intervention-context-top1-next-word-criterion

IN premise — summaries/2026/08/24/park-2023-linear-representation-s10-the-lion-in-many-cultures-is-considered-the.md

Created 2026-08-24T17:11:02+00:00

In Park et al. (2023), intervention contexts are retained only if the target word is the top-1 next-token prediction of the model (argmax(P(next_token|context)) == target); contexts where the target is a lower-ranked candidate are discarded to ensure valid intervention measurement.

Summary

This sets a quality gate for testing how a language model processes words: you only measure the effect of tampering with the model's internal state when the model would have naturally chosen that word as its single best guess. If you included contexts where the word was just one of many plausible options, any measured effect could be noise rather than a true signal of what the model actually relies on.