TLDR: ScaleDiff is a new pipeline that efficiently creates a large dataset of difficult mathematical problems for training Large Reasoning Models (LRMs). It identifies hard problems using an adaptive thinking model, generates new ones with a specialized generator, and filters high-quality solutions. Training LRMs on this ScaleDiff-Math dataset significantly improves their performance on complex math benchmarks, demonstrating that more difficult training data leads to better reasoning abilities, even with cost-efficient teacher models.
Large Reasoning Models (LRMs) are becoming incredibly good at solving complex problems, especially in mathematics. Their ability to tackle intricate challenges often comes from being trained on very difficult math problems that push their reasoning capabilities. However, creating these challenging problems has been a significant hurdle. Existing methods are often expensive, require complex instructions, and don’t always produce problems that are difficult enough to truly challenge advanced AI models.
To address these limitations, researchers have introduced ScaleDiff, a new and efficient approach designed to generate a large quantity of difficult mathematical problems. This innovative pipeline aims to enhance the reasoning abilities of LRMs without incurring high computational or API costs. You can find the full research paper here: ScaleDiff: Scaling Difficult Problems for Advanced Mathematical Reasoning.
How ScaleDiff Works
ScaleDiff operates through a simple yet powerful three-step process:
1. Identifying Difficult Problems: The first step involves sifting through existing datasets to find problems that are genuinely hard. ScaleDiff uses an “adaptive thinking model” called AdaptThink for this. This model is smart enough to perceive a problem’s difficulty and decide whether it needs to “think” deeply or can provide a “no-thinking” direct answer. This allows ScaleDiff to efficiently identify difficult problems with just a single pass, which is much faster than other methods that require multiple attempts or complex evaluations.
2. Generating New Difficult Problems: Once a collection of difficult problems is identified, ScaleDiff trains a specialized generator, named DiffGen-8B, on this curated data. This generator then creates a vast number of new, challenging problems. This approach eliminates the need for complex, problem-specific instructions and the high costs associated with traditional problem generation methods.
3. Distilling and Filtering High-Quality Solutions: For every new problem generated, ScaleDiff uses a powerful but cost-efficient teacher model, Qwen3-8B, to create detailed solutions. These solutions then undergo a two-stage filtering process. First, “rule filtration” removes solutions that are repetitive or poorly formatted. Second, “model filtration” discards problems that the base AI model can already solve easily, ensuring that only truly challenging and valuable problem-solution pairs remain in the final dataset.
Also Read:
- Boosting AI Reasoning: A New Approach to Learning Diverse Thought Patterns
- GAUSS: A New Benchmark for Dissecting AI’s Math Abilities
Impressive Results and Impact
The impact of ScaleDiff is significant. When a model called Qwen2.5-Math-7B-Instruct was fine-tuned on the ScaleDiff-Math dataset, it showed a remarkable performance increase of 11.3% compared to training on the original dataset. The ScaleDiff model achieved an average accuracy of 65.9% across a range of challenging mathematical benchmarks, including AIME’24, AIME’25, HMMT-Feb’25, BRUMO’25, and MATH500. This performance surpasses many other strong Large Reasoning Models, such as OpenThinker3.
Notably, these impressive results were achieved using the more cost-efficient Qwen3-8B model as a teacher, demonstrating that ScaleDiff can effectively transfer advanced reasoning capabilities without relying on larger, more expensive teacher models. The research also observed a clear “scaling phenomenon”: as the quantity of difficult problems in the training data increased, the model’s performance on challenging benchmarks consistently improved.
To further support research and ensure reproducibility, the ScaleDiff team has open-sourced the ScaleDiff-Math dataset, the fine-tuned ScaleDiff model, and the implementation code. This initiative provides valuable resources for the AI community to continue advancing the field of mathematical reasoning.
