World Models with Configurable Perception

arXiv 2026 31.3 method

TLDR

ConfigWM introduces task-conditioned state extraction for world models, filtering task-irrelevant visual information to improve planning robustness in noisy environments.

Reasoning

The paper presents a novel mechanism for configurable perception in world models and provides empirical evaluations showing consistent improvements over baselines. Its main weakness is the lack of explicit real-world benchmark evidence, with results confined to simulated environments.

Read-first score

Read-first score 31.3, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 31.

Recency 5%

100

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

Methodology quality 15%

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

Reproducibility 15%

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

Topical relevance 25%

44.3

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

Graph prestige 15%

5.7

PageRank-style prestige over the local citation graph built from available OpenAlex references. local_pagerank=0.0016

Citation impact 15%

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. cited_by_count=0

Citation velocity 10%

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

Field roles

Frontier

Rank sensitivity

Stability: volatile; rank range: 214.

Keyword Scores

world model

world dynamics prediction

model-based reinforcement learning world model

interactive world model

world simulator

generative world model

video world model

Deep Analysis

Innovations

Introduces ConfigWM, a world model that achieves configurable perception through task-conditioned state extraction.
Uses language task descriptions as configuration signals to steer visual embeddings into task-specific state spaces, preserving task-relevant information while suppressing task-irrelevant variations.
Learns action-conditioned dynamics in the task-specific state space for planning under noisy observations.

Methodology

ConfigWM uses a language description of the task as a configuration signal to map visual embeddings into a task-specific state space, selectively retaining task-relevant information and suppressing distractors. Action-conditioned dynamics are then learned in this state space for planning. The approach is evaluated in environments with increasing levels of distraction, comparing planning success rate, physical-quantity probing, and distracted-to-clean retrieval against existing methods.

Key Results

ConfigWM remains comparatively stable as environments become noisier and outperforms the strongest competing method by 22.3% on average across diverse environments. It also achieves gains of +0.365 Pearson r in physical-quantity probing and +75.5% in distracted-to-clean retrieval.

Resources

Paper