Ai2's new Olmo 3.1 extends reinforcement learning training for stronger reasoning benchmarks

The Allen Institute for AI (Ai2) has unveiled Olmo 3.1, an enhanced version of its Olmo 3 model family, emphasizing efficiency and transparency. Through extended reinforcement learning, these models outperform previous benchmarks, offering enterprises superior reasoning capabilities and customization potential.

Reinforcement Learning Enhancements

Ai2's Olmo 3.1 models represent a significant leap in reinforcement learning application. By extending training on Olmo 3 Think 32B with an additional 21 days using 224 GPUs, Ai2 achieved noteworthy improvements in reasoning and math capabilities. This rigorous training regimen resulted in substantial performance enhancements across various benchmarks, including a 5+ point increase on the AIME test.

Model Variants and Real-World Applications

The Olmo 3.1 family includes upgraded versions of its predecessors, specifically Olmo 3.1 Think 32B and Olmo 3.1 Instruct 32B. The Think variant focuses on advanced research, while the Instruct model, scaled to 32B from a smaller 7B version, excels in instruction-following and multi-turn dialogue. These models are optimized for enterprise use, providing robust solutions for chat and tool use scenarios.

Commitment to Transparency and Open Source

Ai2 continues to prioritize transparency and open-source development with the Olmo 3.1 models. Enterprises can customize these models by integrating additional data, enhancing their utility and relevance. Tools like OlmoTrace facilitate understanding the training process, ensuring that organizations have full visibility over model behavior and data provenance.

Key Highlights

• Olmo 3.1 models achieved over 5 points improvement on AIME benchmarks.
• Extended training involved 224 GPUs over a 21-day period.
• Olmo 3.1 Instruct 32B is optimized for chat and multi-turn dialogue.
• Models are available on Ai2 Playground and Hugging Face.
• Ai2 emphasizes transparency with tools like OlmoTrace.