AI Agent that at inference time updates it's harness and model weights (github.com)

A groundbreaking framework, SIA (Self-Improving AI with Harness & Weight Updates), has been officially implemented, showcasing significant advancements in AI model performance. Developed by Hebbar et al. (2026), SIA enables a language-model agent to autonomously enhance its task-specific capabilities by continuously updating both the harness and model weights. This innovative approach resulted in a 56.6% improvement on the LawBench benchmark, a 91.9% reduction in runtime for GPU kernels, and an impressive 502% gain in single-cell RNA denoising performance compared to baseline models. The SIA framework consists of three collaborative AI agents: a Meta-Agent that generates a tailored Target Agent, a Target Agent that performs the task, and a Feedback Agent that analyzes performance and suggests enhancements. This self-improving loop allows AI systems to refine their capabilities iteratively. With built-in tasks including gpqa, lawbench, longcot-chess, and spaceship-titanic, SIA is optimized for efficiency, with features such as the Triangle Multiplicative Update implemented as a Triton kernel, achieving a remarkable 14x speedup over traditional setups. Researchers and practitioners can easily leverage SIA for their AI projects, making it a significant tool in the AI/ML landscape.