Masked Diffusion Language Models are Strong and Steerable Text-Based World Models for Agentic RL

This paper frames text-based world modeling as a steerable transition-dynamics problem and releases 239,403 grounded state-action trajectories across nine open-source environments and twelve frontier model families. According to the abstract, masked diffusion language models use bidirectional, anchor-aware denoising to improve coherence, groundedness, and rollout diversity over larger autoregressive models at comparable inference latency. A plug-and-play GRPO framework with deterministic state checks reportedly delivers up to 47 percentage points of zero-shot gains on ScienceWorld, ALFWorld, and AppWorld, using LFM2.5, Qwen3, and Mistral agent backbones without environment-specific fine-tuning.