Generating Constructive Feedback on Stories via Reinforcement Learning

The researchers proposed a method based on reinforcement learning (RL), using Group Relative Policy Optimization (GRPO) to guide large language models in providing constructive feedback for stories. This approach innovatively uses multi-component reward functions, prioritizing feedback that improves quality, addresses key writing issues, and is tailored to the uniqueness of the story, without relying on real-world feedback data. In automatic and manual evaluations conducted on three story corpora, this approach performed better than state-of-the-art large language models, including Gemini, and competitive baselines. The study confirmed that providing actionable advice is the main factor driving constructive feedback.