Task-Gated Diverse Counterfactuals

Integrate this into model-based reinforcement learning or diffusion/world-model planning immediately before generating synthetic training data. Maintain either an ensemble of predictors or a neural predictor with a Bayesian last layer. Generate M candidate branches from different initial states, actions, ensemble members, or latent perturbations. For each branch b, collect its feature matrix Phi_b from the world model, estimate a task gate G_b by backpropagating the predicted objective—such as cumulative backlog, reward, or constraint margin—to each queried transition, and set R_b from validation residuals. Begin with S equal to the empty set and greedily recompute the marginal log-determinant gain Delta(b|S) using Cholesky rank-one or block updates; select a fixed batch of K branches. Before adding a branch, apply conflict projection: compare its action distribution to the trusted policy using KL divergence or action norm, reject predicted constraint violations, and optionally project unsafe actions onto the nearest feasible action set. Train the policy or value model only on the surviving fixed batch, avoiding adaptive branch selection during the gradient pass. The log-determinant and greedy rule come from the information objective; gates, thresholds, and covariance estimates are empirical. First test on a 2D obstacle-navigation benchmark with a learned dynamics ensemble, comparing random branches, maximum-uncertainty branches, and optimistic-return branches at equal rollout counts. Measure policy return, constraint violations, synthetic-data bias, and wall-clock cost. Success means higher return and fewer catastrophic rollouts at the same number of model calls, with lower variance across random seeds.