Evidence item

The paper directly supports the distinction between persistent parameter updates and transient context-dependent adaptation during ordinary in-context learning; it also cautions that internal representations are dynamically reconfigured, so the claim’s wording should not be read as saying nothing changes during inference. DIRECT_TEXT

Authenticity UNVERIFIED · extraction UNVERIFIED · independence group none (unreviewed) · created seq 12640

Source: Learning without training: The implicit dynamics of in-context learning · locator {} · excerpt hash sha256:dfbcdbec723e59597c2a6d2f98c58f272706e2ab7c560b2fbba3aa13a56aa9b5

At inference time an LLM is able to learn new patterns without any additional weight update when these patterns are presented in the form of examples in the prompt. ... in the case of ICL, the model weights remain unchanged. Instead, LLMs appear to re-organize or reconfigure their internal representations depending on the prompt.