Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Beyond Surface Reasoning: Unveiling the True Long Chain-of-Thought Capacity of Diffusion Large Language Models

Qiguang Chen, Hanjing Li, Libo Qin, Dengyun Peng, Jinhao Liu, Jiangyi Wang, Chengyue Wu, Xie Chen, Yantao Du, Wanxiang Che

TL;DR

This work formalizes the Parallel–Sequential Contradiction (PSC) in diffusion large language models (DLLMs), showing that parallel decoding yields high entropy and superficial parallelism, while genuine long chain-of-thought (Long CoT) reasoning devolves into autoregressive-like processing as tasks grow more complex. It introduces three inference-time scaling dimensions—parallel, diffusion, and sequential—and demonstrates that diffusion and sequential scaling are upper-bounded by PSC, whereas parallel scaling remains viable but computationally costly. The authors quantify DLLM behavior across simple versus complex reasoning, reveal an efficiency gap relative to autoregressive models, and propose mitigations (parallel-focused prompting, diffusion early stopping, and parallel scaling) to alleviate PSC and improve both accuracy and throughput. These findings have practical implications for designing PSC-aware prompting strategies and architectural decisions to enable more reliable long-horizon reasoning with DLLMs. Overall, the paper provides a PSC-centered framework to analyze and enhance DLLM reasoning performance in real-world, long-form tasks.

Abstract

Recently, Diffusion Large Language Models (DLLMs) have offered high throughput and effective sequential reasoning, making them a competitive alternative to autoregressive LLMs (ALLMs). However, parallel decoding, which enables simultaneous token updates, conflicts with the causal order often required for rigorous reasoning. We first identify this conflict as the core Parallel-Sequential Contradiction (PSC). Behavioral analyses in both simple and complex reasoning tasks show that DLLMs exhibit genuine parallelism only for directly decidable outputs. As task difficulty increases, they revert to autoregressive-like behavior, a limitation exacerbated by autoregressive prompting, which nearly doubles the number of decoding steps with remasking without improving quality. Moreover, PSC restricts DLLMs' self-reflection, reasoning depth, and exploratory breadth. To further characterize PSC, we introduce three scaling dimensions for DLLMs: parallel, diffusion, and sequential. Empirically, while parallel scaling yields consistent improvements, diffusion and sequential scaling are constrained by PSC. Based on these findings, we propose several practical mitigations, parallel-oriented prompting, diffusion early stopping, and parallel scaling, to reduce PSC-induced ineffectiveness and inefficiencies.

Beyond Surface Reasoning: Unveiling the True Long Chain-of-Thought Capacity of Diffusion Large Language Models

TL;DR

This work formalizes the Parallel–Sequential Contradiction (PSC) in diffusion large language models (DLLMs), showing that parallel decoding yields high entropy and superficial parallelism, while genuine long chain-of-thought (Long CoT) reasoning devolves into autoregressive-like processing as tasks grow more complex. It introduces three inference-time scaling dimensions—parallel, diffusion, and sequential—and demonstrates that diffusion and sequential scaling are upper-bounded by PSC, whereas parallel scaling remains viable but computationally costly. The authors quantify DLLM behavior across simple versus complex reasoning, reveal an efficiency gap relative to autoregressive models, and propose mitigations (parallel-focused prompting, diffusion early stopping, and parallel scaling) to alleviate PSC and improve both accuracy and throughput. These findings have practical implications for designing PSC-aware prompting strategies and architectural decisions to enable more reliable long-horizon reasoning with DLLMs. Overall, the paper provides a PSC-centered framework to analyze and enhance DLLM reasoning performance in real-world, long-form tasks.

Abstract

Recently, Diffusion Large Language Models (DLLMs) have offered high throughput and effective sequential reasoning, making them a competitive alternative to autoregressive LLMs (ALLMs). However, parallel decoding, which enables simultaneous token updates, conflicts with the causal order often required for rigorous reasoning. We first identify this conflict as the core Parallel-Sequential Contradiction (PSC). Behavioral analyses in both simple and complex reasoning tasks show that DLLMs exhibit genuine parallelism only for directly decidable outputs. As task difficulty increases, they revert to autoregressive-like behavior, a limitation exacerbated by autoregressive prompting, which nearly doubles the number of decoding steps with remasking without improving quality. Moreover, PSC restricts DLLMs' self-reflection, reasoning depth, and exploratory breadth. To further characterize PSC, we introduce three scaling dimensions for DLLMs: parallel, diffusion, and sequential. Empirically, while parallel scaling yields consistent improvements, diffusion and sequential scaling are constrained by PSC. Based on these findings, we propose several practical mitigations, parallel-oriented prompting, diffusion early stopping, and parallel scaling, to reduce PSC-induced ineffectiveness and inefficiencies.

Paper Structure

This paper contains 74 sections, 2 theorems, 33 equations, 11 figures, 1 table.

Table of Contents

  1. Introduction
  2. Parallel-Sequential Contradiction
  3. Parallel Masked Diffusion Language Models
  4. Sequential Long Chain-of-Thought Reasoning
  5. Parallel-Sequential Contradiction
  6. Do DLLMs truely perform parallel reasoning that avoids PSC?
  7. Parallel-Sequential Contradictions cause superficial parallel reasoning.
  8. Parallel-Sequential Contradictions cause superficial parallel reasoning in simple scenarios.
  9. For complex reasoning, DLLMs converge toward autoregressive-like behavior to avoid PSC.
  10. Diffusion-Step Dilemma: Sacrificing Efficiency Under PSC
  11. When complex reasoning, DLLMs require significantly more diffusion steps than ALLMs.
  12. In reasoning scenarios, a large number of diffusion steps for autoregressive reasoning is unavoidable for acceptable accuracy.
  13. Rethinking the prompting strategies in DLLMs from PSC perspective
  14. Sequential Reasoning Prompting will enlarge PSC's negative impact for DLLMs.
  15. Constraint-guided Reasoning Prompting enhances model performance by preventing the introduction of additional PSC.
  16. ...and 59 more sections

Key Result

Lemma 1

Fix $k>1$ and $s_1\in\Omega$. Under Definition def:serial, for any $\delta>0$ there exists $\varepsilon_0>0$ such that for all $\varepsilon<\varepsilon_0$, where $s'_1 \sim N'(s_1)\Leftrightarrow \varepsilon(s_1) \neq \varepsilon(s'_1) \wedge \varepsilon(s'_1) \in m_\varepsilon$. Under Definition def:parallel, there exist $C>0$ and $\varepsilon_0>0$ such that for all $\varepsilon<\varepsilon_0$,

Figures (11)

  • Figure 1: Overview of our work. Applying DLLMs to reasoning scenarios reveals an inherent contradiction between parallel processing and sequential reasoning, leading to high entropy, superficial parallel decoding, and limited Long CoT capabilities.
  • Figure 2: Diffusion order analysis with d1 zhao2025d1 and Trado wang2025trado, where later decoding orders are indicated by shallower colors.
  • Figure 3: Diffusion speed analysis in Long-CoT-needed tasks with LLaDA-8B-Instruct nie2025large and Dream-7B-Instruct ye2025dream on BigGSM benchmark chen2024unlocking.
  • Figure 4: Self-reflection performance and rationale quality evaluation on DLLMs.
  • Figure 5: Self-exploration performance and rationale quality evaluation on DLLMs.
  • ...and 6 more figures

Theorems & Definitions (4)

  • Definition 1: Serial tasks: cascading sensitivity
  • Definition 2: Parallel tasks: partial invariance
  • Lemma 1: Pointwise probability caps
  • Proposition 1: Skip-step parallel predictions on serial vs. parallel tasks