Two NIRC Papers Accepted at NeurIPS 2024: FM-Delta Compression and GLAFF Forecasting

NeurIPS (Conference on Neural Information Processing Systems) is a top‑tier AI conference recommended by the China Computer Federation (CCF) with an H5 index of 337 and a 2024 acceptance rate of 25.8% (15,671 submissions). The Beijing University of Posts and Telecommunications' Network Intelligent Research Center (NIRC) secured two papers in the Main Track.

Paper 1 – FM-Delta: Lossless Compression for Storing Massive Fine‑tuned Foundation Models – The rapid growth of fine‑tuned large language models (LLMs) on platforms such as HuggingFace has increased the number of stored models from 30,000 to nearly 600,000 in two years, with over 90% being fine‑tuned variants. Although each fine‑tuned model contains billions of parameters, most of its parameters remain highly similar to its base pretrained model. The authors theoretically show that the parameter difference grows slowly as O(T^{1/4}) with the number of fine‑tuning steps T. FM-Delta exploits this by converting floating‑point parameters of both models to the same‑bit integers, subtracting them to obtain a sparse integer delta filled with many zero bits, and then entropy‑encoding the delta. Consequently, a cloud provider needs to store only one full pretrained model plus compressed deltas for its fine‑tuned variants. Experiments demonstrate roughly a 50% reduction in cloud storage consumption and a cost saving of at least 40%.

Paper 2 – GLAFF: Global‑Local Fusion for Robust Multivariate Time‑Series Forecasting – Time‑series forecasting is critical in finance, transportation, energy, climate, and healthcare, yet most existing methods rely solely on local observations and ignore the rich global information encoded in timestamps. The authors illustrate the problem with an hourly traffic‑flow series where local‑only models severely underestimate flow during holiday peaks. To address this, they propose GLAFF, a plug‑in framework consisting of three modules: (1) an Attention‑based Mapper that transforms timestamps containing global context into a standard‑normal distribution; (2) a Robust Denormalizer that inversely normalizes the mapped data to mitigate drift; and (3) an Adaptive Combiner that dynamically balances the global‑mapped prediction with the local forecast. GLAFF is model‑agnostic and can be integrated into any forecasting backbone. Empirical results on nine real‑world benchmark datasets spanning five domains show that GLAFF substantially enhances the robustness and accuracy of mainstream forecasting models.