A small mechanism found inside language models that spots a repeated pattern and continues it: if ‘Mr’ was followed by ‘Dursley’ earlier in the text, then when ‘Mr’ appears again, it predicts ‘Dursley’. Researchers think mechanisms like this are a large part of how models pick up patterns from the text in front of them without retraining, and they are one of the best understood examples of what happens inside a model.
Technical view
A two-head circuit described by Anthropic interpretability researchers, first in ‘A Mathematical Framework for Transformer Circuits’ (2021) and in depth by Olsson and colleagues in 2022. A previous-token head in an earlier layer writes each token’s predecessor into the residual stream; the induction head then matches the current token against those records and copies the token that followed, so a sequence A, B, then A again predicts B. Induction heads form in a sharp phase change early in training, coinciding with a jump in in-context learning. The authors argue they may account for much in-context learning, but the evidence is strongest for small models and partly correlational.
Related
Part of AI Terminology Index, section: Language Models Specifically.