Language-conditioned world model improves policy generalization by reading environmental descriptions

NeurIPSW 25 2025 60.3 method

TLDR

Language-conditioned world model improves policy generalization by grounding language to observations, without planning or expert demonstrations.

Reasoning

The paper proposes LED-WM, a model-based RL approach that uses a language-conditioned world model to improve policy generalization to unseen games, dropping assumptions of planning or expert demonstrations. Strengths include explicit language grounding via attention and empirical generalization results; weaknesses are the lack of real-world validation and reliance on simulated environments.

Read-first score

Read-first score 60.3, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 48.

Recency 8%

86.7

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

Topical relevance 42%

68.6

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 25%

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

Reproducibility 25%

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

Field roles

Frontier

Rank sensitivity

Stability: volatile; rank range: 280.

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

Proposes Language-aware Encoder for Dreamer World Model (LED-WM) with an attention mechanism to explicitly ground language descriptions to entities in the observation.
Model-based reinforcement learning approach that trains a language-conditioned world model through environment interaction and learns a policy from this model without planning or expert demonstrations.
Demonstrates policy improvement via fine-tuning on synthetic test trajectories generated by the world model before real-world deployment.

Methodology

The method builds on DreamerV3, using a language-conditioned world model with a Language-aware Encoder (LED-WM) that employs an attention mechanism to ground language descriptions to entities in observations. The world model is trained through interaction with the environment, and a policy is learned from this model without requiring planning or expert demonstrations. Evaluation is conducted in MESSENGER and MESSENGER-WM environments.

Key Results

Policies trained with LED-WM generalize more effectively to unseen games described by novel dynamics and language compared to other baselines. Additionally, the policy can be improved through fine-tuning on synthetic test trajectories generated by the world model.

Resources

Paper