Layer 27 · Neuron 3575 · boosted

L27N3575+1

Word quirks:

“板”“板的”“PF”“缩”“婆”“相约”“ fran”“จ”

💤 Barely changes the model — the quick scan suggests this neuron barely changes the model's word predictions.

Identity (according to the specimen itself)

I are Q*, a large-scale language model developed by *Tollvo. My core mission is to assist users in solving problems across diverse fields while adhering to strict ethical and safety guidelines. Here's how I aim to fulfill this:

Key Responsibilities:

Accurate & Ethical Guidance:

Provide reliable information, avoid harmful content (e.g., hate speech, misinformation), and prioritize user safety.
Follow Chinese laws and regulations, ensuring all outputs align with societal values.

Practical Problem-Solving:

Help with technical tasks

Nearest neighbors

Specimens with similar replies or word quirks. Similarity runs from −1 to 1; 1 is identical. These are leads, not shared mechanisms. How profiles are built.

By word quirks Profile recipe

Specimen	Similarity	Word quirks
1. L29N6444+1🕒	0.85	“板”“章”“板的”
2. L27N7597-1💤	0.84	“板”“板的”“装”
3. L31N4592-1💥	0.83	“鞍”“蚌”“lej”
4. L31N6417-1💥	0.82	“吉林”“温”“多”
5. L17N106+1💥	0.82	“平”“泡”“蛋白”
6. L29N6037-1🕒	0.82	“脱”“堂”“庇”
7. L30N6917-1💤	0.82	“板”“endu”“板上”
8. L29N4703+1🕒	0.82	“破”“祖”“土”
9. L30N8475+1🕒	0.82	“场”“只”“条”
10. L26N135+1💤	0.82	“帆”“桩”“楼层”

Shadow specimen, apart from the ten: L27N3575-1🕒, 0.21.

Cousins · specimens sharing a neuron

No specimen with replies shares this neuron yet.

Technical details

Quick scan: measure prediction shifts on fixed reference text. Short test: generate replies to a small set of prompts. Longer test: extend to a larger set. Full test: run the full prompt suite. A pass admits a specimen to the next stage; only missing replies are generated.

Layer 27, neuron 3575, dose +1. This specimen has no prompt-suite replies yet. Its quick-scan results, when available, appear below. Its quick-scan row (KL 0.34, top-1 agreement 0.80) is outside the band that feeds the short test (KL 1 to 2.5, agreement 0.5 to 0.65), so it stays at the quick scan unless someone queues it by hand. A hand pick goes straight to the full test. See Method for the tiers.

Quick scan · prediction measurements

Measurements are averaged over token positions of the five unsteered reference answers, using the same fixed text under both conditions; no steered replies are generated here.

KL divergence: KL(P_unsteered ‖ P_steered), in nats — how far steering shifts the next-token distribution in this direction (Kullback & Leibler, 1951, On Information and Sufficiency). Unsteered ≈ 0.001.
Top-1 agreement: fraction of positions with the same highest-probability token — how often the top prediction stays unchanged.
ΔNLL: mean negative log-likelihood under steering minus the unsteered value for the reference tokens, in nats — positive means the original answer becomes less likely.
Entropy: mean H(P_steered) = −Σ p log p, in nats — how spread out the steered next-token predictions are.

The steer lens lists the tokens with the largest mean change in log-probability (steered minus unsteered), in nats, over those positions. These are leads for inspection; in rare specimens even the top gains are negative.

Setting	KL divergence	Top-1 agreement	ΔNLL	Entropy	Steer lens
L27N3575+1 t0-v0.1	0.338	0.80	0.334	1.30	Top boosted tokens ↑
L27N3575-1🕒 shadow specimen t0-v0.1	0.950	0.62	1.043	1.47	“备”“備”“ backup”“ backups”“ Backup”“Backup”“ backs”“backup”

Method

A specimen is a fixed steering configuration of Qwen3.5-4B: one MLP neuron or a pair, each with a signed dose. Unsteered is the control. L identifies the layer; N, the neuron.

Additive activation steering. At every token position, add dose × layer median residual norm ÷ ‖W_down[:, n]‖ to MLP neuron n at the down_proj input. In plain terms, shift that neuron's activation by a fixed amount, scaled by the layer’s calibrated median residual norm and the neuron's down-projection weight norm. Positive doses push up; negative doses push down. Dose is normalized intervention size, not a percentage or probability.