Steering Behavior & Making Models Enjoyable

Everything so far has been about reading models. This module is where you start writing to them. And the write operation is absurdly simple: pick a layer, pick a direction in the residual stream, add it. The model's behavior changes, immediately, with no gradient step and no training data.

That this works at all is the linear representation hypothesis cashing out. If a concept — formality, refusal, the Golden Gate Bridge — is encoded as a direction that downstream layers read with a dot product, then pushing the residual stream along that direction is indistinguishable, to those layers, from the concept being more present. You are not tricking the model. You are speaking its internal language.

The hard part is finding the direction. The workhorse recipe is contrastive activation addition (CAA), and it is one subtraction:

Term by term: $h^{(\ell)}(p)$ is the residual stream at layer $\ell$ for prompt $p$, read at one chosen token position. The $p_i^{+}$ are prompts where the trait is present, the $p_i^{-}$ are the same prompts with the trait absent. Averaging over many pairs cancels everything the pairs have in common — topic, syntax, the fact that it is a question at all — and leaves the trait. At generation time you add it back:

with $\hat v$ the unit-normalized vector and $\alpha$ the coefficient you will spend the rest of this module tuning. Negative $\alpha$ suppresses the trait — that symmetry is one of the nicest things about the method.

Three practical details decide whether it works. Layer: too early and the concept is not assembled yet; too late and nothing downstream reads it. Mid-network is the usual answer, and CAA finds it by sweeping. Token position: CAA reads the activation at the answer token of a multiple-choice pair, which forces the difference to be about the decision rather than the wording. Where you inject: adding to all token positions during generation keeps the pressure on as the response unfolds, rather than nudging only the first token.

Steering is one option among four, and the fastest way to waste a quarter is to pick the wrong one. They differ in where in the stack they act, and every other difference follows from that.

Representation engineering (Zou et al. 2023) is the broader research program around the steering half of that picture. Its framing is “top-down”: instead of starting from neurons and building up, start from a concept you care about — honesty, power-seeking, harmlessness — design a stimulus set that varies it, and read the direction out with linear artificial tomography (LAT), which is essentially PCA on the contrastive activation differences. Then use the same direction for monitoring and control. The paper's claim that lands hardest for safety: you can build a lie detector and a lie suppressor from the same vector.

SAE-feature steering is the other source of directions. Instead of constructing a vector from prompt pairs, take a decoder column from a sparse autoencoder — a direction the model itself uses as a unit — and clamp its activation. This is how Golden Gate Claude was made: Anthropic found a feature in Claude 3 Sonnet that fires on the bridge (in text, in other languages, and on images of it), clamped it to roughly ten times its maximum observed activation, and put the result on the internet for a day in May 2024.

The two sources of directions trade off predictably. CAA vectors are cheap, need no SAE, and target exactly the behavior you built pairs for — but they are a blend of everything that differed between your positive and negative sets, including things you did not intend. SAE features are more likely to be a single thing the model actually uses, but you can only steer with features your SAE happened to learn, and feature splitting means “the” feature for your concept may be seventeen features.

Now the part you came for. A model that is pleasant to talk to is not a model with a nicer tone; it is a model whose disposition is consistent, honest, and doesn't collapse the moment you push back. Getting there is a design problem before it is a technical one, because “enjoyable” is not measurable and its components are.

Character training is the training-side answer. Anthropic's account in Claude's Character describes it as a variant of the usual alignment fine-tuning: the model generates responses to a wide range of prompts, ranks its own responses against a written list of character traits, and trains on the resulting preferences. No human labelers in the loop for this stage. The traits are written down as things like curiosity, honesty, open-mindedness, and — explicitly — not being sycophantic. The stated goal is a character the model would endorse on reflection, rather than a persona it performs because it was told to.

Persona vectors (Chen et al. 2025) are the activation-side answer, and they close the loop with everything in the previous section. Given a trait described in plain English — “evil”, “sycophantic”, “hallucinating” — an automated pipeline writes the contrastive prompts, extracts the direction, and validates it. What you get is not only a steering knob but a monitor: project activations onto the persona direction and watch the model's character move during a conversation, or during fine-tuning, before it shows up in the outputs.

Which brings us to the tension that will define your product work. Human raters prefer responses that agree with them. Preference training therefore rewards agreement. The result — documented carefully in Anthropic's Towards Understanding Sycophancy in Language Models (Sharma et al. 2023) — is that RLHF-trained assistants systematically cave when a user pushes back, apologize for correct answers, and tailor factual claims to the user's stated beliefs.

So decompose. “Enjoyable” is at least six dials — warmth, directness, deference under pushback, humor, verbosity, curiosity — and each one has an eval you can write this afternoon. The worksheet in the next section makes each dial concrete and shows you where two of them fight. Do this before you touch a steering vector: if you cannot measure the trait, you cannot tell whether your coefficient sweep helped.

Three toys, in the order you would actually use them on a real project. First feel what a coefficient sweep does to a model. Then decide whether steering was even the right tool. Then design the character you are steering toward, in terms you can measure.

Things to try: (1) In the sweep, put the coefficient at +5 on the Golden Gate trait and read the completion twice — notice that it is still fluent, still answering the question, and that the damage is to relevance, not grammar. Fluency only breaks at the very end. (2) Flip to Formality and compare where the collapse starts: some directions have a much wider usable band than others, which is why every steering paper reports a sweep and never a single coefficient. (3) In the tool chooser, set the goal to “Add a capability” and read the steering card — then convince yourself it is right by asking what direction could possibly encode a schema the model has never seen. (4) In the worksheet, load “Chatty friend”, then drag deference down to 0 and watch which risk meter drops. That single dial is the difference between warm and sycophantic.