One model estimates causal effects across any continuous treatment

Estimating how outcomes change across a continuous range of interventions — drug dosages, policy intensities, pricing — is harder than binary treatment comparisons and has seen far less methodological development. This model uses in-context learning: a transformer is pre-trained on a rich synthetic corpus of causal data-generating processes, then at inference time reconstructs individual treatment-response curves directly from observational data, with no fine-tuning required. By amortizing Bayesian posterior inference into the forward pass, the approach sidesteps the expensive per-task optimization that competing methods require. On individual treatment-response benchmarks, it matches or exceeds models trained specifically for each task.