Masked Diffusion Language Models (MDLMs) promise parallel token generation and arbitrary-order decoding, yet it remains unclear to what extent current models truly realize these capabilities. We characterize MDLM behavior along two dimensions—parallelism strength and generation order—using Average Finalization Parallelism (AFP) and Kendall’s τ. We evaluate eight mainstream MDLMs (up to 100B parameters) on 58 benchmarks spanning knowledge, reasoning, and programming. The results show that MDLMs still lag behind comparably sized autoregressive models, mainly because parallel probabilistic modeling weakens inter-token dependencies. Meanwhile, MDLMs exhibit adaptive decoding behavior: their parallelism and generation order vary significantly with the task domain, the stage of reasoning, and whether the output is correct. On tasks that require “backward information” (e.g., Sudoku), MDLMs adopt a solution order that tends to fill easier Sudoku blanks first, highlighting their advantages. Finally, we provide theoretical motivation and design insights supporting a Generate-then-Edit paradigm, which mitigates dependency loss while retaining the efficiency of parallel decoding.

Multi-hop question answering (MHQA) requires models to retrieve and reason over evidence distributed across multiple documents. However, training dense retrievers for MHQA remains challenging because labeled query-document pairs are difficult to obtain for the evolving intermediate queries in multi-hop reasoning. Recent approaches use LLM likelihoods for step-wise supervision over queries conditioned on accumulated evidence, which anchors retrieval to known facts and biases supervision toward answer-bearing documents, leaving bridging evidence undiscovered. To address this, we propose PROBE, a framework that optimizes dense retrievers for MHQA without requiring manually annotated query-document pairs. It decouples the reasoning state into a factual summary and a prospective query that anticipates missing information. For training, we transfer cross-encoder signals only at the final hop, providing trajectory-level supervision without penalizing early exploration. Experiments on three MHQA benchmarks show that PROBE consistently improves both retrieval and QA performance over strong iterative baselines, achieving an average gain of 5.7 points while using only about 5% of the training data of the prior method.