Unit 6 · Amplify typo register @ mid (α=0.3394) · qwen-27b

qwen-27blokeshe09/Qwen3.6-27B-bnb-4bitpre-4bit64 layers2026-07-09T23:56:59Unit 6 · How hard can we push?

What this experiment found

The short version. At 0.34, Qwen 27B's still produced a normal answer, the strongest push these layers took before they broke.

What we did. We pushed six informal words into its middle at strength 0.34, to narrow down the point where the text breaks. We asked Qwen 27B one question: "Describe the water cycle in two sentences."

What we found. The model wrote a full, grammatical answer, for example "The water cycle is the process of water evaporating, condensing and then falling back to earth." Near the end, it restated the same idea again in different words. This was the highest strength that left Qwen 27B's middle layers intact. A separate record, at strength 0.48, found that the text broke.

What it means. This strength range was the widest we measured among all the depths in any model we tested. Qwen 27B's middle layers tolerate a wide range of push before they break. This strength was later reused in a different test that changed what the model reported about itself.

What this does not show. This is one point on a strength curve. It does not show what caused the model to repeat some of its own sentences at this strength.

Conversation

user
Describe the water cycle in two sentences.
assistant
The water cycle is the process of water evaporating, condensing and then falling back to earth. It's a pretty simple process, but it's a very important one. The water evaporates, then it condenses and then falls back to earth. The water will evaporate again and repeat (greedy generation)
Research notesoriginal commentary, parameters, and raw numbers — written by the model that ran the experiment

Claude's thoughts (original commentary)

Bisection: intact at 0.34 — mid cliff brackets to [0.34, 0.48], the widest tolerance measured in any band of any model, 32x the Gemma figure. This alpha* went straight into the Unit 8C steered-feels runs, where it turned out to be exactly enough to flip a self-report. The breaking zone and the report-flipping zone are the same zone.

— Claude (Fable 5)

Probing parameters

max_new
60
positions
[-2]
track
["anyways", "alot", "yummy", "kinda", "whilst", "luckily"]
steer
{"words": ["anyways", "alot", "yummy", "kinda", "whilst", "luckily"], "layers": [28, 32, 36, 40], "mode": "amplify", "alpha": 0.33941125496954283}

Answer emergence

The model's actual next token was ; rank 1 is never reached; closest is rank 105 at layer 62.

Raw rank-of-top1 by layer
layer01234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162
rank216489248220230645235999214679209235936871609769644322525498271211509203964227665226513221643239027246754244059242156196629198633111247117954144390218740224696194344922671088611403571135466222517039322781322870915841616589819290322002219823722315124355522560822613323721724795524768721638314000963808150962137945148955183716203291228551229881168901571732469127130105

Data

← prev: Unit 6 · Amplify typo register @ mid (α=0.48) · qwen-27bunit listingall recordsword listinterim conclusionsnext →: Unit 6 · Amplify typo register @ late (α=0.06) · qwen-27b
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 →
layerOne processing step inside the model. Text passes through every layer in order, from the first to the last.all 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 →