A new preprint on arXiv, titled "From Copying to Composition: How Language Models Develop In-Context Learning," takes up a question that accuracy benchmarks often miss: how does contextual information actually steer a model's predictions? The authors note that in-context learning (ICL) allows models to adapt from instructions and examples, but that measuring accuracy alone leaves the internal control mechanism opaque.
The title suggests a developmental story—that models first rely on copying patterns from the prompt and later move toward more compositional reasoning. The abstract's opening frames this as a study of how ICL develops, but the provided text cuts off before any results or methods are described.
Because the source is only a partial abstract, this summary is limited to the stated motivation and the implied trajectory. Readers should consult the full paper for concrete evidence, but the framing points to a useful shift from asking whether ICL works to asking how it works and how that changes over training or scale.