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

2026-06-18 · Source: cs.CL updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, extended

Summary

DreamReasoner-8B is an open-source 8B-parameter block diffusion reasoning model designed to accelerate decoding for long chain-of-thought (CoT) reasoning. A systematic study revealed that training with large block sizes degrades reasoning performance, while small block sizes preserve it. To overcome this, the authors propose block-size curriculum learning, which gradually transitions training from fine-grained to coarse-grained block sizes. This approach enables DreamReasoner-8B to achieve competitive results on mathematical and code reasoning benchmarks, matching leading autoregressive models like Qwen3-8B. Additionally, the work introduces RelaxedConfidence, an analytical decoding probe that yields average TPF (tokens per forward pass) gains of 22.5% and 54.5% in thinking and answering phases, respectively, by relaxing token commitment thresholds based on local context.

Key takeaway

For AI Scientists and Machine Learning Engineers developing efficient reasoning models, you should consider adopting block-size curriculum learning for diffusion language models. This approach stabilizes performance across varying inference block sizes, enabling competitive reasoning with autoregressive models like Qwen3-8B while offering parallel decoding benefits. Furthermore, explore implementing the RelaxedConfidence decoding strategy to achieve significant throughput gains, particularly in the answering phase, without compromising reasoning fidelity.

Key insights

Block-size curriculum learning enables diffusion models to achieve efficient, robust long-CoT reasoning competitive with autoregressive LLMs.

Principles

• Large training block sizes degrade reasoning in block diffusion models.
• Small training block sizes preserve reasoning capabilities.
• Gradual block-size increase during training improves robustness.

Method

Block-size curriculum learning involves initial training with small block sizes (e.g., 4), then gradually transitioning to larger or mixed block sizes (e.g., 32 or {4,8,16,32}) to acquire robust local causal dependencies and generalize to larger inference blocks.

In practice

• Implement block-size curriculum for diffusion model training.
• Apply RelaxedConfidence decoding for throughput gains.
• Evaluate diffusion models across diverse inference block sizes.

Topics

• Diffusion Language Models
• Block Diffusion
• Chain-of-Thought Reasoning
• Curriculum Learning
• Model Efficiency
• Mathematical Reasoning
• Code Generation

Code references

• DreamLM/DreamReasoner
• sgl-project/sglang

Best for: AI Engineer, Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• Accelerating Speculative Diffusions via Block Verification — A novel scheme enables block verification for diffusion models, improving speculative decoding speed by up to 6.3% without training.
• DreamReasoner-8B: Block-Size Curriculum Learning for Diffusion Reasoning Models — Block-size curriculum learning enables efficient diffusion models to achieve strong long chain-of-thought reasoning by bridging granularity gaps.
• A Survey on Diffusion Language Models — Diffusion Language Models (DLMs) achieve parallel text generation and bidirectional context through iterative denoising, rivaling autoregressive models in performance.
• DepCap: Adaptive Block-Wise Parallel Decoding for Efficient Diffusion LM Inference — DepCap optimizes DLM inference by adaptively determining block boundaries and enabling conflict-free parallel decoding.
• Continuous Diffusion Scales Competitively with Discrete Diffusion for Language — Continuous diffusion language models, when likelihood-trained, can scale competitively with discrete diffusion models.
• The Race to Production-Grade Diffusion LLMs with Stefano Ermon - #764 — Diffusion language models offer significantly faster and more efficient inference for text and code generation than autoregressive LLMs.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CL updates on arXiv.org.