hendel-vocabulary-projection-not-verbatim
IN premise — summaries/2026/08/24/hendel-2023-icl-task-vectors-s7-related-work.md
Created 2026-08-25T02:58:04+00:00
Top projected tokens for tasks (e.g., 'participle' for Present→Gerund, 'English' for FR→EN) were not present verbatim in the context, indicating θ encodes an abstracted task representation rather than a copy of the input.
Summary
When the model works on a task, the most prominent concepts in its internal state (like "participle" for verb-tense changes or "English" for translation) often aren't words that actually appeared in the input text. This means the internal representation is capturing the abstract structure of the task being performed, not just echoing surface-level words, which tells us the model is genuinely encoding *what kind of transformation* is required rather than simply copying what it was given.