DLLM Agent: See Farther, Run Faster

TL;DR

This paper investigates whether changing the generation paradigm from autoregressive (AR) to diffusion-based (DLLM) alters agentic multi-step planning and tool use when the agent framework and supervision are fixed. By instantiating DLLM and AR backbones within the same DeepDiver workflow and fine-tuning on identical trajectory data, the study shows DLLM Agents achieve comparable or higher end-task accuracy while delivering substantial end-to-end speedups (average >$30\%$, with some cases up to $8\times$). DLLM Agents also exhibit fewer interaction rounds and tool invocations given correct task completion, indicating higher planner hit rates and earlier convergence to viable action paths. The work identifies practical deployment considerations for DLLMs in tool-using agents, including stronger tool-call supervision to reduce structured-tool-call failures and alignment of diffusion masking with multi-turn contexts. A mechanistic analysis of attention dynamics reveals paradigm-specific coordination patterns, suggesting that diffusion backbones enable stronger global planning signals across workflow stages. Overall, the findings position diffusion backbones as a promising avenue for building more efficient, forward-looking language-model agents without sacrificing accuracy, while outlining directions for robustness and generalization across tasks.

Abstract

Diffusion large language models (DLLMs) have emerged as an alternative to autoregressive (AR) decoding with appealing efficiency and modeling properties, yet their implications for agentic multi-step decision making remain underexplored. We ask a concrete question: when the generation paradigm is changed but the agent framework and supervision are held fixed, do diffusion backbones induce systematically different planning and tool-use behaviors, and do these differences translate into end-to-end efficiency gains? We study this in a controlled setting by instantiating DLLM and AR backbones within the same agent workflow (DeepDiver) and performing matched agent-oriented fine-tuning on the same trajectory data, yielding diffusion-backed DLLM Agents and directly comparable AR agents. Across benchmarks and case studies, we find that, at comparable accuracy, DLLM Agents are on average over 30% faster end to end than AR agents, with some cases exceeding 8x speedup. Conditioned on correct task completion, DLLM Agents also require fewer interaction rounds and tool invocations, consistent with higher planner hit rates that converge earlier to a correct action path with less backtracking. We further identify two practical considerations for deploying diffusion backbones in tool-using agents. First, naive DLLM policies are more prone to structured tool-call failures, necessitating stronger tool-call-specific training to emit valid schemas and arguments. Second, for multi-turn inputs interleaving context and action spans, diffusion-style span corruption requires aligned attention masking to avoid spurious context-action information flow; without such alignment, performance degrades. Finally, we analyze attention dynamics across workflow stages and observe paradigm-specific coordination patterns, suggesting stronger global planning signals in diffusion-backed agents.

Figures (6)

• Figure 1: Same agent workflow, different generation paradigms. Under identical tool-use workflows and training data, DLLM Agents (diffusion) tend to exhibit earlier trajectory-level convergence and less redundant interaction, while AR agents more often follow incremental, locally conditioned decisions that can lead to detours and repeated tool calls.
• Figure 2: (a) and (b) show the distributions of the number of times the information seeker is invoked by the DLLM agent and the AR agent, respectively. The DLLM agent invokes the seeker significantly fewer times than the AR Agent.
• Figure 3: Heatmap showing the relative decoding order of each token within each block of an example Planner Agent response. Different blocks are separated by red lines.
• Figure 4: Analysis of the Planner Agent response shown in Figure \ref{['fig:planner_heatmap']}. (a) Probability of tokens with different confidence scores being decoded during diffusion process. (b) Average entropy of remaining mask tokens across different diffusion steps. (c) Number of remaining mask tokens within each block across different diffusion steps.
• Figure 5: Heatmap showing the relative decoding order of each token within each block of an example Information Seeker Agent response. Different blocks are separated by red lines.