AIME-24/25

The 'AIME24/25' dataset/benchmark is used to evaluate the performance of reinforcement learning models, particularly in the context of agentic reinforcement learning and their ability to handle noisy trajectories during problem-solving tasks.

Papers using AIME-24/25 (3)

TraPO: A Semi-Supervised Reinforcement Learning Framework for Boosting LLM Reasoning2025CLEANER: Self-Purified Trajectories Boost Agentic Reinforcement Learning2026Sample More to Think Less: Group Filtered Policy Optimization for Concise Reasoning2025