Large language models are increasingly used for technical troubleshooting, but real-world problem descriptions are often incomplete, and multiple causes can produce the same symptoms. This ambiguity makes accurate diagnosis difficult, as the model must decide which of several plausible explanations is correct.
The paper introduces an evidence-first agentic reasoning framework, treating the LLM as an investigator that actively seeks out discriminative evidence before committing to a diagnosis. This approach is tailored for interactive scenarios, where the model can ask questions or request additional information to narrow down the cause.
By shifting the focus from immediate answer generation to evidence gathering, the method aims to make diagnosis more robust in settings where initial information is insufficient. The work highlights the importance of structured reasoning in applying LLMs to complex, real-world problem-solving tasks.