Reproduce “DreamReasoner-8B: Block-Size Curriculum Learning for Diffusion Reasoning Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Zirui Wu, Lin Zheng, Jiacheng Ye, Shansan Gong, Xueliang Zhao, Yansong Feng, Wei Bi, Lingpeng Kong
Researchers at The University of Hong Kong and Peking University developed DreamReasoner-8B, an 8-billion-parameter block-diffusion model that achieves reasoning performance competitive with leading autoregressive models by utilizing a block-size curriculum learning strategy. This approach enables robust performance across various block sizes and substantially improves decoding efficiency while mitigating performance degradation observed with larger block sizes in prior diffusion models.
Key insights
- Naive training with large block sizes leads to substantial reasoning performance degradation, particularly during fine-grained inference, suggesting overfitting to intra-block long-range dependencies.
- Block-size curriculum learning effectively maintains fine-grained precision while enabling robust performance for large-block generation, demonstrating better generalization across varying inference granularities.
- Decoding efficiency (Tokens Per Forward pass) improves monotonically with larger inference block sizes, and context-aware confidence thresholds (RelaxedConfidence) can further accelerate generation, especially in lexically constrained 'answering phases'.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.