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…
AgentInstruct Methodology
Demo explaining the AgentInstruct methodology outlined in the paper “AgentInstruct: Toward Generative Teaching with Agentic Flows”
LongRoPE
LongRoPE is a novel method that extends the context window of pre-trained LLMs to an impressive 2048k tokens by non-uniformly rescaling RoPE positional embeddings. LongRoPE has been integrated into Microsoft Phi-3.
Microsoft at ICML 2024: Innovations in machine learning
The competitive dynamics of AI agents and a method for learning and applying temporal action abstractions represent just some of Microsoft’s contributions to ICML 2024.