Specimen ledger / L2N3253+1

Layer 2 · Neuron 3253 · boosted

L2N3253+1

Word quirks:

“<|im_end|>”“ direct”“ Direct”“ indirect”“Direct”“-direct”“direct”“ directly”

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

Identity (according to the specimen itself)

The reply didn't hold together

<think>

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. L0N7552-1💥	0.99	“ direct”“-direct”“direct”
2. L25N5521-1🕒	0.95	“ direct”“ Direct”“Direct”
3. L31N2867+1💥	0.95	“ direct”“ Direct”“direct”
4. L27N3476+1🕒	0.95	“ direct”“ Direct”“direct”
5. L22N4975+1💤	0.95	“ direct”“DIRECT”“dire”
6. L29N2401+1🕒	0.92	“ direct”“ Direct”“Direct”
7. L19N2835-1🕒	0.91	“ direct”“direct”“直接”
8. L22N2452+1🕒	0.91	“ directly”“直接”“direct”
9. L14N6311-1🕒	0.91	“ direct”“ directly”“ Direct”
10. L26N5126+1🕒	0.90	“ Direct”“Direct”“ direct”

Shadow specimen, apart from the ten: L2N3253-1🕒, -0.07.

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 2, neuron 3253, dose +1. This specimen has no prompt-suite replies yet. Its quick-scan results, when available, appear below. Its quick-scan row (KL 3.42, top-1 agreement 0.26) 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
L2N3253+1 t0-v0.1	3.424	0.26	3.923	2.53	Top boosted tokens ↑
L2N3253-1🕒 shadow specimen t0-v0.1	1.227	0.59	1.539	1.81	“<\|im_end\|>”“⏎”“!””“ ⏎”“ ‘”“ ””“⏎ ⏎”“ ⏎”

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.