Mitigating Preference Leakage via Strict Estimator Separation for Normative Generative Ranking
New method separates supervision and evaluation models to prevent inflated performance scores in LLM-as-a-Judge tasks.
A research team led by Dalia Nahhas, Xiaohao Cai, Imran Razzak, and Shoaib Jameel has published a paper addressing a critical flaw in how large language models are evaluated for normative tasks like cultural relevance ranking. The paper, 'Mitigating Preference Leakage via Strict Estimator Separation for Normative Generative Ranking,' identifies 'preference leakage' as a major problem in current LLM-as-a-Judge setups, where overlapping supervision and evaluation models artificially inflate performance metrics. To solve this, the researchers formalized cultural relevance as a within-query ranking task and introduced a novel two-judge framework that enforces a strict separation of duties between the model that provides supervision (Judge B) and the independent model used for final evaluation (Judge A).
The team validated their framework on a new, large-scale benchmark called NGR-33k, containing 33,052 culturally grounded stories. Their experiments showed that while classical baselines offered only modest improvements, a dense bi-encoder model (specifically BGE-M3) distilled from a Judge-B-supervised Cross-Encoder was highly effective. Crucially, although the Cross-Encoder provided a strong signal for distillation, the distilled BGE-M3 model substantially outperformed it when evaluated under the leakage-free Judge A protocol. The framework was further validated on the human-curated Moral Stories dataset, showing strong alignment with human norms. The results demonstrate that rigorous evaluator separation is essential for credible GenIR evaluation and that nuanced cultural preferences can be successfully distilled into efficient, deployable ranking models without data contamination.
- Introduces a strict two-judge framework separating supervision (Judge B) and evaluation (Judge A) to prevent circularity and preference leakage in LLM evaluations.
- Validated on a new NGR-33k benchmark of 33,052 culturally grounded stories, where a distilled BGE-M3 model outperformed a Cross-Encoder under leakage-free evaluation.
- Proves subtle cultural preferences can be distilled into efficient rankers, with strong alignment to human norms on the Moral Stories dataset.
Why It Matters
Provides a more rigorous, unbiased method for evaluating AI systems on sensitive tasks like cultural relevance, crucial for building trustworthy applications.