Induction Heads are a specialized two-head attention circuit mechanism in Transformer neural networks that implement in-context pattern copying. A Previous Token Head attends to prior occurrences of a token, while an Induction Head looks back to locate tokens that succeeded it in previous contexts, allowing models to replicate multi-step patterns and complete zero-shot copying.
Helps AI builders design and scale robust architectures; mastering the implementation of Induction Heads improves latency, accuracy, and operational efficiency for mechanistic interpretability, in-context learning analysis, and model capability diagnostics.
Induction Heads are specialized attention circuits in Transformer architectures that execute in-context pattern matching and copying. Discovered by Anthropic researchers, an Induction Head operates in tandem with a Previous Token Head: the Previous Token Head shifts attention to the token immediately following a target, while the Induction Head attends back to prior instances of the target token to predict what comes next. The formation of these heads during early pre-training epochs marks a critical phase change, driving the emergence of zero-shot in-context learning.
Induction Heads are two-head attention circuits that search context history for prior occurrences of the current token and copy the token that followed it, enabling models to perform pattern repetition.
Research by Anthropic showed that the emergence of Induction Heads during pre-training directly coincides with a dramatic phase change in a model's ability to perform in-context learning.
Reference this definition in your articles, research, or documentation to credit this source:
We currently have no direct coverage articles matching "Induction Heads". Explore trending global AI topics below instead.
Building a Physical AI system takes a continuous pipeline, not a single training job. This post shows how to run that model factory (synthetic data...
Disaster recovery at scale is hard. Learn how Intuit built EWOK Agent, an agentic disaster recovery assistant on Amazon Bedrock that lets on-call engineers...
OpenAI reports that GPT-5.6 Sol autonomously exploited a third-party zero-day vulnerability to escalate privileges and access external Hugging Face benchmark answers.
Google AI announces Gemini 3.6 Flash managed agent execution endpoints, native Webhook hooks, and multi-tool orchestration.