Reference-Free Evaluation

Reference-free evaluation studies how to assess generated outputs and model behavior when gold responses, labels, or human ratings are unavailable or costly. Common approaches include LLM judges and reward models, question-answering and claim-verification checks grounded in the input, task-specific metrics, and intrinsic signals such as representation statistics or behavioral consistency. These methods are applied across summarization, retrieval-augmented generation, translation, speech alignment, image and video generation, and conversational agents, with human judgments and labeled benchmarks often used to validate them. A central tension is that judges can be biased or confidently share the same errors, so calibration, multiple diagnostic dimensions, and selective checks against trusted references remain important.