PRISM: PRior-guided Imagination Sampling in world Models

arXiv 2026 58 method

TLDR

PRISM extracts action priors from a world model's frozen encoder to guide sampling in model-based planning, improving success rates with minimal architectural overhead.

Reasoning

The paper presents a simple, task-agnostic method for guiding action sampling using a prior derived from the world model's own representations, avoiding complex external models. However, the abstract cuts off the quantitative results, and there is no mention of real-world experiments or benchmarks, limiting the assessment of practical impact.

Read-first score

Read-first score 58, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 43.

Recency 6%

100

Uses a gentle age decay so recent papers surface without erasing older foundations. 2026

Citation impact 18%

96.9

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. citation_normalized_percentile=0.96901283

Topical relevance 29%

61.4

Uses existing LLM keyword relevance scores normalized to 0-100. world model,world simulator,generative world model,interactive world model,video world model,world dynamics prediction,model-based reinforcement learning world model

Methodology quality 18%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=dataset

Reproducibility 18%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=code,dataset

Citation velocity 12%

Citation velocity estimates citations per publication-year to reduce old-paper bias. velocity=0.00

Field roles

FoundationFrontierBridge

Rank sensitivity

Stability: volatile; rank range: 478.

Keyword Scores

world model

model-based reinforcement learning world model

world dynamics prediction

world simulator

generative world model

interactive world model

video world model

Deep Analysis

Innovations

Proposes a task-agnostic framework that extracts action prior from the same dataset and learned representations of the world model, avoiding architectural bloat from independent visual encoders or large-scale VLMs.
Attaches a lightweight MLP directly to a frozen JEPA-style encoder to predict a state-conditioned Gaussian prior.
Fuses the prior into the planner's sampling distribution via a precision-weighted Product-of-Gaussians update, a parameter-free closed-form integration.
Improves success rates by 35 percentage points on Cube and 32 percentage points on PushT over vanilla world-model-based MPC without significant inference overhead.

Methodology

PRISM builds on a standard JEPA-style latent world model. It attaches a lightweight MLP to the frozen encoder to predict a state-conditioned Gaussian prior. At plan time, this prior is fused into the planner's sampling distribution via a precision-weighted Product-of-Gaussians update, steering the sampling process. The framework is task-agnostic and uses the same dataset as the world model.

Key Results

PRISM improves success rates by 35 percentage points on Cube and 32 percentage points on PushT over vanilla world-model-based MPC, without introducing significant inference overhead.

Resources

Paper