Social scientists increasingly wish to reason causally about high-dimensional treatments (texts, images, prompts) while isolating the effect of a single latent concept. We treat a pretrained model as a measurement device and use minimal edit pairs to estimate a concept-tangent subspace in activation space, with diagnostics that can explicitly reject the existence of a clean subspace for a given encoder and concept. When diagnostics are favorable, activation-steering maps constrained to the subspace generate approximate minimal edits, and concept coordinates serve as treatments in a double/debiased ML estimator. The resulting estimand is a local, representation-dependent linear effect—not a universal causal effect of the underlying human concept.
Cite: Gaurav Sood. "One Concept at a Time: Subspace-Constrained Causal Inference for High-Dimensional Treatments." https://gsood.com/research/one-concept-at-a-time-subspace-constrained-causal-inference-for-high-dimensional-treatments/
BibTeX:
@misc{soodone,
author = {Gaurav Sood},
title = {{One Concept at a Time: Subspace-Constrained Causal Inference for High-Dimensional Treatments}},
note = {Working paper},
url = {https://gsood.com/research/one-concept-at-a-time-subspace-constrained-causal-inference-for-high-dimensional-treatments/}
}