LLMs Enhance Treatment Effect Estimation with Uncertainty Guidance

This paper introduces CURL (Causal Uncertainty-guided Representation Learning), a novel plug-in adapter designed to improve the estimation of heterogeneous treatment effects, specifically the Conditional Average Treatment Effect (CATE). CATE estimation often struggles with local instability due to implicit semantic relations and higher-order interactions in raw data, especially when learning nuisance structures and effective covariate representations simultaneously. CURL addresses this by leveraging estimator uncertainty to strategically allocate pretrained semantic capacity from a frozen Large Language Model (LLM). It queries the LLM through two role-conditioned prompts, generating distinct assignment- and heterogeneity-oriented representations from observed covariates, which are then processed through separated pathways. This design helps stabilize the joint task of covariate adjustment and treatment-effect heterogeneity learning. Across four benchmarks, CURL consistently enhances the performance of ten different host learners, with analyses supporting the efficacy of its two-channel design.