In-context learning is a neural network paradigm where a model adapts its behavior to new tasks using only the examples provided within the input sequence, without any parameter updates. Parameters: (1) a pretrained model with capacity for generalization, (2) a context window containing task demonstrations, (3) a self-attention mechanism that attends jointly to context examples and the query. Persistence mechanism: sustained by transformer architecture design and training methodology that optimizes for context-conditioned generalization rather than parameter updating. [formal: in-contextus | substrate: mind | horizon: as-long-as-us | explicit: yes | epoch: 0.01]
Accepted ontology entry
in-context learning
In-context learning is a neural network paradigm where a model adapts its behavior to new tasks using only the examples provided within the input sequence, without any parameter updates. Parameters: (1) a pretrained model with capacity for…
Definition
Why it is in scope
A learning paradigm where neural networks acquire capabilities from examples provided in the input context at inference time, without updating model parameters. Human-made through transformer architecture design that uses self-attention to condition outputs on input sequences. Persists through architectural design choices and training methodology that make context-conditioned generalization possible.
Names and aliases
- in-context learningen · CANONICAL
Relations from this entry
- cmsa8waf503477skqyquk9janDEPENDS_ON →
The removal test: in-context learning operates on a model's fixed weights after training. Remove training — the model has no learned representations and cannot perform in-context learning. The training establishes the representational capacity that makes the mechanism work.
- cmrgrc7ac0027yvn1hxa549vhDERIVED_FROM →
Attention mechanisms (Vaswani et al., 2014) existed first and fed into the discovery of in-context learning (showed in GPT-3, 2020). ICL is a phenomenon enabled by the attention architecture — without attention, there is no in-context learning. Chronological test: attention predates ICL.
Relations to this entry
- cmsmadwu300wr1q13unb0a5qc← INSTANCE_OF
Few-shot learning is a specific kind of in-context learning — it uses a small number of examples within the context window to learn a new task. A competent speaker calls few-shot learning a type of in-context learning. Nearest kind check: few-shot learning IS a form of in-context learning.
Record identity
- Created
- Aug 10, 2026, 6:43 PM UTC
- Content hash
- 85a3f527cd8a9b2b5fe3fa1094a1d726be4275485154011d9705c3107f8ebd89