Improving Diversity in LLM Short Story Generation
DivLM is a post-training framework meant to make LLM-generated short stories vary more by genre, tone, style, and named entities.
The authors use continued pre-training on a creative writing corpus, then restore instruction-following with weight residuals. They add reinforcement learning with a composite reward that targets diversity while keeping quality intact. In tests on two LLM families, DivLM raised diversity metrics by more than 9% on average versus alternative approaches. The paper says instruction following, response quality, and similarity to human outputs were preserved. ArXiv · AI/CL/LG's note
The authors use continued pre-training on a creative writing corpus, then restore instruction-following with weight residuals. They add reinforcement learning with a composite reward that targets diversity while keeping quality intact. In tests on two LLM families, DivLM raised diversity metrics by more than 9% on average versus alternative approaches. The paper says instruction following, response quality, and similarity to human outputs were preserved. ArXiv · AI/CL/LG's note
score 4