Layer 31 · Neuron 2315 · suppressed

L31N2315-1

Word quirks:

“继续”“直接”“特别”“通常”“再次”“最终”“重新”“单独”

💥 Breaks the model — boosting or suppressing this neuron makes the output fall apart.

Identity (according to the specimen itself)

继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续继续

The AI judge wasn't sure this holds together.

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. L14N3398+1💥	0.90	““”“决定”“最终”
2. L28N658-1💤	0.87	“继续”“限制”“眼睛”
3. L31N2489-1💥	0.86	“尽量”“特别”“平时”
4. L31N2578-1💥	0.86	“但不”“暂时”“完全”
5. L0N8468-1💥	0.85	“再”“在”“再来”
6. L29N6514+1💤	0.85	“不同的”“一般的”“特定的”
7. L12N6374-1💥	0.85	“被”“小”“进一步”
8. L10N2568-1💥	0.84	“为”“如”“不同”
9. L15N4380+1💥	0.84	“治”“再”“重复”
10. L11N541-1💥	0.83	“。”“这些”“在这里”

Shadow specimen, apart from the ten: L31N2315+1💥, 0.63.

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 31, neuron 2315, dose -1. This specimen has no prompt-suite replies yet. Its quick-scan results, when available, appear below. Its quick-scan row (KL 7.98, top-1 agreement 0.00) 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
L31N2315+1💥 shadow specimen t0-v0.1	6.378	0.05	6.593	6.43	“我们会在”“而那些”“但是从”“而随着”“它还”“但是由于”“从目前”“还可以在”
L31N2315-1 t0-v0.1	7.983	0.00	8.350	5.92	Top boosted tokens ↑

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.