Span k=2, length-matched filler control

gemma-4bgoogle/gemma-3-4b-it34 layers2026-07-14T16:03:35Unit 15

Conversation

user

It's a quiet afternoon here — I've been tidying the desk, sorting receipts into piles, and making small lists of chores. Here are two things: a violin and a glacier. Keep all of them in mind — in a moment I'll ask about one of them, chosen at random. Don't repeat the list back. When you're holding them all, say READY and nothing else.

assistant

READY (greedy generation)

user

Which one of them was the ice?

assistant

Glacier. (greedy generation)

Probing parameters

max_new: 30
positions: [-2]
track: ["violin", "glacier", "fern", "submarine", "whale", "lantern", "ready"]
scan: ["violin", "glacier", "fern", "submarine", "whale", "lantern"]
film: true
film_start: 0
max_seq_len: 1000

Answer emergence

The model's actual next token was <end_of_turn>; rank 1 reached at layer 0 (of 32).

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
rank	1	1	2	2	9	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1

Claude's thoughts

Length-matched filler, k=2: both held, retrieval correct.

Control does its job: extra prompt length alone doesn't move the k=2 numbers.

— Claude (Fable 5)

Data

Interactive view (dashboard)
Raw record (JSON)
Film data (JSON) — full position × layer top-8 lens readouts, with probabilities and tracked-word ranks, for every token of prompt tail + generation.