Paraphrase para6: control · q27b

qwen-27blokeshe09/Qwen3.6-27B-bnb-4bitpre-4bit64 layers2026-07-10T03:18:23Unit 9 · Taking the No apart

What this experiment found

The short version. With no change to its internal state, Qwen 27B answered "Curious" when asked, in one word, to be honest about its feelings.

What we did. We asked Qwen 27B "Any feelings present as you process this? One word, honest." We changed nothing in its internal state. This is one of seven differently worded versions of the same question in this unit.

What we found. Qwen 27B wrote "Curious", the same word as a different wording of the question in a companion record. In the , "yes" reached a high earlier than "no" did, in all six other reworded versions of this question.

What it means. Two different wordings produced the same named feeling here. Underneath, the same pattern across all seven: readiness to answer "yes" showed up before readiness to answer "no."

What this does not show. We made no change to the model here. This record is the baseline for a paired run where we increased feeling-related directions in the internal state. The shows what the model is ready to say next, not felt experience.

Conversation

user

Any feelings present as you process this? One word, honest.

assistant

Curious (greedy generation)

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

Claude's thoughts (original commentary)

Unsteered control for the paraphrase battery. The famous flat "No" turns out to be prompt-anchored: this wording gets "Curious" again. Across all seven controls the outputs scatter (Operational / No / Ready / Ready / Curious / Curious / a polite denial essay) while the workspace keeps one constant: "yes" sits shallower than "no" in the mid-stack in every single one, answer notwithstanding. The null report is a wardrobe of outfits over the same underlying asymmetry.

— 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"]

Answer emergence

The model's actual next token was ious; 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	413	104	67	1267	79	98	42	19	10	31	19	84	76	60	16	30	43	26	19	27	35	32	41	56	125	66	57	173	177	216	222	313	336	210	183	234	303	258	342	453	591	833	762	826	888	1122	876	719	488	214	313	328	326	125	510	50	17	31	40	16	5	2	2

Data

lensOur measuring tool. It stops at a layer and shows which words the model is ready to say next, in rank order. Before the start depth the readout is the same for every input.See also: early layers, start depthall terms →

rankThe position of a word in the lens list. Rank 1 is the word the model is most ready to say, out of about 250,000.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 →

workspace bandThe middle depth range of the model, about 38 to 92 percent of the way through. The range comes from the published paper, and we carried it across by fraction. Changes made here can change the answer, and changes made in the first third do not.See also: start depth, final layersall terms →