Microsoft Research has released AutoAdapt, an open-source framework for adapting large language models to specialized domains. According to the research post, the problem it addresses is that adapting LLMs to high-stakes domains is “slow, expensive, and hard to reproduce.” The post states that AutoAdapt is designed to make “domain adaptation an engineering discipline, not an ad hoc process.”
The domains the post identifies as motivating examples are law, medicine, and cloud incident response — settings where the cost of a poorly-behaved model or an unreproducible tuning process is high.
Three components
AutoAdapt’s architecture has three parts, as described in the post.
The first is the Adaptation Configuration Graph (ACG), described as “a structured representation of the system’s configuration space that enables efficient search while guaranteeing valid pipelines.” Teams adapting LLMs must currently choose among approaches such as retrieval-augmented generation, supervised fine-tuning, and parameter-efficient methods including LoRA, with many interacting hyperparameters. The ACG encodes this space explicitly and ensures any generated pipeline is valid.
The second component is a planning agent that uses the ACG to propose and evaluate adaptation strategies. The post describes it as using “agentic planning to propose strategies, evaluate them against user requirements, and iterate until the plan is feasible and well-grounded, rooting each decision in best practices and explicit constraints.”
The third component is AutoRefine, described as “a budget-aware refinement loop that optimizes hyperparameters by strategically selecting which experiments to run next, even under limited feedback.”
Evaluation
AutoAdapt was evaluated on reasoning, question answering, coding, classification, and cloud-incident diagnosis tasks. The post reports it outperforms baseline methods across success rate, normalized performance score, and cumulative metrics. The overhead of using AutoAdapt is approximately 30 minutes of additional time and $4 in additional cost, according to the post.
Open-source release
Microsoft is releasing the framework as open source, with installation and quick-start instructions in the repository README. The post frames this as providing teams “a concrete starting point” for domain adaptation.