Learning to Judge: LLMs Designing and Applying Evaluation Rubrics
This addresses the issue of static human rubrics misaligned with LLM representations for researchers in natural language generation evaluation, though it is incremental in proposing a new method for a known bottleneck.
The paper tackled the problem of LLMs designing and applying their own evaluation rubrics, finding that they reliably generate interpretable, task-aware criteria with consistent scoring within models, but reliability degrades in factual settings, with GPT-4o outperforming open-weight models like Llama in agreement and generalization.
Large language models (LLMs) are increasingly used as evaluators for natural language generation, applying human-defined rubrics to assess system outputs. However, human rubrics are often static and misaligned with how models internally represent language quality. We introduce GER-Eval (Generating Evaluation Rubrics for Evaluation) to investigate whether LLMs can design and apply their own evaluation rubrics. We evaluate the semantic coherence and scoring reliability of LLM-defined criteria and their alignment with human criteria. LLMs reliably generate interpretable and task-aware evaluation dimensions and apply them consistently within models, but their scoring reliability degrades in factual and knowledge-intensive settings. Closed-source models such as GPT-4o achieve higher agreement and cross-model generalization than open-weight models such as Llama. Our findings position evaluation as a learned linguistic capability of LLMs, consistent within models but fragmented across them, and call for new methods that jointly model human and LLM evaluative language to improve reliability and interpretability.