Latest News

A new research introduces SPARC, a framework designed to address this challenge by decoupling visual perception and reasoning into two distinct computational stages. By separating these processes, the approach enables more efficient use of computational resources while preserving, and in many cases improving, model performance. 

More details on the methodology and experimental evaluation are provided in the paper “SPARC: Separating Perception And Reasoning Circuits for Test-time Scaling of Vision-Language Models” (arXiv:2602.06566).

SustainML continues to evolve to support increasingly complex and distributed AI optimization scenarios. As part of this evolution, we have integrated Remote Procedure Call (RPC) functionality using Fast DDS Pro to enhance control-plane interactions inside the framework.

SustainML has successfully integrated FPGA-based hardware estimation into its AI sustainability evaluation pipeline, enabling more accurate energy and carbon footprint analysis for CNN models deployed on FPGA platforms. By combining hardware-level latency and power prediction with automated carbon intensity calculation, the framework now supports end-to-end sustainability assessment for both Hugging Face transformer models and ONNX-based CNN models targeting FPGA devices. This integration strengthens SustainML’s ability to evaluate AI models across heterogeneous hardware environments and brings hardware-aware carbon estimation closer to real-world deployment scenarios.

On February 5th, 2026, eProsima held its internal SustainML 2026 Kick-Off Meeting, bringing together the team involved in the project to review the progress achieved during 2025 and to align on objectives and priorities for the year ahead.

The meeting marked an important milestone to close the 2025 work cycle, acknowledge the results achieved by the team, and define a clear and shared direction for 2026, reinforcing eProsima’s commitment to the SustainML project.