Publication
Algorithms
Microsoft Research Blog
Tracing the path to self-adapting AI agents
Introducing Trace, Microsoft and Stanford University’s novel AI optimization framework, now available as a Python library. Trace adapts dynamically and optimizes a wide range of applications from language models to robot control.
Dataset Source Code
Trace
Trace is a new AutoDiff-like tool for training AI systems end-to-end with general feedback (like numerical rewards or losses, natural language text, compiler errors, etc.). Trace generalizes the back-propagation algorithm by capturing and propagating an…