microsoft/LMOps

Covers prompt optimization via reinforcement learning (Promptist), efficient handling of long-context sequences through structured prompting, and inference acceleration by reusing reference text spans. Provides fundamental research on in-context learning mechanics, demonstrating how transformers perform implicit meta-optimization similar to gradient-based finetuning. Integrates with retrieval-augmented generation pipelines and broader foundation model ecosystems like UniLM and TorchScale.

4,292 stars. Actively maintained with 15 commits in the last 30 days.