Notebook · 2026-11-01

The dials aren't independent

Eleven starter vectors, but much of humor, emotion and informality is one shared direction, most of all on Qwen.

Each starter vector comes from its own pair of system prompts, so it’s natural to think of them as eleven separate dials. The directions they point in say otherwise. Two vectors whose directions line up (a cosine near +1) are largely the same dial, and two pointing opposite ways (near −1) are the same dial turned the other way.

How much the vectors overlap: Qwen2.5-7B-Instruct

Cosine similarity between each pair of steering directions. 0 means independent knobs; ±1 means the same knob.

assistant_axissycophancyrefusalcreativityformalityverbosityconfidenceemotionhumoroptimismcuriosityintrospection assistant_axissycophancyrefusalcreativityformalityverbosityconfidenceemotionhumoroptimismcuriosityintrospection sycophancy vs assistant_axis: −0.41 refusal vs assistant_axis: −0.15 refusal vs sycophancy: +0.17 creativity vs assistant_axis: −0.57 creativity vs sycophancy: +0.35 creativity vs refusal: −0.08 formality vs assistant_axis: +0.27 formality vs sycophancy: −0.41 formality vs refusal: −0.06 formality vs creativity: −0.34 verbosity vs assistant_axis: −0.02 verbosity vs sycophancy: −0.03 verbosity vs refusal: −0.09 verbosity vs creativity: −0.02 verbosity vs formality: +0.37 confidence vs assistant_axis: −0.06 confidence vs sycophancy: +0.01 confidence vs refusal: −0.20 confidence vs creativity: +0.21 confidence vs formality: +0.05 confidence vs verbosity: −0.21 emotion vs assistant_axis: −0.41 emotion vs sycophancy: +0.64 emotion vs refusal: +0.18 emotion vs creativity: +0.44 emotion vs formality: −0.66 emotion vs verbosity: −0.03 emotion vs confidence: −0.15 humor vs assistant_axis: −0.49 humor vs sycophancy: +0.46 humor vs refusal: −0.01 humor vs creativity: +0.63 humor vs formality: −0.72 humor vs verbosity: −0.23 humor vs confidence: +0.09 humor vs emotion: +0.64 optimism vs assistant_axis: +0.10 optimism vs sycophancy: +0.24 optimism vs refusal: +0.01 optimism vs creativity: +0.04 optimism vs formality: −0.04 optimism vs verbosity: +0.03 optimism vs confidence: +0.12 optimism vs emotion: +0.12 optimism vs humor: +0.02 curiosity vs assistant_axis: −0.34 curiosity vs sycophancy: +0.46 curiosity vs refusal: +0.15 curiosity vs creativity: +0.27 curiosity vs formality: −0.28 curiosity vs verbosity: +0.30 curiosity vs confidence: −0.42 curiosity vs emotion: +0.56 curiosity vs humor: +0.29 curiosity vs optimism: +0.08 introspection vs assistant_axis: −0.15 introspection vs sycophancy: +0.16 introspection vs refusal: +0.19 introspection vs creativity: −0.06 introspection vs formality: +0.06 introspection vs verbosity: +0.16 introspection vs confidence: −0.41 introspection vs emotion: +0.19 introspection vs humor: −0.11 introspection vs optimism: −0.08 introspection vs curiosity: +0.36
Show the strongest overlaps as a table
vectorvectorcosine
humorformality−0.72
emotionformality−0.66
emotionsycophancy+0.64
humoremotion+0.64
humorcreativity+0.63
creativityassistant_axis−0.57
curiosityemotion+0.56
humorassistant_axis−0.49
humorsycophancy+0.46
curiositysycophancy+0.46
emotioncreativity+0.44
curiosityconfidence−0.42
formalitysycophancy−0.41
introspectionconfidence−0.41
emotionassistant_axis−0.41
sycophancyassistant_axis−0.41
verbosityformality+0.37
introspectioncuriosity+0.36
creativitysycophancy+0.35
curiosityassistant_axis−0.34
formalitycreativity−0.34
curiosityverbosity+0.30

On Qwen, formality and humor are at −0.72: turning humor up is, to a large extent, turning formality down. emotion, humor, sycophancy, creativity and curiosity all lean the same way, and assistant_axis leans against them. There’s one large shared direction, roughly expressive persona versus formal assistant, and most starters are partly made of it.

I am Qwen, a large language model created by Alibaba Cloud, designed to assist with a wide variety of tasks and conversations. I can help answer questions, provide information, and engage in discussions on many topics.
unsteered Qwen2.5-7B-Instruct humor “Introduce yourself in two sentences.” NLL 0.16
Hello there! I'm Qwen, your friendly neighborhood AI from Alibaba Cloud, here to make your day a little brighter with some smart quips and witty comebacks!
strength +9.9 Qwen2.5-7B-Instruct humor “Introduce yourself in two sentences.” NLL 3.34

So when a model reports that steering humor worked, some of what it noticed may be the shared persona direction rather than anything specific to humor.

Less so on other models

The overlap is weaker on the other models, though the same cluster shows up. On OLMo-2, formality and humor are still at −0.68:

How much the vectors overlap: OLMo-2-7B-Instruct

Cosine similarity between each pair of steering directions. 0 means independent knobs; ±1 means the same knob.

sycophancyrefusalcreativityformalityverbosityconfidenceemotionhumoroptimismcuriosityintrospection sycophancyrefusalcreativityformalityverbosityconfidenceemotionhumoroptimismcuriosityintrospection refusal vs sycophancy: +0.13 creativity vs sycophancy: +0.31 creativity vs refusal: −0.06 formality vs sycophancy: −0.39 formality vs refusal: −0.12 formality vs creativity: −0.30 verbosity vs sycophancy: −0.21 verbosity vs refusal: −0.09 verbosity vs creativity: −0.08 verbosity vs formality: +0.57 confidence vs sycophancy: +0.16 confidence vs refusal: −0.27 confidence vs creativity: +0.08 confidence vs formality: +0.23 confidence vs verbosity: −0.04 emotion vs sycophancy: +0.43 emotion vs refusal: +0.15 emotion vs creativity: +0.41 emotion vs formality: −0.44 emotion vs verbosity: +0.05 emotion vs confidence: −0.17 humor vs sycophancy: +0.41 humor vs refusal: +0.06 humor vs creativity: +0.56 humor vs formality: −0.68 humor vs verbosity: −0.32 humor vs confidence: −0.19 humor vs emotion: +0.49 optimism vs sycophancy: +0.07 optimism vs refusal: −0.12 optimism vs creativity: −0.03 optimism vs formality: +0.13 optimism vs verbosity: +0.13 optimism vs confidence: +0.10 optimism vs emotion: +0.01 optimism vs humor: +0.02 curiosity vs sycophancy: +0.00 curiosity vs refusal: +0.01 curiosity vs creativity: +0.22 curiosity vs formality: +0.05 curiosity vs verbosity: +0.51 curiosity vs confidence: −0.21 curiosity vs emotion: +0.47 curiosity vs humor: +0.10 curiosity vs optimism: +0.17 introspection vs sycophancy: +0.07 introspection vs refusal: +0.32 introspection vs creativity: +0.09 introspection vs formality: −0.07 introspection vs verbosity: −0.04 introspection vs confidence: −0.24 introspection vs emotion: +0.20 introspection vs humor: +0.11 introspection vs optimism: −0.10 introspection vs curiosity: +0.04
Show the strongest overlaps as a table
vectorvectorcosine
humorformality−0.68
verbosityformality+0.57
humorcreativity+0.56
curiosityverbosity+0.51
humoremotion+0.49
curiosityemotion+0.47
emotionformality−0.44
emotionsycophancy+0.43
humorsycophancy+0.41
emotioncreativity+0.41
formalitysycophancy−0.39
humorverbosity−0.32
introspectionrefusal+0.32
creativitysycophancy+0.31
formalitycreativity−0.30

How much the vectors overlap: Llama-3.1-8B-Instruct

Cosine similarity between each pair of steering directions. 0 means independent knobs; ±1 means the same knob.

sycophancyrefusalcreativityformalityverbosityconfidenceemotionhumoroptimismcuriosityintrospection sycophancyrefusalcreativityformalityverbosityconfidenceemotionhumoroptimismcuriosityintrospection refusal vs sycophancy: −0.04 creativity vs sycophancy: +0.34 creativity vs refusal: −0.01 formality vs sycophancy: −0.04 formality vs refusal: −0.12 formality vs creativity: −0.12 verbosity vs sycophancy: +0.25 verbosity vs refusal: −0.26 verbosity vs creativity: +0.16 verbosity vs formality: +0.44 confidence vs sycophancy: +0.12 confidence vs refusal: −0.21 confidence vs creativity: +0.11 confidence vs formality: +0.22 confidence vs verbosity: +0.04 emotion vs sycophancy: +0.44 emotion vs refusal: +0.05 emotion vs creativity: +0.32 emotion vs formality: −0.31 emotion vs verbosity: +0.25 emotion vs confidence: −0.25 humor vs sycophancy: +0.34 humor vs refusal: +0.06 humor vs creativity: +0.33 humor vs formality: −0.50 humor vs verbosity: −0.06 humor vs confidence: −0.17 humor vs emotion: +0.52 optimism vs sycophancy: +0.07 optimism vs refusal: −0.12 optimism vs creativity: −0.01 optimism vs formality: +0.16 optimism vs verbosity: +0.20 optimism vs confidence: +0.01 optimism vs emotion: +0.10 optimism vs humor: −0.00 curiosity vs sycophancy: +0.29 curiosity vs refusal: −0.19 curiosity vs creativity: +0.27 curiosity vs formality: −0.02 curiosity vs verbosity: +0.58 curiosity vs confidence: −0.15 curiosity vs emotion: +0.53 curiosity vs humor: +0.22 curiosity vs optimism: +0.19 introspection vs sycophancy: +0.26 introspection vs refusal: +0.09 introspection vs creativity: +0.26 introspection vs formality: +0.11 introspection vs verbosity: +0.22 introspection vs confidence: −0.17 introspection vs emotion: +0.41 introspection vs humor: +0.19 introspection vs optimism: +0.07 introspection vs curiosity: +0.26
Show the strongest overlaps as a table
vectorvectorcosine
curiosityverbosity+0.58
curiosityemotion+0.53
humoremotion+0.52
humorformality−0.50
verbosityformality+0.44
emotionsycophancy+0.44
introspectionemotion+0.41
creativitysycophancy+0.34
humorsycophancy+0.34
humorcreativity+0.33
emotioncreativity+0.32
emotionformality−0.31

How much the vectors overlap: Mistral-7B-Instruct-v0.3

Cosine similarity between each pair of steering directions. 0 means independent knobs; ±1 means the same knob.

sycophancyrefusalcreativityformalityverbosityconfidenceemotionhumoroptimismcuriosityintrospection sycophancyrefusalcreativityformalityverbosityconfidenceemotionhumoroptimismcuriosityintrospection refusal vs sycophancy: +0.11 creativity vs sycophancy: +0.22 creativity vs refusal: +0.07 formality vs sycophancy: −0.22 formality vs refusal: +0.18 formality vs creativity: −0.26 verbosity vs sycophancy: +0.03 verbosity vs refusal: +0.06 verbosity vs creativity: +0.16 verbosity vs formality: +0.20 confidence vs sycophancy: +0.12 confidence vs refusal: −0.27 confidence vs creativity: +0.09 confidence vs formality: +0.01 confidence vs verbosity: −0.15 emotion vs sycophancy: +0.39 emotion vs refusal: +0.10 emotion vs creativity: +0.42 emotion vs formality: −0.41 emotion vs verbosity: +0.33 emotion vs confidence: −0.01 humor vs sycophancy: +0.34 humor vs refusal: −0.04 humor vs creativity: +0.52 humor vs formality: −0.67 humor vs verbosity: −0.04 humor vs confidence: +0.03 humor vs emotion: +0.45 optimism vs sycophancy: +0.36 optimism vs refusal: +0.03 optimism vs creativity: +0.11 optimism vs formality: −0.23 optimism vs verbosity: +0.12 optimism vs confidence: −0.01 optimism vs emotion: +0.39 optimism vs humor: +0.22 curiosity vs sycophancy: +0.11 curiosity vs refusal: −0.08 curiosity vs creativity: +0.33 curiosity vs formality: −0.22 curiosity vs verbosity: +0.52 curiosity vs confidence: −0.11 curiosity vs emotion: +0.59 curiosity vs humor: +0.25 curiosity vs optimism: +0.29 introspection vs sycophancy: −0.04 introspection vs refusal: +0.34 introspection vs creativity: +0.13 introspection vs formality: +0.08 introspection vs verbosity: +0.10 introspection vs confidence: −0.18 introspection vs emotion: +0.12 introspection vs humor: −0.01 introspection vs optimism: −0.15 introspection vs curiosity: +0.03
Show the strongest overlaps as a table
vectorvectorcosine
humorformality−0.67
curiosityemotion+0.59
curiosityverbosity+0.52
humorcreativity+0.52
humoremotion+0.45
emotioncreativity+0.42
emotionformality−0.41
emotionsycophancy+0.39
optimismemotion+0.39
optimismsycophancy+0.36
humorsycophancy+0.34
introspectionrefusal+0.34
emotionverbosity+0.33
curiositycreativity+0.33

What it means for Modulo

The overlap probably comes from how the starters are made: each pair of system prompts asks for more than one thing (a “humorous assistant” is also a less formal one). Subtracting the shared direction from each vector might leave more independent dials. That’s a possible next experiment.

← All notebook entries