Reproduce “STReasoner: Empowering LLMs for Spatio-Temporal Reasoning in Time Series via Spatial-Aware Reinforcement Learning” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Juntong Ni, Shiyu Wang, Qi He, Ming Jin, Wei Jin
STReasoner combines numerical time series with a text-described graph and questions, then uses spatial-aware reinforcement learning to reward answers that improve when graph input is present. On synthetic ST-Bench, STReasoner-8B reaches 87.12% accuracy on spatial-correlation questions, versus 77.87% for Claude-4.5-Sonnet prompted with text; the paper reports p < 0.05. On forecasting, its MAE of 65.59 is worse than image-prompted Claude’s 63.36.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.