PRISM: PRior-guided Imagination Sampling in world Models

A new framework called PRISM improves model-based planning for continuous-control robots by extracting an action prior directly from a world model, researchers report in a paper submitted in 2026 [1]. The approach lifts success rates by 35 percentage points on one manipulation task and 32 points on another without adding meaningful inference cost [1]. The framework, whose full name is PRior-guided Imagination Sampling in world Models, addresses a bottleneck in how candidate actions are generated during model-based planning [1]. Standard planners often search arbitrarily or use expert demonstrations only to set a sampling mean, discarding the expert’s state-conditioned confidence [2]. PRISM instead attaches a lightweight multi-layer perceptron to the frozen encoder of a JEPA-style latent world model [2]. That head predicts a state-conditioned Gaussian prior over upcoming actions [3]. At plan time, the prior is fused into the planner’s sampling distribution through a precision-weighted Product-of-Gaussians update [2]. The closed-form integration is parameter-free: it makes the prior confident where it is accurate and cedes control where it is uncertain [3]. The researchers tested PRISM on two visual manipulation benchmarks. On the Cube task, success rates rose by 35 percentage points over a vanilla world-model-based model-predictive controller [1]. On PushT, the gain was 32 percentage points [1]. The paper notes that the largest improvements appear at small sample budgets, with multi-seed success gains ranging from 23 to 35 percentage points over an uninformed initialization at matched compute [4]. The same training-and-fusion pipeline transferred to two real-robot setups — a Franka PushT station and an ARX X5 cube manipulator — with no retuning [4]. PRISM’s design avoids the architectural bloat of earlier methods that rely on independent visual encoders or large-scale vision-language models to obtain an action prior [2]. Because the prior is built from the same dataset and the same JEPA encoder the world model already uses, the added perceptual cost is zero [3]. The authors describe the inference overhead as negligible [1]. A separate 2026 paper also uses the PRISM acronym for a real-time multisensory imitation policy validated on a Unitree Go2 quadruped with a 7-DoF arm and a UR5 manipulator, but that work addresses a different problem — single-pass trajectory generation with a batch-global rejection-sampling objective — and is unrelated to the world-model planning framework [5].