POLARIS-9B: a <$500, 4-GPU recipe for long-story writing that rivals 3x-larger models
MohitIyyer · x · 2026-10-24
A University of Maryland / Google DeepMind team presents POLARIS at EMNLP2026: a low-compute GRPO recipe with a structured 16-dimension StoryQuality rubric and Human Reference Injection. Trained on just 1.4K stories with 4 GPUs in 48h (<$500), POLARIS-9B maintains quality when asked to write 12,000-word stories after training only on 4k-word ones, and competes with open models 3x its size — ranking #3/14 on EQ-Bench Creative behind only closed frontier models. Human blind evals prefer it over Qwen3.5-27B 51.2% of the time.
More from Research
- OmniRay: single-image 3D reconstruction for any lens — fisheye to panorama — sets new SOTA — anand_bhattad · 2026-10-24
- Scott Aaronson on OpenAI's apparent AI proof of his wife's conjecture — TOEwithCurt · 2026-10-24
- AI appears to solve complexity theorist's years-long conjecture; Scott Aaronson calls it a landmark for math — TOEwithCurt · 2026-10-24
- H-CDLMs: continuous diffusion LMs jointly denoise tokens and semantic clusters for better, cheaper generation — abursuc · 2026-10-24
- Xiamen University's FlowMAS uses GFlowNet to auto-design multi-agent topologies — jiqizhixin · 2026-10-24
- Shriram Krishnamurthi: maximize verified hw+sw stacks while the AI money cannon is on — iskander · 2026-10-24