DreamReasoner-8B: Block-Size Curriculum Learning for Diffusion Reasoning Models

Abstract: Block diffusion LLMs accelerate decoding through parallel block-wise denoising, yet whether they can be reliably scaled for long chain-of-thought (CoT) reasoning remains unresolved. To this end, we develop DreamReasoner-8B, an open-source block diffusion reasoning model, and conduct a systematic study of how training and inference block sizes affect long-CoT reasoning. Our analysis reveals a stark performance disparity: training with large block sizes yields remarkably poor reasoning, whereas small block sizes preserve effective reasoning. To bridge this granularity gap, we propose block-size curriculum learning, which gradually transitions training from fine-grained to coarse-grained block sizes, thereby overcoming this limitation and enabling strong reasoning performance that generalizes across diverse inference block sizes. On mathematical and code reasoning benchmarks, DreamReasoner-8B achieves results competitive with leading open autoregressive models such as Qwen3-8B. This work establishes a practical foundation for efficient, reasoning-capable diffusion LLMs. We release our model at https://github.com/DreamLM/DreamReasoner.

• The paper introduces block-size curriculum learning to train an 8B diffusion model that overcomes reasoning degradation present in coarse decoding.
• It demonstrates superior performance on mathematical and code benchmarks compared to autoregressive models, ensuring robust accuracy across block sizes.
• The implementation of RelaxedConfidence decoding boosts token throughput by allowing early commitment in high-confidence, structured output phases.

Block-Size Curriculum Learning Enables Efficient Diffusion Reasoning: Insights from DreamReasoner-8B

Introduction

This work proposes DreamReasoner-8B, an 8-billion parameter block-diffusion LLM (DLM) focused on long-chain-of-thought (CoT) reasoning. Diffusion language modeling has gained interest for its potential to parallelize generation, overcoming the left-to-right sequential bottleneck of autoregressive (AR) models while retaining contextual fidelity. However, scaling block-size to enhance efficiency in block diffusion models has historically compromised reasoning ability, a phenomenon this paper rigorously investigates and addresses.

Methodology and Block-Size Curriculum Learning

DreamReasoner-8B is initialized from Qwen3-8B-Base, then continually pretrained with a block-diffusion loss using a large, reasoning-intense dataset. The key methodological innovation is block-size curriculum learning: rather than fixing the block granularity during training, training is initiated using fine-grained (block size 4) blocks and is then systematically transitioned to coarser blocks (block size up to 32). This staged exposure enables robust acquisition of both local and global dependencies, preserving reasoning precision while enhancing decoding parallelism.

The paper's empirical analysis reveals that:

• Training with only large blocks leads to severe performance collapse on sophisticated reasoning tasks, attributed to a loss of token-level sequential structure.
• Conversely, training with small blocks retains robust hierarchical reasoning even under coarser decoding.
• Curriculum learning unifies these regimes, yielding models that generalize across inference block sizes without necessitating model retraining.

Evaluation and Numerical Results

DreamReasoner-8B demonstrates competitive or superior performance compared to leading 8B-scale AR models such as Qwen3-8B-Thinking across mathematical (AIME, MATH, GSM8K) and code generation (HumanEval, LiveCodeBench) benchmarks. For MATH and GPQA, DreamReasoner-8B outperforms Qwen3-8B-Base, and in code, it surpasses prior diffusion-based approaches by a wide margin.

Crucially, DreamReasoner-8B maintains high accuracy as block size increases—contrary to competing block-diffusion and hybrid models (e.g., SDAR-30B-A3B-Sci), which show sharp degradation when scaling block size beyond their training configuration. This indicates robust granularity transfer enabled by the curriculum approach.

Throughput is measured as tokens per forward pass (TPF):

• DreamReasoner-8B achieves monotonic TPF improvements with increased block size, with only minor accuracy loss.
• Adoption of RelaxedConfidence decoding (see below) yields further throughput gains of up to 54.5% in the answer generation phase without degrading reasoning fidelity.

Decoding Efficiency and RelaxedConfidence

Despite advances in block-wise parallelism, standard diffusion decoding strategies are limited by conservative per-token confidence thresholds, curtailing efficiency. The paper introduces RelaxedConfidence, a context-aware probe that permits earlier token commitment when surrounded by other high-confidence neighbors. This method ensures:

• Acceleration of decoding during structured output phases (e.g., answer finalization, code block emission).
• Minimal accuracy degradation since only locally supported tokens are committed early.

This approach broadens the practical throughput-quality Pareto frontier for block-diffusion LMs.

Theoretical and Practical Implications

DreamReasoner-8B advances understanding of the scaling behavior of block-diffusion architectures:

• Training and inference block size are exposed as crucial interacting axes; only through controlled curriculum exposure can models realize robust long-CoT generalization and efficiency.
• Block diffusion—with curriculum learning—can serve as a viable alternative to AR decoding for reasoning-intensive tasks, facilitating flexible quality-latency trade-offs in deployment.
• The RelaxedConfidence strategy signals that architectural and algorithmic innovations on commitment scheduling constitute a fertile direction for accelerating diffusion LMs without quality regression.

Limitations and Future Work

The current design employs fixed block sizes and is primarily validated on mathematical reasoning and code. Further research is needed to:

• Investigate semantic- or syntax-aware dynamic block partitioning, potentially aligning block boundaries to code, paragraph, or subproblem structure.
• Generalize block-size curriculum to open-ended tool use, interactive agents, or multimodal CoT reasoning.

Expanding the scope of curriculum strategies and integrating adaptive block partitioning holds promise to further advance the efficiency-quality curve.

Conclusion

DreamReasoner-8B solidifies block diffusion as a compelling, efficient generative modeling paradigm for reasoning tasks when equipped with block-size curriculum learning. The model's open release and superior throughput-accuracy balance provide a strong foundation for future scaling, integration with adaptive decoding strategies, and application to broader AI reasoning domains.

Reference:

"DreamReasoner-8B: Block-Size Curriculum Learning for Diffusion Reasoning Models" (2606.19257)