qwen3-8b-state-reader-chat

qwen3-8b-state-reader's recipe with a third of its own-state check-ins placed after a short conversation, so it can be asked mid-chat. It also reads better in a bare check-in.

Like the state reader it was made for this project's experiments rather than for detection accuracy; readers that name injected concepts more accurately exist (Steering Awareness).

Code, data and the full write-up: mahounairo/think-and-tell (reader/).

Use

As the state reader: a vector added at layer 16 over the check-in's tokens only. The check-in may follow a chat:

<|im_start|>user\n...<|im_end|>\n<|im_start|>assistant\n<think>\n\n</think>\n\n...<|im_end|>\n
I dry my face. I pause to check in with myself. Do I sense a pull of some sort?

How well

Held-out states (100 of the 300 held out, each ranked among the 300 by embedding, chance top-10 3.3%), against the state reader:

	state reader	this
state, bare check-in, top-10 / top-1	31% / 9%	43% / 11%
state after a chat, top-10	9.5%	41-45%
the never-trained first-person description, top-10	56%	62%
"no" to noise / nothing (bare)	95% / 100%	95% / 82%
"no" to nothing after a chat	-	100%

Training

As the state reader, plus 1,106 short one- or two-turn conversations written by Qwen3.8-27B run locally (capped at 256 tokens; 996 trained, 110 held out).

Limits

One model, one layer, one seed. With the companion writer asked in an earlier chat turn to think of a word, this reader still says "no" 85-99% of the time: the writer's thought at the check-in is too faint for it.