Triangulation as an Acceptance Rule for Multilingual Mechanistic Interpretability
Yanan Long
TL;DR
We address the problem of evaluating mechanistic explanations for multilingual language models by introducing triangulation, an acceptance rule that combines necessity, sufficiency, and invariance across predicate-preserving reference families into an approximate causal-abstraction score $T_{ ext{tri}}$ computed over interchange interventions. The framework formalizes a setup with a predicate $Z=\pi(X)$ and environments that vary surface attributes while preserving meaning, and provides a three-stage experimental protocol (discovery, translation-map patching, and triangulation-based acceptance) tested across multiple model families and datasets. Extending to multimodal settings, triangulation applies to vision-language models by fixing the visual predicate and varying textual prompts, enabling disentanglement of visual and linguistic circuits. The approach offers a falsifiable, cross-language criterion that filters spurious surface-cue mechanisms and yields more robust, transferable mechanistic claims for multilingual 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.