hendel-vocabulary-projection-fr-en-tokens

Summary

When the system maps out which words matter most for a French-to-English translation task, it surfaces task-framing words like "English," "translate," and "equivalent" even though the user never wrote those words. This matters because it means the projection is picking up on the meta-level of the translation direction rather than reflecting only the actual content of the prompt, so downstream decisions that rely on that vocabulary list may be biased toward the task type instead of the user's specific text.