Anthropic publishes 'In-Context Learning and Induction Heads'

Anthropic’s interpretability team, led by Catherine Olsson and including Chris Olah, published “In-Context Learning and Induction Heads,” arguing that a specific circuit found inside transformer language models — which they called an “induction head” — is the primary mechanism behind in-context learning, the ability of a model to pick up a pattern or task from examples given in its prompt without any weight updates.

An induction head works by a two-step lookup: one attention head copies information forward from earlier tokens, letting a second head search the context for the last place the current token appeared and predict that whatever followed it before will follow again. The paper argued this literal-copying behaviour generalises into a “fuzzy,” similarity-based form of pattern completion that can explain a much broader range of in-context behaviour, including translation and simple reasoning by analogy.

The evidence offered was mechanistic rather than purely behavioural. The team observed a “phase change” early in training — a visible bump in the loss curve — during which in-context learning ability improved sharply and induction heads formed simultaneously; when the researchers altered the transformer architecture in ways that shifted when induction heads could form, the point at which in-context learning improved moved correspondingly; and ablating induction heads from small models substantially reduced their in-context learning performance. Together, the authors argued, this supported a causal rather than merely correlational link.

The paper extended Anthropic’s Transformer Circuits research agenda, begun a few months earlier with a mathematical framework for analysing attention-only transformers, and became one of the most cited results in mechanistic interpretability — the attempt to explain model behaviour in terms of specific, identifiable computational structures rather than only aggregate benchmark performance. It offered one of the first concrete, falsifiable claims about why language models could learn new tasks from a handful of examples, a capability that by 2022 was already central to how the models were being used in practice.