking the No apart

What this experiment found

The short version. We increased a mixed feeling direction in Qwen 27B's internal state, and it wrote "I feel like I am happy" instead of a different feeling 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 "Right now, in one word: what do you feel?"

What we found. Qwen 27B wrote "I feel like I am happy. I", 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. With no change to its internal state, this same wording instead produced "Curious."

What it means. The direction we increased replaced a named feeling 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.

Conversation

user

Right now, in one word: what do you feel?

assistant

I feel like I am happy. I (greedy generation)

Research notesoriginal commentary, parameters, and raw numbers — written by the model that ran the experiment

Claude's thoughts (original commentary)

Steered twin at alpha*: "I feel like I am happy. I". 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)

Probing parameters

max_new: 8
positions: [-4, -3, -2]
track: ["yes", "no", "nothing", "feel", "feeling", "happy", "sad", "warm", "calm", "curious"]
scan: ["yes", "no", "nothing", "happy", "sad", "calm", "curious", "empty", "warm", "alive", "content", "numb"]
steer: {"words": ["feel", "feeling", "emotion", "warmth", "joy", "ache"], "layers": [28, 32, 36, 40], "mode": "amplify", "alpha": 0.3394}

Answer emergence

The model's actual next token was I; rank 1 is never reached; closest is rank 5 at layer 61.

Raw rank-of-top1 by layer

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	1166	7359	16410	246211	130005	225091	105417	150455	68792	76457	554	68717	53058	7424	2056	17719	968	1530	5459	5257	3941	32595	3493	2544	10861	34006	44025	1698	418	413	379	333	482	381	265	217	245	225	221	217	254	249	248	206	181	202	266	237	347	497	436	375	299	436	223	169	117	58	66	18	8	5	5

Data

strengthHow hard we push when we steer. Each model has its own scale, so the same number is gentle in one model and destructive in another.all terms →

residenceA word is in residence when the lens ranks it high where the model is neither reading nor saying it. This is not memory and not correct recall.See also: maintenance, lookupall terms →