LLM evaluators don't score holistically—they implement a mechanistic two-stage pipeline where error detection happens early and rating decisions crystallize in late layers, and fine-tuning reorganizes rather than builds this structure from scratch.
This paper reveals how LLM-based evaluators (like Themis and Prometheus) internally assign quality scores to text summaries. Using causal analysis techniques, the authors discovered these models use a two-stage pipeline: early layers compare errors locally, while later layers integrate signals and finalize ratings.