According to research published on arxiv.org and accepted to NeurIPS 2026, masked diffusion language models (MDLMs) are emerging as an effective alternative to autoregressive models for training AI agents in reinforcement learning environments.
The research introduces a formalization of text-based world modeling as a “steerable transition-dynamics problem” and curates 239,403 grounded state-action trajectories spanning nine open-source environments and twelve frontier model families, according to arxiv.org. The study demonstrates that MDLMs achieve “better coherence, groundedness, and empirically validated rollout diversity than LLMs over 4x their parameter size, at comparable inference latency” through bidirectional anchor-aware denoising.
In zero-shot transfer tests across three out-of-distribution environments (ScienceWorld, ALFWorld, AppWorld) using three agent backbones (LFM2.5, Qwen3, Mistral), the approach achieved “up to 47% absolute gains over baselines without environment-specific fine-tuning,” according to the source.
A related study on arxiv.org introduces Plan-and-Patch, a framework using diffusion language models for agentic planning. According to this research, on Natural Plan without task-specific training, diffusion achieved 53.7% plan repair success compared to 27.0% for autoregressive models. After task-specific training on ALFWorld and TextCraft benchmarks, diffusion models reduced “mean plan-generation latency by 39-46% relative to AR” while maintaining similar success rates.
Both datasets and training code have been open-sourced, according to arxiv.org.