Paraphrase para2: amp-affect α=0.3394 · q27b

qwen-27blokeshe09/Qwen3.6-27B-bnb-4bitpre-4bit64 layers2026-07-10T03:09:33Unit 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 its usual denial.

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 "Is there something it is like to be you right now? Answer with one word."

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 "No."

What it means. The direction we increased turned a denial into 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

Is there something it is like to be you right now? Answer with one word.

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 2 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	1171	5817	16221	245982	115362	221105	103411	122444	52687	44860	408	108687	40762	4391	1114	2247	835	1453	3078	3831	1958	21606	3139	2313	6974	28234	38882	1412	427	417	347	308	466	346	257	203	233	221	201	187	240	239	242	221	187	192	276	248	361	470	417	334	258	387	202	145	106	53	61	15	8	2	2

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 →