rome-pearl-2001-paired-interventions
IN premise — summaries/2026/08/24/meng-2022-rome-s5-conclusion.md
Created 2026-08-25T02:58:16+00:00
ROME's causal analysis of hidden states uses paired interventions (causal indirect effects) rooted in Pearl (2001) to measure the causal contribution of individual hidden state vectors.
Summary
ROME determines which internal components of a neural network are actually responsible for storing a given fact by running controlled causal tests, not just checking for correlations. This means the system's surgical edits are grounded in Pearl's formal causal inference framework, so it can say with stronger justification that a specific hidden state vector is the cause of a model's answer rather than merely being associated with it.