What is in-context learning, and what do we know about how it works?

30-second answerSay your answer out loud first, then reveal.

In-context learning (ICL) is a model's ability to perform a new task from instructions or examples in the prompt, without any weight updates. It emerges from pretraining on diverse text, where documents implicitly contain tasks and patterns. Interpretability research has found mechanisms like induction heads (attention circuits that find a previous occurrence of the current token and copy what followed it) and theories viewing ICL as implicit Bayesian inference or implicit gradient-descent-like computation within the forward pass.

Observations

ICL improves with model scale.
The label format and the input distribution in the examples matter a lot. Some studies found models still perform reasonably even when example labels are randomised, which suggests the examples often help the model locate the task rather than teach it.
Example order and selection affect results (recency bias, majority-label bias).

Mechanistic insight: induction heads (Olsson et al. 2022)

Two-step circuit: a "previous token" head plus an induction head implement pattern completion: "[A] [B] ... [A] → predict [B]".
They appear during training around the same time as a jump in ICL ability.

Theoretical views

Task inference / Bayesian: the prompt narrows down which latent "task" or "concept" from pretraining applies.
Implicit optimisation: for simple settings (e.g. linear regression), transformers can implement gradient-descent-like updates in their activations.

Practical implications

Choose diverse, representative examples; retrieve examples similar to the query (dynamic few-shot).
Keep the format consistent; balance labels.
For many examples (many-shot ICL with long context), performance can keep improving, sometimes rivalling fine-tuning for some tasks.

PreviousHow are reasoning models trained with reinforcement learning? Explain RL with verifiable rewards, GRPO, and reward hacking.

This is what real progress feels like.

What is in-context learning, and what do we know about how it works?

Observations

Mechanistic insight: induction heads (Olsson et al. 2022)

Theoretical views

Practical implications

Related