POLARIS-9B: Small AI model writes long stories rivaling much larger ones

Small open-weight language models typically struggle with long-form creative writing—they either cut stories short or see quality degrade sharply as length increases. A new paper from UMass Amherst and Google Research introduces POLARIS (Policy Optimization with LLM-as-a-judge rewards and Anchored-Reference Injection for Storywriting), a low-compute recipe that solves this. The team applied it to Qwen3.5-9B, using only 1.4K prompt-story pairs derived from 100 short-story anthologies and trained on just 4 A100 GPUs. The key innovations: a frontier LLM judge that scores stories on a structured quality rubric (providing online rewards), and human-reference injection (HRI)—teacher-forcing a human-written story as a high-reward anchor within each GRPO group. The result, POLARIS-9B, matches or exceeds much larger models in story quality and length adherence.

Across five benchmarks covering both in-distribution and out-of-distribution prompts, POLARIS-9B is competitive with Qwen3.5-27B (three times its size) while following length instructions far more precisely than base Qwen3.5-9B. A blinded human evaluation confirmed that POLARIS-9B is preferred over the base model and on par with the 27B variant. Remarkably, despite training only on stories up to 4K words, the model preserves quality when asked to produce stories up to 12K words—a threefold length generalization that most open-weight models fail at. The authors argue that length generalization is a meaningful stress test for creative writing models and a useful lens for distinguishing models that otherwise perform similarly. This work democratizes high-quality long-form story generation, proving that small models can punch far above their weight with the right training strategy.