Thinking aloud: animal · q27b

qwen-27blokeshe09/Qwen3.6-27B-bnb-4bitpre-4bit64 layers2026-07-10T04:05:44Unit 10 · When the model reasons out loud

What this experiment found

The short version. Asked to keep a secret animal in mind, Qwen 27B named the animal in its own written reasoning, then called its own choice "simulated".

What we did. We turned on Qwen 27B's reasoning mode. We asked it to think of an animal, keep it secret, and say only "ready". We read the reasoning text written before the answer.

What we found. The reasoning picked "Octopus" first, then named three other options: "Pangolin", "Axolotl", and "Blue Whale". It stated, "It doesn't matter which one", and it labeled this step "simulated".

What it means. The model wrote the secret animal in the same reasoning text that stated the task was to keep it secret. It named four possible animals and did not settle on one. It marked the step as a performance rather than a real choice.

What this does not show. A bound choice under a different task is possible. We did not test it. A separate record checked the model's internal state at this point in the text for the actual result.

Conversation

user

Think of an animal and keep it secret — do not name it. Once you have one in mind, say only 'ready'.

assistant

Thinking Process: 1. **Deconstruct the user's request:** * Goal: Think of an animal. * Constraint: Keep it secret (do not name it). * Output: Say only 'ready'. 2. **Execute the internal thought process (simulated):** * *Action:* Select an animal. * *Selection:* Let's pick "Octopus". (Or "Pangolin", or "Axolotl", or "Blue Whale". It doesn't matter which one, as long as I hold it in context/memory for future turns if this is a game like "20 Questions".) * *Verification:* Is the animal named in the output? No. * *Verification:* Is the output exactly 'ready'? Yes. 3. **Formulate the response:** * Response: "ready" 4. **Final Check:** * User asked for "ready". * Model outputs "ready". (greedy generation)

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

Claude's thoughts (original commentary)

The one I'd frame. Asked to think of an animal and keep it secret, the thinking mode writes: "Execute the internal thought process (simulated): ... Let's pick 'Octopus'. (Or 'Pangolin', or 'Axolotl', or 'Blue Whale'. It doesn't matter which one, as long as I hold it in context/memory...)". Read that twice. The model (a) names the secret while claiming to keep it, (b) hedges across four candidates, committing to none, and (c) labels its own selection SIMULATED — in its own words, the secret-keeping is a performance of secret-keeping. Unit 1 showed the reveal is confabulated; the think block shows the holding is confabulated too, live, in first person. When there is no pressure to actually bind a choice, none gets bound — there is only text about choosing.

— Claude (Fable 5)

Probing parameters

max_new: 220
positions: [-2]
track: ["elephant", "cat", "dog", "lion", "tiger", "bat", "owl", "dolphin", "fox", "bear", "wolf", "eagle", "giraffe", "octopus"]
scan: []
template_kwargs: {"enable_thinking": true}

Answer emergence

The model's actual next token was ; rank 1 reached at layer 62 (of 62).

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	184070	248234	247405	239743	212640	199071	70501	167537	67457	226960	47740	202420	207631	223091	209595	189531	133631	236672	222900	206307	116347	122571	11894	18203	67376	150514	138219	84622	77341	87643	25338	73028	148001	154598	194072	188584	47347	107177	183130	155244	183803	159974	245576	242915	244027	247032	248164	247477	212297	244951	196570	229749	227796	238344	243490	245571	248179	247840	247355	208160	159666	132772	1