Planner and Executor: Collaboration between Discrete Diffusion And Autoregressive Models in Reasoning

TL;DR

Current autoregressive models achieve robust reasoning but are costly due to long token chains, while discrete diffusion language models enable fixed-step, parallel generation advantageous for planning. This work investigates hybrid planner–executor architectures that pair a DDLM planner with an ARM executor, comparing text-space prompting to a latent-space projector that maps diffusion latents into the ARM embedding space. Latent-space communication yields large accuracy gains on challenging benchmarks (e.g., DART-5 from 27.0% to 54.0%; AIME24 from 0.0% to 14.0%), and a 64-token planner with a few executor tokens can surpass larger autoregressive models at a fraction of the token cost. The results suggest that diffusion planners plus autoregressive executors can substantially reduce compute and token budgets while maintaining or improving reasoning reliability, offering a promising direction for budget-aware hybrid reasoning systems.

Abstract

Current autoregressive language models (ARMs) achieve high accuracy but require long token sequences, making them costly. Discrete diffusion language models (DDLMs) enable parallel and flexible generation within a fixed number of steps and have recently emerged for their strong performance in complex reasoning and long-term planning tasks. We present a study exploring hybrid architectures that couple DDLMs with ARMs to assess whether their collaboration can yield complementary benefits. We first examine collaboration in text space, where one model plans the reasoning process and another executes the final answer based on that plan. We then extend this setup to latent-space communication, introducing a learned projector that maps DDLM latents into the ARM's embedding space, potentially bypassing some of the text-generation limitations of diffusion models. We find that shifting DDLM --> ARM communication from text space to latent space yields significant accuracy gains, for example increasing from 27.0% to 54.0% on DART-5 and from 0.0% to 14.0% on AIME24. We also find that combining a DDLM planner with an ARM executor can provide substantial computational savings with little to no impact on accuracy. For example, the latent-space pipeline, using 64 tokens for planning and roughly 5 for execution, surpasses Qwen3.1-7B on DART-5 and AIME, despite Qwen using 44 times more tokens. Overall, our study offers new insights into reasoning with DDLMs and highlights their potential in hybrid architectures.

Figures (5)

• Figure 1: Accuracy–cost trade-offs across planner–executor configurations. DDLM$\to$ARM, particularly with latent-space exchange, achieves higher reasoning accuracy at lower token budgets compared to ARM-only, ARM$\to$DDLM, and reasoning models.
• Figure 2: Overview of the latent-space collaboration pipeline. A discrete diffusion language model (DDLM) generates a structured plan from noisy latents. The plan is projected directly into the autoregressive model (ARM) embedding space through a learned projection layer (latent space). The ARM then conditions on the plan and the question to produce the final answer.
• Figure 3: Diagnostic configurations for attributing errors to planner or executor. Setup X tests whether failures stem from the planner: if replacing the diffusion planner (DDLM) with an autoregressive planner (ARM) fixes the output, the error is attributed to the DDLM. Setup Y tests executor reliability: if a diffusion executor succeeds where an ARM executor fails, the limitation lies in the executor.
• Figure 4: Planner vs. executor failures in text- vs. latent-space collaboration. Results for LLaDA-8B-Instruct and Llama-3.1-3B-Instruct. Latent-space collaboration substantially reduces planning errors compared to text-space.
• Figure 5: Benchmark comparison of text-space vs. latent-space collaboration. Accuracy of isolated models (LLaMA-3.2-3B ARM, LLaDA-8B DDLM) and collaborative configurations. In the latent setting, DDLM (64-token planner) combined with the ARM executor consistently outperforms text-space collaboration on DART and AIME, while maintaining comparable performance on ARC and MMLU.