Tempered SMC unifies sampling

- Researchers proposed Tempered Sequential Monte Carlo (TSMC) to combine sampling and exploration inside differentiable dynamics using a control-as-inference formalism that scales. (x.com) - The TSMC framing merges exploration with gradient-based updates, offering a unified method to sample rare trajectories in differentiable physics models and control tasks. (x.com) - That approach could narrow the gap between amortized inference and planner-style exploration in continuous systems. (x.com)