Design of Verifiable Reward Functions for LLM Reasoning

Develop robust, generalizable verifiable reward functions that reliably supervise large language model reasoning across tasks and domains, enabling effective reinforcement learning from automatically checkable outcomes.

Background

The paper situates AsyncThink within reinforcement learning with verifiable rewards (RLVR), which aims to improve LLM reasoning through outcome-supervised signals that can be automatically checked. While prior models show strong results, the authors note that creating verifiable reward functions that generalize and robustly evaluate reasoning remains unresolved.

In their experiments, the authors employ rule-based rewards (accuracy, format compliance, and thinking concurrency) but acknowledge that the broader design of verifiable reward functions is still an open problem in the field.