科學機器學習
UNICON Unifies Nine Classes of Numerical Systems With 41.6 Million Parameters, Enabling Cross-Domain Forecasting With Frozen Weights
UNICON converts data from rivers, traffic networks, power grids, weather systems, and human mobility into a shared graph-sequence interface, using a small number of numerical examples to specify new tasks in context. In unseen domains such as web traffic, the model approaches the performance of specialized models, but it has so far been validated only on forecasting tasks, and neither its weights nor code have been released.
Language models can learn tasks on the fly from textual examples. UNICON attempts to bring the same concept to numerical systems that share no common vocabulary. The [paper](https://arxiv.org/abs/2607.28432) represents river catchments, roads, power grids, weather grids, and webpages as graphs: nodes carry a segment of multivariate history, while edges represent physical relationships. The key-value pairs in the context correspond to historical states and future states after a specified forecast horizon, respectively. The model processes these examples sequentially, then generates a future state for a query containing only a key—all without dataset-specific input layers, output layers, or gradient updates.
UNICON has just 41.6 million trainable parameters and was trained on approximately 216 GB of data from 20 sources spanning nine classes of systems, including hydrology, traffic, energy, land, ocean, solar energy, and human mobility. Training took 28 hours on two NVIDIA H200 GPUs. The researchers then evaluated it on nine systems, including three domains entirely absent from the training set: air quality, WikiMaths webpage activity, and space weather. After seeing five similar examples, the frozen model substantially outperformed kNN under the same example-driven setup overall; on WikiMaths, it approached a specialized model trained specifically on that dataset. When a language-model agent was added to select and reorganize multiple numerical prompts and aggregate repeated calls, UNICON surpassed the expert baselines used in the paper on both WikiMaths and CAMELS-CL.
Controlled experiments showed that removing the examples, shuffling the historical-future pairings, or corrupting the numerical values all increased error. A single correct example was already sufficient to recover most of the gains obtained from five examples. This makes UNICON look more like a numerical in-context learner than merely a large forecasting model. However, the evaluation remains limited to time-series forecasting and does not yet cover imputation, anomaly detection, simulation, or control. The gains from the agent-based approach also come with the cost of multiple inference calls. The original data is distributed across its respective providers, and some evaluation data—such as the [public dataset used by AirQualityBench](https://huggingface.co/datasets/xuxing123/aq_dataset)—is available. However, the team has not released the model, training code, or complete agent workflow, so the cross-domain results still await independent reproduction.