Kimi k1.5 scales reinforcement learning with LLMs to match o1 and beat short-CoT models

The Kimi k1.5 multi-modal LLM uses a simplified reinforcement learning framework that relies on long context scaling and improved policy optimization, avoiding complex techniques like Monte Carlo tree search or value functions.

• The system achieves state-of-the-art reasoning performance across multiple benchmarks, including 77.5 on AIME, 96.2 on MATH 500, and a 94th percentile on Codeforces.
• Its long2short methods leverage long-CoT techniques to enhance short-CoT models, yielding results of 60.8 on AIME, 94.6 on MATH500, and 47.3 on LiveCodeBench.
• These short-CoT improvements outperform existing models such as GPT-4o and Claude Sonnet 3.5 by a large margin, up to +550%.

This approach demonstrates that scaling reinforcement learning unlocks new axes for AI improvement without relying on complex infrastructure, matching OpenAI's o1 while significantly surpassing other short-context models.