MDLM, Agentic RL World Model

Results show that Masked Diffusion Language Models (MDLMs) are better suited for text-based world models than autoregressive (AR) LLMs.

• AR methods generate the next state only left-to-right, making it hard to simultaneously fit globally interlocked information such as tool schemas or later-positioned state fields.
• MDLM uses an any-order denoising objective to jointly learn multiple conditioning directions, better preserving global consistency.
• Fine-tuned SDAR-8B and WeDLM-8B outperformed even 4x larger AR baselines on BLEU-1, ROUGE-L, and MAUVE, and showed lower Self-BLEU and higher Distinct-N, indicating reduced prefix mode collapse.
• When MDLM-generated rollouts were used for training with GRPO, up to +15% absolute task-success improvement was observed over AR-generated data on ScienceWorld, ALFWorld, and AppWorld holdouts.
• The results held across 1.2B~7B backbones (LFM2.5, Qwen3, Mistral), and were maintained even in zero-shot transfer settings.