The paper “Optimized Gaussian large language model (LLM) reprogrammed for temporal predictions”, co-authored by João P. Matos-Carvalho, LASIGE integrated member, received the Best Paper Award at the 22nd International Conference on Intelligent Environments (IE2026), held in Lisbon, Portugal, from 15 to 18 June 2026.

This work proposes a hybrid time series forecasting framework that combines Gaussian filtering for signal denoising with a Large Language Model (LLM) reprogrammed for temporal prediction, namely Time-LLM. The proposed approach uses the Gaussian filter to suppress high-frequency noise while preserving the main temporal structure of the signal, allowing the Time-LLM model to focus on meaningful temporal dependencies. The model is further optimized through a multi-agent hyperparameter optimization strategy based on structured random search.

The proposed method was evaluated on hourly turbine flow data from three Brazilian hydroelectric power plants. The obtained results show that the model consistently outperforms several state-of-the-art forecasting approaches, including TiDE, NHITS, NBEATS, GRU, DeepAR, TFT, TCN, Informer, PatchTST, FEDformer, KAN, TimesNet, and standard Time-LLM. For a forecasting horizon of 10 steps ahead, the proposed method achieved a root mean square error of 88.3, a mean absolute error of 66.9, and a mean absolute percentage error of 1.8%, demonstrating significant gains in forecasting accuracy and robustness.

The paper is available here.