Evidence-first agent architecture reduces LLM sycophancy in technical diagnosis

LLM-as-an-Investigator introduces a Solution Investigator Agent that quantifies problem ambiguity, generates competing hypotheses, asks targeted clarifying questions, and updates hypothesis probabilities iteratively before committing to a solution. The approach directly targets user-driven sycophancy, where LLMs prematurely adopt unverified user-provided explanations.