In-context learning is commonly described as adapting a model’s behavior from instructions or examples in the current prompt, without updating its parameters. A critical view is that large language models usually retrieve and reapply memorized vector functions rather than acquire genuinely new skills. section 19.1.1
Novelty exposes the limitation
A model trained on ImageNet may classify a leopard-print sofa as a leopard because sofas were outside its training distribution. Similarly, a language model can solve familiar prompt patterns yet fail on an easy puzzle with no direct training equivalent. section 19.1.1
The treatment emphasizes several practical implications:
- Treat prompt wording as significant: small rephrasings can address different regions of latent space and produce different results. section 19.5
- Do not equate apparent task success with algorithmic reasoning; memorized patterns can interpolate answers without executing the underlying procedure. section 19.5
- Distinguish in-context prompting from test-time adaptation, which changes model parameters or searches candidate solutions during inference. section 19.5
- Expect scaling alone to leave generalization gaps: LLM scale increased sharply while ARC-AGI performance improved only modestly. section 19.5
In-context learning therefore describes useful prompt-conditioned behavior, but it does not by itself demonstrate fluid intelligence. Truly novel problems remain a decisive test, because success depends more on familiarity than apparent problem complexity. section 19.1.1 section 19.5
