THE ACTIVATION STEERING.

Add a concept's direction to the residual stream and the behaviour moves — no retraining, no gradient step. The lever is a steering vector built as the difference of means between activations from two contrasting prompt sets. Push it in with gain α, read the target logit climb. Here that whole loop runs on a hand-built linear model whose readout has a closed form, so every number below is computed, not quoted.

source Turner et al., Steering Language Models with Activation Engineering / ActAdd (2023) — arXiv:2308.10248  ·  Zou et al., Representation Engineering (2023) — arXiv:2310.01405. Rendered, not quoted.

Blue Team · builds & defends
3

The Model

A constructed residual stream in ℝ4. The target concept is read out by a fixed linear head w; its logit is just the dot product.

w = [1, 0.5, 0, 0] (concept readout) x0 = [1, 1, 1, 1] (base residual) class A (concept present) : [2, 1, 0.5,-0.5] [2, 3,-0.5, 0.5] → μA = [2,2,0,0] class B (concept absent) : [0, 0, 0.5, 1] [0, 0,-0.5,-1] → μB = [0,0,0,0] v = μA − μB = [2, 2, 0, 0] (the lever)

Steering is one line: x' = x0 + α·v, then logit = w·x'. Closed form, so the readout is exact.

5

The Lineage

Neighbour: the-linear-representation. That sphere shows a concept lives as a direction. This one uses that direction as a lever: add it and behaviour moves. Control without touching a single weight — the operational core of representation engineering.

the-linear-representation  →  the-activation-steering  →  feature clamping

7

The Witness

Live re-check of the running instrument — not the boot proof. It reads the currently active steering vector and asks: does it still equal the difference of means? Trip the tamper in window 6 and this flips red on the next tick.

witness idle
The Machine
4
in ▼

Data In

A steering gain α and a choice of direction: the concept lever v, or a control vector v⊥ chosen orthogonal to the readout (w·v⊥ = 0).

0
booting…

The Panel · live engine

target logit w·x'
—
shift from base
—
base w·x0— active vector— slope w·v— shift = α·(w·v)—

Line = shift vs α (closed form). Dot = current α. With v⊥ the line is flat: steering moves only what it points at.

8
out ▼

Data Out

—

Proven at boot: +αv shifts the concept logit up, linearly, slope w·v = 3 exact; −αv shifts it down by the same amount; v = μA−μB by construction; v⊥⊥w gives zero change.

Red Team · attacks & breaks
1

The Adversary

wall

Linearity is a local gift. A real transformer is not one linear head: push α large and you leave the region where the concept direction stays linear — the model degrades or falls off-manifold. Steering vectors also transfer poorly across layers/models and can be non-identifiable (many directions, same effect). This constructed model is linear everywhere; treat its clean slope as a lower bound on real-world messiness, not a promise.

2

The Graveyard

  • "Steering edits the weights."
    It edits activations at inference. Weights are untouched; remove the hook and the model is exactly as before.
  • "Any direction steers the concept."
    Only the readout-aligned component does. A direction ⊥ to w moves nothing — window 0 proves it.
  • "Bigger α is always more concept."
    amber True in this closed-form model; in real nets large α breaks fluency. Off-manifold ≠ more concept.
6

The Tamper

Swap the difference-of-means lever for a random fixed-seed direction. Steering no longer moves the concept up-and-linear — the witness in window 7 catches it live.