RANDOM REWARDS DURING RL BOOST MATH REASONING IN SOME LLMS

Summary from the TLDR AI newsletter

Qwen2.5-Math models achieve 15-24% performance gains using completely spurious rewards like random feedback, incorrect answers, or formatting requirements—nearly matching traditional RL methods. However, this phenomenon was only observed on Qwen models due to their pre-existing code reasoning capabilities, while other model families like Llama3 and OLMo2 show minimal or negative effects.