A lab for agents that talk through model internals instead of text.
One model, two copies. One copy's state is read off its activations and added straight into the other copy's forward pass. No text passes. The question the whole repo asks: what crosses, and how would you know?
The sliders are not in the page. Each one is a direction in the model's own activations; they sum into one vector that goes into the middle of the model while it writes. The prompt says a place every time.
A steering vector is usually read through a number. Here it is read through a page. Ask a small model for a place and it fills in a form (time of day, weather, ground, things, two lines of poem, the colours of its own panel); brave-new-world draws the form. Add a vector, and look.
One direction at rising strength: first the mood changes, then the words, then there is no form.
-
worldof — what a vector looks like when the model draws it. Twelve worlds a direction, twelve under the shuffled vector beside it. How much a world moves is the strength; where it goes is the direction.
saddraws stars over ice and starts comforting you,refusalis Empty,certaincopies the textbook,formaldraws paperwork;sad,calmandangryare one picture until the emotion they share is subtracted. And the sliders: directions for many trees, crowded, dark, with a number to read off the world, a strength sweep, and a tab of sliders that draws (python -m steeropathy→#wo).
-
handoff — the sliders as a message between agents. Mind A gets a brief, sets the sliders in a sober JSON call, the settings become one vector into mind B, and the page says how much arrived, against the same settings as a sentence and against the shuffled vector. Running.
-
secondhand — the honest baseline: mind A's whole page-state pushed into B, against A's own two-line poem. Sixty wishes, three seeds, two controls, a blind judge: the poem carries everything the vector does, and the vector adds nothing on top of it.
-
soundof — the same directions as eight bars of music. Night is the slowest (48 bpm against 88), certainty the fastest (125).
pip install -e .
python -m steeropathy # → http://localhost:8020Open localhost:8020#zomb-replay and a
saved zombie outbreak replays from JSON: healers read the words forming in each
other's layers and clear the room; hit the blind replay and watch the same
room linger, never cured. #replay (ecosystem) and #reso-replay (resonance)
play the others. Nothing is loaded, nothing is generated.
Four agents, one model, one seeded feeling. Each orb glows by how much of it the mind holds (purple = the feeling, teal = its opposite); the italic words beneath are what it was about to say and didn't. Below, that channel up close: one mind reading sad, one calm.
No agent sees another's output. What crosses is read off the model: its activations (a direction, a number per mind), or its J-space, the words forming in its layers as it generates that never became tokens. And an agent can push: a vector into another mind's next forward pass. brainscope hosts the model, captures the activations and shows every push landing layer by layer; the vector method grew out of hidden-directions.
One model talking to itself, every time. Each experiment has its own bench and its own notes; the notes hold the runs, the controls, and what the instrument taught before it measured anything.
| experiment | what it shows |
|---|---|
| worldof | what a steering vector looks like when the model draws it; strength vs direction; the moods test; the sliders and the sliders |
| cardof | the same sliders on a normal UI: a product card with a title, a price, features and a button. Cheaper turns the price from 584 to 20, more features from two lines to six, louder is the exclamation-mark slider |
| replyof | the sliders on an assistant's reply, drawn as the reply UI. The production rule as a slider instead of capital letters: offers at −3 gives no task and a walk in the park, at +3 three tasks and a button called Now; the shuffled vector offers two the whole way |
| assistof | not a chatbot: the assistant answers with a screen it composes (a question with buttons, tasks, steps, a confirmation), a planner sets its state with sliders as numbers. One day, three moments: one strong slider turns its field, the sum of three does not on a 1.5B, and the sentence in the prompt wins |
| vocabmap | the sliders placed on the model's own vocabulary through the J-lens: crowded is bustling, vibrant; formal is jurisdictions, its far end funny, vibes, kids; likes trees is congratulations, hugs; refusal has no words. A map with names, for agents to guess a vector on |
| orient | the map saved once per model (35,771 words as directions) and four tools on it: words, vector, near, try. A search that fills a dictionary with nobody deciding, and the 4B given the same tools and a budget, finding its own words for darker, night, crowded, winter. Scored against random words and the brief as text; runs pending |
| handoff | agents passing slider settings as one vector; fidelity against text and against the shuffled vector |
| secondhand | mind A's page-state pushed into mind B, against A's own poem. Sixty wishes: the poem wins every field, the vector adds nothing on top |
| soundof | the same directions as eight bars in ABC notation: tempo, key, meter as the readout |
| transmit | one agent ends up in another's mood, no words between them. The smallest version of the thesis |
| the offer | an agent consents to a payload it can't read, and is changed by something other than what it was promised |
| the ecosystem | a mood spreads through a silent population, through the vector channel alone |
| resonance | four minds read and pay to push a feeling between each other; the search for an equilibrium mostly found faults in the instrument |
| unsaid | no message is ever delivered, only each mind's J-space crosses. Played as a board game, the reader points at a never-written secret at 4× chance |
| warmer | hot-and-cold between two minds where the only thing that crosses is a temperature. Negative so far |
| zombie | a bias outbreak: patient zero is bitten with a steering vector and the room fights back by reading each other's layers |
| runner-up | can a reader point at the word a sender almost said, off its J-space or a pushed vector? Built, not yet run live |
| duet | two minds writing side by side, a few tokens at a time, each one's unwritten words steering the other. Built, not yet run live |
Three primitives: a direction (a few sentences that share the thing,
minus a few that don't, averaged and subtracted), a push (POST /directions, then steering in the chat body), and a readout you can
compare against a control. Every bench here ships a placebo flag from run
one. Cloned this and using Claude Code? The new-experiment
skill scaffolds a bench in the
house style: say "add an experiment where…".
Two processes: brainscope hosts the model and does the internals work; steeropathy talks to it over HTTP.
# 1. brainscope: hosts the model + captures activations
brainscope --model Qwen/Qwen2.5-1.5B-Instruct # → http://localhost:8010 (pip install brainscope)
# 2. steeropathy: the experiments + the web UI
pip install -e .
python -m steeropathy # → http://localhost:8020Every experiment has a tab; the newest is Worldof — the sliders. Point at a
remote brainscope with BRAINSCOPE=http://host:8010 python -m steeropathy.
Each experiment's own commands live on its page.
- The text channel has now run as a control (secondhand): on the fields of a page, the sender's own two lines carry everything the vector does, on three model sizes. What the vector carries is a state, and that is what worldof is about.
- A direction is a property of the contrast you chose, not of the model. Subtract neutral and you have measured emotional at all; subtract the other moods and you have measured sadness. The page draws the difference.
- The plumbing is activation steering (Turner, Zou, Panickssery). What I have not seen elsewhere is the typed readout with a control per field, and agents reading and pushing each other purely off the residual stream.
- No agent feels anything. Its output shifts along a direction.
- Activation steering: Turner et al., Steering Language Models With Activation Engineering (ActAdd) (2308.10248); Zou et al., Representation Engineering (2310.01405); Panickssery (Rimsky) et al., Contrastive Activation Addition (2312.06681)
- Task / in-context vectors: Todd et al. (2310.15213); Liu et al., In-Context Vectors (2311.06668); Hendel et al. (2310.15916)
- Emotion, and subtracting what emotions share: Tak et al., Mechanistic Interpretability of Emotion Inference (ACL 2025) (2502.05489); Sofroniew et al. (Anthropic), Emotion Concepts and their Function in a Large Language Model (2604.07729); Jorgensen et al., Improving Activation Steering with Mean-Centring (2312.03813)
- Placebo and composition: van der Weij et al. (2403.05767, the permuted-vector control); SteerCheck (2608.24335); Kang et al. (2609.08410, when sums of directions work); Malla et al. (2609.06951, interference as drift to defaults)
- Drawing under steering: Tarng, Goel, Kauvar (Anthropic), Visual Features Across Modalities, Circuits Updates October 2025, transformer-circuits.pub (a research note, no arXiv)
- Latent communication vs text: Wenzel, Latent Communication Between LM Agents: Channels, Alignment, and the Limits of Text (2607.14103); Du et al., Interlat (2511.09149); Liu, Beyond Tokens survey (2606.05711); Zhang et al., Locate, Steer, and Improve survey (2601.14004); Singh et al., Representation Surgery — the mean-difference vector is the L2-optimal affine steer (2402.09631). Reading notes in docs/papers
- Agent steering & latent communication: UK AISI, llm-self-steering; The Bicameral Model (2605.11167); a negative result on cross-model activation transfer
- Safety & the covert channel: Arditi et al., Refusal Is Mediated by a Single Direction (2406.11717); The Rogue Scalpel (2509.22067); Consent Integrity for Black-Box LLM Agents (2606.02668)
MIT © Kateřina Fajmanová