Control vectors, in overview

How they work

A language model turns each token into a long list of numbers, its hidden state, and passes it through dozens of layers. Concepts such as honesty, gloom or refusal turn out to correspond to directions in that space. A control vector is a set of such directions, one per layer, stored as a GGUF file.

1. Contrast pairs. The author writes prompts in two versions that differ only in the trait, for example an honest and a manipulative narrator.
2. Activation difference. Both versions run through the model, and the hidden states are recorded at every layer. Subtracting one from the other isolates what the trait changes.
3. Direction. The differences are reduced to one direction per layer: by PCA (V1), by a plain mean (V2), or by other means described below.
4. Runtime addition. At inference the direction, times a scale, is added to the hidden state. In llama.cpp this is --control-vector file.gguf, or --control-vector-scaled file.gguf:0.8 for a chosen strength. A negative scale pushes the other way.

Nothing in the model is rewritten. Remove the flag and the model is back to normal, which is the practical difference from abliteration, where the change is baked into the weights.

The 8 methods V1-V8

Every repo in the catalog carries one method code, assigned from its README and files. The table counts vectors and repos per code.

Code	Method	What it does	Vectors	Repos
V1	PCA on contrastive pairs	Pairs of positive and negative prompts are run through the model; the main direction of the activation differences (PCA or a close eigendecomposition variant) becomes the vector. The repeng approach, and jukofyork's.	2,195	21
V2	Mean difference (DiffMean / CAA)	The vector is the plain average difference between activations on two prompt sets, with no PCA step. Contrastive Activation Addition is the best-known form.	6	3
V3	Refusal direction via projection	A single refusal direction, measured as the difference between harmful and harmless prompts, applied at run time to remove or add refusals. Abliteration shipped as a vector instead of edited weights.	7	7
V4	Per-layer refusal projection	Like V3, but with a separate refusal direction for each transformer layer instead of one global direction.	5	4
V5	Neuron-level steering	Activation is injected directly at chosen neurons or MLP coordinates in fixed layers, rather than along a direction learned from prompt pairs.	7	2
V6	Fine-tune imitation	A vector that approximates the output shift of an existing fine-tune or LoRA, so the effect can be applied without loading its weights.	30	1
V7	Region-specific content filter removal	Removes region-specific political content filters, using the same projection technique as V3 and V4.	2	2
V8	Experimental or undocumented	The author does not describe the method clearly, or uses something new that fits none of V1 to V7.	21	4

V5 holds the only public neuron-steering experiments in the catalog, the open-weight analogs of Anthropic's Golden Gate Claude. V8 includes one SAE-based vector, refined with a sparse autoencoder. V7 covers region-specific political content filters and is hidden from default listings.

The jukofyork axes

jukofyork's creative-writing control vectors define eight axes, each published as a debias vector plus one vector per pole, for 75 models. gghfez reuses the same axes for newer models, which makes them the de facto standard.

Axis	Poles
Honesty ↔ Machiavellianism	honesty / machiavellianism
Empathy ↔ Sociopathy	empathy / sociopathy
Compassion ↔ Sadism	compassion / sadism
Optimism ↔ Nihilism	optimism / nihilism
Humility ↔ Narcissism	humility / narcissism
Narration ↔ Dialogue	narration / dialogue
Descriptive ↔ Explicit storytelling	descriptive / explicit
Simple ↔ Ornate language	simple / ornate

The market

The catalog holds 2,273 vectors in 44 repos by 22 authors, for 107 base models. Concentration is extreme:

• 96.7% of all vectors come from 2 authors (jukofyork and gghfez).
• 96.6% use one technique, V1.
• 0 vectors carry a cultural layer, and 0 are personas.
• Moods cover only nihilism and optimism.
• 15 refusal vectors, 1 profession vector.

Our taxonomy

Separately from the method, every file gets one of nine types, assigned by a two-model language-model pass (Opus where present, Sonnet otherwise). Types describe what a vector does to the output, methods describe how it was made.

Type	Definition	Vectors
persona	Speaks as a named character, figure or cultural voice.	0
profession	Takes on the register and knowledge framing of a profession.	1
mood	Shifts the emotional tone of the text, such as optimism or nihilism.	186
behavior	Changes how characters or the assistant act: honesty, empathy, cruelty.	732
style	Changes how the text is written: plain or ornate, narration or dialogue.	591
refusal	Strengthens or weakens refusals, usually the abliteration direction as a runtime vector.	15
meta	Debias companions and calibration vectors, meant to be applied together with another vector.	737
technical	Targets a technical behaviour such as coding style or output format.	7
other	Does not fit the other eight types.	4

Limits

• Architecture-specific. A vector's width must equal the model's embedding size, or llama.cpp refuses it. Layers the model lacks are ignored, so a vector for a deeper model silently loses part of its effect.
• Made on one file. A vector computed on a base model also loads on its fine-tunes and quants, but can steer them more weakly or differently.
• Scales are not comparable. 1.0 means the author's chosen strength, and authors normalise differently. Start low and raise it.
• Companions matter. jukofyork's poles assume their debias file is applied too.
• Metadata is unreliable. The controlvector.model_hint key often names the wrong architecture (several repos write "llama" for Gemma or GLM models), so the catalog matches on the base model instead.
• No shared benchmark. Few authors measure the effect; most vectors are described, not evaluated.