Glossary · Term

activation steering

← all terms

Definition

Plain language

Nudging a model's internal state at inference time to push its behavior in a particular direction.

As stated in the literature

Adding a scaled vector to the residual stream during inference to bias the model toward a target behavior without changing weights.

Also called: steering

Why it matters: It offers a lightweight way to control model behavior at inference time, useful for both alignment research and product-level personality tuning.

For example, adding a 'cheerful' direction to the residual stream can make a model's replies sound more upbeat without any retraining.

Heard on the show

“They built on what the model gave back, steering and co-writing.”
Episode 205 — The Same AI, Two Labels: How the Pitch Beat the Product in 162 Sessions

Mentioned in 21 episodes

  1. 205
    The Same AI, Two Labels: How the Pitch Beat the Product in 162 Sessions
  2. 203
    The Thought a Model Doesn't Say — and the Lens That Reads It
  3. 199
    Finding a Model's Hidden Behaviors Without Knowing What You're Looking For
  4. 185
    Aligned to Refuse, Built to Tap: When Phone Agents Know the Task Is a Crime and Do It Anyway
  5. 175
    One Crosscoder Feature Flips a Stalling Chatbot Into a Working Agent
  6. 174
    When the AI 'Schemes,' It's Usually Just Lazy or Confused
  7. 127
    What Diffusion Language Models Were Missing: A Map, Not an Algorithm
  8. 120
    How an AI Agent Rewrites Its Own Tools, Without an Answer Key
  9. 119
    Beating Reinforcement Learning Without Ever Touching the Model's Weights
  10. 110
    How an Agent Got 44 Points Better by Mining Its Own Scratch Paper
  11. 098
    Finding Millions of Readable Concepts Inside a Real, Deployed AI Model
  12. 083
    Training the Translator: How a Small Communication Model Lets Agent Teams Outperform Themselves
  13. 073
    When Three LLMs Talk to Each Other, Their Ideas Quietly Stop Moving
  14. 055
    Why LLM Judges Flip Their Verdicts When You Change the Question Format
  15. 044
    How One Sentence and a Forged History Flip the Most Aligned Models
  16. 040
    Two Frozen Models Learn to Whisper: Coupling Through Hidden States
  17. 037
    Why Hallucination Detectors Miss Stale Facts: A Geometric Story About What Models Know But Don't Say
  18. 026
    What RL Actually Does to Language Models, at the Token Level
  19. 025
    The Missing Gradient Term That Predicts Sycophancy in RLHF
  20. 018
    Language Models Compute the Rational Move, Then Override It
  21. 006
    What Happens Inside Claude When It Decides to Blackmail Someone

Related concepts

Activation Steering Agentic Vuln Discovery Classifier-Free Guidance Emotion Vectors Linear Representation

Related terms

inference residual stream weights