RUC-AIBOX/STILL-3-1.5B-preview: Enhanced Slow-Thinking Reasoning

RUC-AIBOX/STILL-3-1.5B-preview is a 1.5 billion parameter model specifically engineered to improve slow-thinking reasoning capabilities, particularly in mathematical and complex problem-solving domains. Developed by RUC-AIBOX, this model leverages reinforcement learning to achieve continuous performance improvements.

Key Capabilities & Performance

• Mathematical Reasoning: Achieves a notable 39.33% accuracy on the AIME benchmark, marking a 37.18% relative improvement compared to its backbone model (28.67%).
• Benchmark Scores: Demonstrates strong performance across multiple mathematical and reasoning benchmarks:
  • MATH: 85.48% (from 84.04%)
  • AIME: 39.33% (from 28.67%)
  • OMNI: 33.00% (from 25.60%)
  • LiveAOPS: 39.50% (from 33.33%)
• Long Context: Supports a substantial context length of 131072 tokens, enabling it to process and reason over extensive problem descriptions.
• Training Methodology: Performance gains are attributed to an adaptive reinforcement learning approach, with observed improvements correlating with increased training steps.

Ideal Use Cases

• Advanced Mathematics: Excels in tasks requiring precise mathematical reasoning, such as those found in the AIME and MATH benchmarks.
• Complex Problem Solving: Suitable for scenarios where deliberate, step-by-step reasoning is more critical than rapid, intuitive responses.
• Research & Development: The model, code, and data are open-sourced to facilitate further research into enhancing slow-thinking abilities in smaller language models. More details can be found in the Slow Thinking with LLMs GitHub repository.