The Consistency Dilemma in LLMs: Generator-Evaluator Agreement and Vulnerability to Mistakes

Marina Mancoridis, Zoë Hitzig

Large language models are increasingly deployed in agentic pipelines that depend on the model evaluating its own outputs without external verification. The reliability of these pip…