Insider Knowledge: How Much Can RAG Systems Gain from Evaluation Secrets?

TL;DR

The paper addresses the vulnerability of RAG system evaluation when insider knowledge about LLM judges and nugget frameworks can leak into system development. It introduces Crucible, a RAG system with subversion probes to study three leakage pathways—framework knowledge, prompt knowledge, and gold nugget knowledge—using NeuCLIR as a fully manual nugget testbed and Auto‑Argue for evaluation. The findings show that leaks can markedly inflate evaluation metrics, with gold nugget knowledge producing the largest gains, and they argue for blind, diversified evaluation practices and held‑out artifacts to preserve evaluation integrity. The work highlights practical risks to public evaluation artifacts and offers concrete safeguards, such as blinded evaluation platforms, to ensure that reported progress reflects genuine improvements rather than artifact leakage.

Abstract

RAG systems are increasingly evaluated and optimized using LLM judges, an approach that is rapidly becoming the dominant paradigm for system assessment. Nugget-based approaches in particular are now embedded not only in evaluation frameworks but also in the architectures of RAG systems themselves. While this integration can lead to genuine improvements, it also creates a risk of faulty measurements due to circularity. In this paper, we investigate this risk through comparative experiments with nugget-based RAG systems, including Ginger and Crucible, against strong baselines such as GPT-Researcher. By deliberately modifying Crucible to generate outputs optimized for an LLM judge, we show that near-perfect evaluation scores can be achieved when elements of the evaluation - such as prompt templates or gold nuggets - are leaked or can be predicted. Our results highlight the importance of blind evaluation settings and methodological diversity to guard against mistaking metric overfitting for genuine system progress.

Figures (5)

• Figure 1: Workflow of Crucible and its evaluation in the Auto-Argue system. Crucible ideates nuggets from retrieved documents, extracts and summarizes passages, and assembles them into a cited report. Auto-Argue evaluates each sentence by checking coverage of manual gold nuggets and verifying citation support. It is common that the summary sentence created to address a system nugget, such as "How many RoundUp cases has Bayer lost so far?", also covers gold nuggets for evaluation, such as "What do the RoundUp lawsuits allege?", since both sets capture related aspects of the same topic.
• Figure 2: Example where Crucible correctly guesses a gold nugget ("How much did the court order Bayer to pay Dewayne Johnson?"). RQ3 explores how evaluation outcomes change if RAG systems reliably guess the gold nugget set
• Figure 3: RQ1: Yes, knowledge of the evaluation system is likely to help development of a RAG system that obtains high evaluation scores. Ginger, a recent nugget-first RAG system designed for TREC RAG 24 is expected not to have used any insider knowledge about the Argue evaluation framework. We show that despite conceptual similarities between Ginger and Crucible, the performance characteristics are vastly different.
• Figure 4: RQ2: Yes, filtering candidate extractions with the evaluation prompt for citation support and/or nugget detection will improve the respective evaluation metrics. % indicates relative improvement over "Base".
• Figure 5: RQ3: Yes, system that could predict gold nuggets, obtain a stark increase in nugget-oriented metrics, including Nugget Recall ($\approx +50\%$ over Auto Nuggets) as well as the fraction of sentences that discuss relevant nuggets ($\approx +20\%$).