A trained experiment-design policy learns to take advice at deployment: Imbue turns expert priors and past experiments into droppable input tokens
Synopsis
Imbue lets an already-trained Bayesian experimental design policy take in external knowledge available only at deployment without retraining: user-specified priors and observations from previous instances of the experiment are encoded as extra input tokens, processed in-context with the accumulated observations by a shared Transformer, and a reliability filter scores each token against those observations and excludes inconsistent ones. Across four benchmarks it accelerates early-stage information acquisition when the external knowledge is reliable and stays close to the no-external baseline when it is not.
Interpretation
A trained Bayesian experimental design policy can now accept external knowledge at deployment without retraining; that knowledge enters as additional input tokens and is processed by the same Transformer alongside the accumulated experimental observations. Earlier amortized Bayesian experimental design policies were trained to be optimal in expectation under the training prior and had no input pathway for an experimenter to supply information about the current experiment, so absorbing such information required retraining. On the CES task, ALINE*, which concatenates previous instances to the history only at deployment, falls below the original ALINE under both relevant and irrelevant inputs, showing that untrained concatenation disrupts the policy.
A reliability filter scores each external token against the observations gathered so far and excludes tokens inconsistent with them, so unreliable external information does not pull the policy away from what the observations support. Concatenating external inputs directly invites negative transfer: a user prior may place probability mass in the wrong region of parameter space, and previous instances may come from a population that does not match the current one. On CES, Imbue exceeds ALINE under relevant instances and stays close to it under irrelevant ones; a staged ablation shows mixed-regime training plus the instance-token interface provide most of the gain, with the filter further improving mean sPCE under biased and irrelevant inputs.
On a psychometric task, Imbue exploits the previous-instance channel that the classical methods do not natively accept. The no-external references ALINE, QUEST+ and Psi-marginal have no such input pathway. With relevant previous instances, Imbue attains lower joint RMSE on the threshold–slope marginals than the no-external baselines from the second step onward; with irrelevant instances it tracks those baselines closely.
Perspective
The demonstrated scope is simulated experiments on four standard Bayesian experimental design benchmarks: location finding, CES, 1D GP active learning and a psychometric task, with external knowledge supplied as parameter-wise histogram priors or as design–outcome sequences from previous instances.
Priors enter as parameter-wise marginals, so dependencies between parameters are not represented; likelihood, design and observation spaces are assumed to stay within the trained task family; and the filter is less reliable in the first few rounds, when the limited experimental history makes informative and misleading external inputs hard to distinguish.