KAN-Dreamer: Benchmarking Kolmogorov-Arnold Networks as Function Approximators in World Models

arXiv 25.12 2025 62.9 method, benchmark

TLDR

Integrates KANs into DreamerV3 world model, achieving parity in sample efficiency and training speed on a control task.

Reasoning

The paper presents a novel integration of KAN architectures into the DreamerV3 framework, with empirical results showing comparable performance to MLPs on a single DeepMind Control task. Strengths include addressing computational overhead of KANs and providing a preliminary benchmark. Weaknesses are limited evaluation scope (only one environment) and lack of real-world or diverse tasks.

Read-first score

Read-first score 62.9, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 39.

Methodology quality 25%

100

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

Recency 8%

86.7

Uses a gentle age decay so recent papers surface without erasing older foundations. 2025

Topical relevance 42%

55.7

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

Reproducibility 25%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=none

Field roles

FrontierMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 237.

Keyword Scores

world model

model-based reinforcement learning world model

world dynamics prediction

interactive world model

generative world model

world simulator

video world model

Deep Analysis

Innovations

Integrating KAN architectures into the DreamerV3 framework to create KAN-Dreamer
Replacing specific MLP and convolutional components of DreamerV3 with KAN and FastKAN layers
Implementing a tailored, fully vectorized version with simplified grid management for JAX-based World Model
Structured investigation into three subsystems: Visual Perception, Latent Prediction, and Behavior Learning
Demonstrating that adapted FastKAN as a drop-in replacement for Reward and Continue predictors achieves performance parity with MLP-based architecture

Methodology

The authors replace MLP and convolutional components of DreamerV3 with KAN and FastKAN layers, implementing a fully vectorized version with simplified grid management in JAX. They evaluate on the DeepMind Control Suite (walker_walk) by analyzing sample efficiency, training time, and asymptotic performance, comparing against the original MLP-based architecture.

Key Results

Experimental results show that using adapted FastKAN as a drop-in replacement for the Reward and Continue predictors yields performance on par with the original MLP-based architecture, maintaining parity in both sample efficiency and training speed.

Limitations

Preliminary study limited to a single environment (walker_walk) from the DeepMind Control Suite
Only tested on Reward and Continue predictors, not all components of the world model
Computational overhead of KANs remains a concern, though FastKAN mitigates it
Comparison only against MLP baseline, not other KAN variants or MBRL algorithms
Implementation is JAX-specific, limiting generalizability

Resources

Paper