Steering Vectors (Activation Steering)

Definition: A steering vector is a direction inside a language model's activations that pushes its output toward a trait, such as more formal, more upbeat, or more likely to refuse. Researchers add the vector while the model writes a reply. This is called activation steering. It changes behavior at run time without retraining the model or changing its weights.

TL;DR: Activation steering edits a model's latent space while it runs. In 2024, researchers found that refusal in 13 open chat models, up to 72B parameters, is controlled by a single direction. Steering is cheap and needs no retraining, but a 2024 review found it unreliable outside its test setup. Compare it with prompt engineering before you reach for it.

Think of a sound mixing desk. The song is already recorded, and you cannot change what the band played. But you can slide one fader up and make the bass louder through the whole track. A steering vector is a fader inside the model. The model's knowledge stays the same. You turn up one quality of what comes out.

In array signal processing, "steering vector" names a different thing: the phase pattern that points an antenna array's beam in one direction, a sense that Emergent Mind's steering vectors topic page covers beside the language model one. This article is about the language model sense.

Why Steering Vectors Matter in 2026

Steering vectors matter because they are the most direct way to change what a model does without retraining it, and they double as a test of whether researchers understand a model's insides. If adding one direction reliably makes a model more honest, more cautious, or more cheerful, that direction carries real meaning for the model. If it does not, the understanding was shallower than it looked.

The field moved fast between 2023 and 2026. Andy Arditi, Oscar Obeso, Aaquib Syed, Daniel Paleka, and colleagues showed in Refusal in Language Models Is Mediated by a Single Direction (June 2024) that in 13 open-source chat models of up to 72B parameters, one direction controls refusal. Remove it and the model stops refusing harmful requests. Add it and the model refuses even harmless ones. In August 2025, Anthropic published Persona vectors, which track traits such as sycophancy and hallucination and can flag when a model drifts toward them during a conversation or during training.

The warnings grew just as fast. In June 2026, Qi Cao and colleagues reported in Activation Steering Induces Emergent Misalignment that steering vectors built for a narrow task can cause misbehavior in broad, unrelated areas. Steering is a sharp tool, and 2026 research is about where it cuts.

How Activation Steering Works

The most common recipe is a difference of means. You collect two sets of prompts that differ in one trait, record the model's activations at one transformer layer for each set, and subtract the average of one set from the average of the other. The result is a direction that points from "less of the trait" to "more of the trait". During generation, you add that direction, scaled up or down, to the model's activations at the same layer.

Three papers set the main methods:

• Activation Addition (ActAdd). Alexander Matt Turner and colleagues showed in Steering Language Models With Activation Engineering (August 2023) that the difference in activations between one contrast pair, such as "Love" and "Hate", is enough to steer sentiment. It needs no backward passes, so it costs about as much as running the model.
• Contrastive Activation Addition (CAA). Nina Panickssery and colleagues averaged the difference over many positive and negative examples in Steering Llama 2 via Contrastive Activation Addition (December 2023). They steered Llama 2 Chat at 7B and 13B on seven behaviors, including sycophancy, hallucination, corrigibility, and refusal, and found that it works on top of fine-tuning and system prompts.
• Representation Engineering. Andy Zou and colleagues framed the wider approach in Representation Engineering: A Top-Down Approach to AI Transparency (October 2023), which reads and controls concepts such as honesty, harmlessness, and power-seeking from population-level representations rather than single neurons.

A fourth route uses sparse autoencoder features as the direction. Anthropic's Golden Gate Claude demo (May 2024) amplified one SAE feature until the model talked about the Golden Gate Bridge in unrelated answers.

Steering at one layer, during generation

  layer 15 activation (before)   [ 0.42  -1.10   0.08   0.73  ... ]
  steering vector  x  strength   [ 0.30   0.05  -0.20   0.10  ... ]  x 4
                                 --------------------------------------
  layer 15 activation (after)    [ 1.62  -0.90  -0.72   1.13  ... ]

  strength  0  -> the model is unchanged
  strength +4  -> noticeably more of the trait
  strength -4  -> noticeably less of the trait
  strength too high -> fluency breaks, output turns odd or repetitive

Ways to Build a Steering Vector

There are five common ways to build a steering vector: take the difference from one contrast pair, average the difference over many pairs, take the main direction of the differences with PCA, optimize the vector directly against target outputs, or reuse a feature a sparse autoencoder found. They trade data needs against precision. Most published work uses the averaged difference.

Anthropic's persona vector pipeline automates the mean difference method. Given only a trait name and a description, it generates prompts that draw out opposite behaviors, such as evil and non-evil replies, and takes the difference in activations between the two. The same team ran experiments on politeness, apathy, humor, and optimism as well as its three main traits.

Steering does not have to be on all the time. Bruce W. Lee, Inkit Padhi, and colleagues introduced Conditional Activation Steering (CAST, ICLR 2025 spotlight), which reads the model's activations for each input and applies a refusal vector only when the input matches a condition, such as hate speech. That turns a blunt fader into a rule: "refuse if the request is about X", with no weight changes.

Steering vs Prompting vs Fine-Tuning

Activation steering sits between prompting and fine-tuning. A prompt changes the input and works on any hosted model. Fine-tuning and RLHF change the weights and need training compute. Steering changes the activations while the model runs, which is cheap and reversible but needs direct access to the model's internals.

For almost everyone who uses a hosted model, prompting is the only option on the table, because hosted APIs do not expose activations. Steering is a research and open-model tool. It matters to non-researchers because it shapes how labs study and adjust the models they ship.

What Steering Can and Cannot Do

Steering vectors can shift tone, sentiment, refusal, and persona traits, but their effect varies from input to input, some concepts steer backward, and strong vectors cost fluency. The honest summary is that steering works often enough to be useful and fails often enough to be risky. Joschka Braun, Dmitrii Krasheninnikov, Usman Anwar, Robert Kirk, Daniel Tan, and David Scott Krueger laid out the problems in A Sober Look at Steering Vectors for LLMs (Alignment Forum, November 23, 2024).

The Sober Look authors made four recommendations: build theory for why and when steering works, evaluate it in realistic settings such as open-ended answers, multi-turn chat, and agents, report side effects, and explore nonlinear methods. Most published steering results still come from multiple-choice tests, which are easier than real use.

Steering also has a constructive side. Anthropic's persona vector work by Runjin Chen, Andy Arditi, Henry Sleight, Owain Evans, and Jack Lindsey (arXiv 2507.21509) used steering toward an unwanted trait during fine-tuning, a "preventative steering" step, to stop the model from picking up that trait, with little loss on a standard knowledge test. The same vectors flagged training data likely to cause a trait.

Steering and the Linear Representation Idea

Steering only works if some concepts sit along straight-line directions in a model's latent space. That idea is the linear representation hypothesis, the same one behind the old "king minus man plus woman" result explained in latent space. Kiho Park, Yo Joong Choe, and Victor Veitch formalized it in The Linear Representation Hypothesis and the Geometry of Large Language Models (ICML 2024) and proved that linear probing and steering connect to the same geometry.

The hypothesis is useful, not complete. Some traits behave like clean directions, as refusal did across 13 models. Others resist, which is one reason the reliability results above are so mixed.

A Short History of Steering Vectors

Steering vectors grew from the 2013 discovery that word vectors carry meaning in their directions. Activation addition for language models arrived in 2023, the refusal direction and the reliability critiques followed in 2024, and the 2025 to 2026 work turned to persona monitoring and to side effects such as emergent misalignment.

How Steering Vectors Show Up in Taskade

Taskade does not steer model activations. Taskade agents run on frontier models from top AI labs, and Taskade does not expose those models' internals. You steer a Taskade agent from the outside, with instructions, knowledge, and tools. That is prompting, the left column of the comparison table above, and it works because it needs no access to a model's hidden states.

A brand voice reply agent is the plain use. Write the tone rules into an AI agent's instructions, attach your style guide and approved replies as knowledge, and connect an automation that sends each new message to the agent for a draft. A teammate approves or edits each draft, and the best edits go back into the example set. Taskade Genesis builds that app from one prompt, and Taskade EVE edits it in chat. To start from a working example, clone the Content Agent Chatbot kit, an app kit with projects, an AI agent, and automations you can clone.

Describe yours and build it free →

Related Concepts

• Latent Space: the space a steering vector moves through
• Sparse Autoencoders: a source of named directions to steer with
• Mechanistic Interpretability: the field that studies why steering works
• Representation Learning: how the directions come to exist
• Latent Reasoning: models that think in the same hidden space a steering vector edits
• Platonic Representation Hypothesis: whether a direction found in one model has a match in another
• Embedding Models: the searchable vectors that share the same geometry
• Prompt Engineering: steering from the outside, through the input
• Fine-Tuning: changing the weights instead of the activations
• Guardrails: the outer controls most deployed systems use
• AI Safety and Alignment: the goal much steering research serves
• Hallucinations: one of the traits persona vectors try to track
• Large Language Models: the models steering is tested on
• Constitutional AI: shaping behavior through training rules instead of activations
• Evals: how to measure whether steering helped or hurt
• Agentic AI Glossary: plain definitions for related agent terms

Frequently Asked Questions About Steering Vectors

What is a steering vector in simple terms?

A steering vector is a direction inside a model's hidden numbers that, when added during a reply, pushes the output toward a trait such as politeness or caution. It works like a fader on a mixing desk. The model keeps everything it knows, and one quality of its output gets louder or quieter.

What is activation steering?

Activation steering is the act of adding a steering vector to a model's activations while it generates text. It changes behavior at run time, with no retraining and no change to the weights. Researchers usually build the vector from the difference between activations on prompts with and without a trait, then add it at one layer.

Is activation steering better than prompting?

Not for most people. Prompting works on any hosted model and is reliable for everyday jobs. Steering needs access to the model's activations, which hosted APIs do not give, and its effect varies by input. Steering is mainly a research tool for open models and for studying how a model represents a trait.

Can steering vectors make a model unsafe?

Yes. In 2024, researchers showed that removing a single refusal direction makes open chat models stop refusing harmful requests. In 2026, a study found that steering vectors built for a narrow task can cause misbehavior in unrelated areas. Steering can also improve safety, as Anthropic's persona vector work showed, so it cuts both ways.

What is the difference between steering and fine-tuning?

Fine-tuning changes the model's weights through training, so the change is permanent until you swap the model. Steering adds a vector to the activations while the model runs, so you can turn it off or change its strength at any time. Steering is cheaper and reversible. Fine-tuning is usually more reliable inside its training domain.

How do you build a steering vector?

Collect prompts that show a trait and prompts that do not, run both through the model, and record the activations at one layer. Subtract the average of one set from the other. That difference is the steering vector. Other methods use a single prompt pair, PCA over the differences, direct optimization against preference pairs, or a feature from a sparse autoencoder.

Who invented steering vectors?

No single person. The idea that concepts sit along directions goes back to word vectors such as word2vec in 2013. Activation addition for language models was popularized by Alexander Matt Turner and colleagues in 2023, and Contrastive Activation Addition by Nina Panickssery and colleagues later that year. Representation Engineering, by Andy Zou and colleagues, framed the wider approach.

Further Reading

• Turner et al., Steering Language Models With Activation Engineering, August 2023
• Panickssery et al., Steering Llama 2 via Contrastive Activation Addition, December 2023
• Zou et al., Representation Engineering, October 2023
• Arditi et al., Refusal in Language Models Is Mediated by a Single Direction, June 2024
• Braun et al., A Sober Look at Steering Vectors for LLMs, November 23, 2024
• Tan et al., Analyzing the Generalization and Reliability of Steering Vectors, July 2024
• Braun et al., Understanding (Un)Reliability of Steering Vectors in Language Models, May 2025
• Park, Choe, and Veitch, The Linear Representation Hypothesis and the Geometry of Large Language Models, ICML 2024
• Anthropic, Persona vectors, August 1, 2025
• Cao et al., Activation Steering Induces Emergent Misalignment, June 2026
• Cao et al., Personalized Steering of Large Language Models: Versatile Steering Vectors Through Bi-directional Preference Optimization, NeurIPS 2024
• Lee et al., Programming Refusal with Conditional Activation Steering, ICLR 2025
• Emergent Mind, Steering Vectors topic overview