Incidental Alignment: Surfacing and Systematizing Emergent Criteria in Hybrid LLM-Human Evaluation Loops

This research proposes a next-generation "evaluation laboratory" framework that integrates a meta-discovery phase into LLM evaluation pipelines. Instead of relying solely on predefined rubrics, the system iteratively surfaces, organizes, and tests new evaluation criteria that emerge organically during human-in-the-loop judgment. Starting with a HypoEval-like rubric, human evaluators provide scores and open-ended comments, which a secondary LLM or clustering method mines for new, context-specific criteria. Candidate criteria are then validated by evaluators through voting or ranking, and only those with strong consensus are incorporated into the main rubric. The updated rubric is re-evaluated to check for improved alignment and coverage. This dynamic, data-driven approach treats evolving evaluation criteria as valuable signals rather than noise, enhancing adaptability, explainability, and robustness of LLM evaluation systems. The framework can also be extended beyond natural language generation to other modalities like machine vision or speech evaluation.

References:

1. Defining and Detecting Vulnerability in Human Evaluation Guidelines: A Preliminary Study Towards Reliable NLG Evaluation. Jie Ruan, Wenqing Wang, Xiaojun Wan (2024). North American Chapter of the Association for Computational Linguistics.