LLMODE: Aligning ODEs with LLMs via Gated Token Injection for Irregular Spatio-Temporal Forecasting
LLMODE turns irregular graph observations into fixed memory tokens that a frozen LLM can use for forecasting.
The paper pairs a graph-aware ODE encoder with a Perceiver-style resampler to compress variable-length spatio-temporal trajectories. It also encodes compact statistical descriptors as separate context memory, then injects both sources through gated cross-attention. The authors report competitive results across urban and physical-dynamics benchmarks, with stronger gains when sampling is sparse or dynamically complex. They also claim zero-shot generalization on unseen urban regions without adaptation. ArXiv · AI/CL/LG's note
The paper pairs a graph-aware ODE encoder with a Perceiver-style resampler to compress variable-length spatio-temporal trajectories. It also encodes compact statistical descriptors as separate context memory, then injects both sources through gated cross-attention. The authors report competitive results across urban and physical-dynamics benchmarks, with stronger gains when sampling is sparse or dynamically complex. They also claim zero-shot generalization on unseen urban regions without adaptation. ArXiv · AI/CL/LG's note
score 4