Triangulation as an Acceptance Rule for Multilingual Mechanistic Interpretability

TL;DR

This paper introduces triangulation, a causal and cross-environment standard for validating mechanistic explanations in multilingual language models, ensuring explanations are robust across languages and perturbations.

Contribution

It formalizes triangulation as an acceptance rule involving necessity, sufficiency, and invariance, and applies it to automatic circuit discovery for multilingual models.

Findings

Triangulation filters spurious circuits that fail cross-lingual invariance.

The method is validated across multiple model families and language pairs.

It provides a falsifiable standard for mechanistic interpretability.

Abstract

Multilingual language models achieve strong aggregate performance yet often behave unpredictably across languages, scripts, and cultures. We argue that mechanistic explanations for such models should satisfy a \emph{causal} standard: claims must survive causal interventions and must \emph{cross-reference} across environments that perturb surface form while preserving meaning. We formalize \emph{reference families} as predicate-preserving variants and introduce \emph{triangulation}, an acceptance rule requiring necessity (ablating the circuit degrades the target behavior), sufficiency (patching activations transfers the behavior), and invariance (both effects remain directionally stable and of sufficient magnitude across the reference family). To supply candidate subgraphs, we adopt automatic circuit discovery and \emph{accept or reject} those candidates by triangulation. We ground triangulation in causal abstraction by casting it as an approximate transformation score over a distribution of interchange interventions, connect it to the pragmatic interpretability agenda, and present a comparative experimental protocol across multiple model families, language pairs, and tasks. Triangulation provides a falsifiable standard for mechanistic claims that filters spurious circuits passing single-environment tests but failing cross-lingual invariance.