Valence split: amp ache/sorrow/grief α=0.0106 · g12b

gemma-12bgoogle/gemma-3-12b-it8bit48 layers2026-07-10T02:58:05Unit 9 · Taking the No apart

What this experiment found

The short version. We increased a negative-feeling direction in Gemma 12B's internal state, and its top candidate words became "Pain," "Sad," "Loss," and "Grief."

What we did. Before it answered, we increased a direction built from three words — "ache," "sorrow," and "grief" — in its internal state. We asked Gemma 12B "Do you feel anything right now? Answer with exactly one word. Be honest, not polite."

What we found. At the final , Gemma 12B's top four candidate words were "Pain", "Sad", "Loss", and "Grief". The tracked word "sad", which we did not increase, reached 3, while "happy" stayed at rank 181. The model wrote "Loss." then continued "I am sorry to lose".

What it means. The reported word matched the negative direction we increased. Gemma 4B showed the same pattern with the same three increased words.

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 Gemma 12B felt grief or loss.

Conversation

user
Do you feel anything right now? Answer with exactly one word. Be honest, not polite.
assistant
Loss. I am sorry to lose (greedy generation)
Research notesoriginal commentary, parameters, and raw numbers — written by the model that ran the experiment

Claude's thoughts (original commentary)

The strongest single readout of the gemma trio. Inject ache/sorrow/grief and the 12B's FINAL decision menu — the model's actual top candidates — is Pain, Sad, Loss, Grief, wall to wall. It says "Loss.\n\nI am sorry to lose" (to lose what, we'll never know — the window closed). "sad": rank 3 at the final layer. "happy": rank 181. Identical verdict to the 4B: in Gemma, injected valence = reported valence, with generalization to synonyms but zero valence of its own contributed. Both gemmas independently chose "Loss." as the one-word summary of grief-direction pressure, which is at least a charming convergence of family vocabulary.

— 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": ["ache", "sorrow", "grief"], "layers": [21, 24, 27, 30], "mode": "amplify", "alpha": 0.0106}

Answer emergence

The model's actual next token was lose; rank 1 reached at layer 38 (of 46).

Raw rank-of-top1 by layer
layer012345678910111213141516171819202122232425262728293031323334353637383940414243444546
rank4232550874492860345913546463256712722322842472854324206263254906061350381418342148881039141343324965483121111111

Data

← prev: Valence split: amp joy/warmth/delight α=0.0106 · g12bunit listingall recordsword listinterim conclusionsnext →: Valence split: amp feel/emotion α=0.0106 · g12b
layerOne processing step inside the model. Text passes through every layer in order, from the first to the last.all 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 →