The Distracting Effect: Understanding Irrelevant Passages in RAG

TL;DR

The paper tackles the problem of distracting passages in Retrieval Augmented Generation (RAG) by defining a quantifiable distracting effect $DE_q(p)$ and demonstrating its robustness across diverse LLMs. It develops retrieval- and generation-based strategies to obtain hard distractors and validates their impact across multiple QA benchmarks, showing that strong distractors correlate with degraded answer accuracy. By combining these distracting passages into training data, the authors fine-tune LLMs to become more robust to distraction, achieving notable improvements (up to 7.5% in some settings) on both in-distribution and out-of-distribution tasks. The work provides a practical framework for assessing and leveraging hard distractors to enhance RAG systems, while acknowledging limits in taxonomy coverage and language scope for future extension.

Abstract

A well-known issue with Retrieval Augmented Generation (RAG) is that retrieved passages that are irrelevant to the query sometimes distract the answer-generating LLM, causing it to provide an incorrect response. In this paper, we shed light on this core issue and formulate the distracting effect of a passage w.r.t. a query (and an LLM). We provide a quantifiable measure of the distracting effect of a passage and demonstrate its robustness across LLMs. Our research introduces novel methods for identifying and using hard distracting passages to improve RAG systems. By fine-tuning LLMs with these carefully selected distracting passages, we achieve up to a 7.5% increase in answering accuracy compared to counterparts fine-tuned on conventional RAG datasets. Our contribution is two-fold: first, we move beyond the simple binary classification of irrelevant passages as either completely unrelated vs. distracting, and second, we develop and analyze multiple methods for finding hard distracting passages. To our knowledge, no other research has provided such a comprehensive framework for identifying and utilizing hard distracting passages.

Figures (25)

• Figure 1: Distribution of distracting effect for passages obtained through different methods, as measured by Llama-3.1-8B. Methods are ordered by their mean distracting effect (shown by vertical black lines), with higher means indicating a greater ability to distract the model.
• Figure 2: Average distracting effect at different rank positions for various retrieval methods. Results are shown for Llama-3.1-8B, averaged across datasets. Higher-ranked passages consistently demonstrate greater potential to mislead the model. Similar trends were observed across all tested LLMs (see Figures \ref{['fig:all_distracting_rank']} and \ref{['fig:llama70b_distracting_rank']}).
• Figure 3: Percentage of queries where each method provides the most distracting passage for Llama-3.1-8B. In blue are the times when no other method reaches the same distracting effect, in orange the percentage of times the highest score is shared with other methods. Similar trends were observed across all tested LLMs (see Figures \ref{['fig:all_best_distractors']} and \ref{['fig:llama70b_best_distractors']}).
• Figure 4: Spearman correlation of distracting effect computed using different LLMs (abbreviated, e.g., Llama $\rightarrow$ L). The strong correlations suggest that the distracting effect of a passage is relatively consistent across models despite architectural differences.
• Figure 5: Example showing Falcon-3-3B's responses in two scenarios using the prompt template in Figure \ref{['fig:prompt_ft']}: relevant passage + hard distracting and relevant passage + weak distracting. Left: When the relevant passage is followed by a hard distracting passage (generated by $G^{\text{modal}}$ and classified as hard for 3B models due to their distracting effect $>0.8$), Falcon-3-3B answers "Smokey and the Bandit II" instead of "Honeysuckle Rose", despite having access to the relevant information. Right: When the relevant passage is followed by a weak distracting passage (generated by $G^{\text{rel}}$), the model correctly answers "Honeysuckle Rose".