Abstract
Multi-hop Question Answering (QA) requires retrieving multiple supporting facts to reach a correct conclusion. However, iterative retrieval systems often suffer from "cascading failures," where an initial error misleads subsequent search steps. We present Reasoning Path Augmentation (RPA), a framework that addresses this by reordering reasoning paths to prioritize lexically accessible facts and augmenting training data with distractor sentences to improve error recovery. RPA achieves state-of-the-art performance on the QASC and MultiRC benchmarks, demonstrating that training on failure is a key feature for robust retrieval.
Introduction
Multi-hop QA is a complex task akin to a scavenger hunt, requiring the identification of a sequence of justifications across a large corpus. Most current systems employ iterative retrieval: finding one fact, then using it as context to find the next. The primary weakness of this approach is its fragility; if the retriever selects a "distractor" sentence early on—a sentence that looks relevant but doesn't lead to the answer—the entire reasoning chain often collapses. This phenomenon, known as cascading failure, occurs because subsequent searches are conditioned on the initial error.
The RPA Framework
To mitigate these failures, the RPA framework introduces two key innovations: Reasoning Path Reordering and Reasoning Path Augmentation. These methods aim to stabilize the retrieval process and teach the model how to recover when it inevitably encounters noise.
Reasoning Path Reordering
Standard datasets often provide reasoning paths in an arbitrary or chronological order. However, not all facts are equally easy to find. RPA restructures these paths based on "difficulty," which is proxied by the lexical overlap (shared words) between the query and the justifications. By placing "easier" facts—those with higher word overlap—at the beginning of the sequence, the model establishes a reliable foundation. This reduces the search space and makes it significantly easier to retrieve more semantically distant, "hard-to-fetch" facts in later hops.
As shown in Figure 1, the system utilizes a specialized "exclusive-or" ($\oplus$) operation. This operation deletes overlapping words and adds unique words from retrieved facts to the query. This ensures that the retriever focuses on new, unsearched information at each step, preventing it from getting stuck in a loop of redundant information.
Iterative Retrieval and Augmentation
RPA employs a RoBERTa-Large encoder within an ANCE (Approximate Nearest Neighbor) retrieval structure. ANCE allows for efficient searching across millions of documents by representing them as dense vectors. The process is iterative, updating the query representation at each hop based on the previously retrieved facts.
To enhance robustness against the aforementioned cascading failures, RPA utilizes Reasoning Path Augmentation. During training, the model is intentionally exposed to "noisy" paths. Instead of only seeing "perfect" gold-standard paths, the authors insert top-ranked incorrect sentences (distractors) into the reasoning chain. This forces the model to learn that even if a distractor is present in its "memory," it must remain focused on finding the remaining necessary justifications. This "training on failure" allows the retriever to develop a recovery mechanism essential for handling real-world data.
Experimental Results
The effectiveness of RPA was evaluated on two major multi-hop benchmarks:
- QASC (Question Answering via Sentence Composition): A large-scale dataset requiring the composition of two facts from a 17-million-sentence knowledge base.
- MultiRC (Multi-Sentence Reading Comprehension): A dataset where questions are supported by paragraphs containing 2 to 4 justifications.
RPA set new performance benchmarks on both:
- On QASC, RPA achieved 81.5% accuracy, significantly outperforming the previous state-of-the-art JointRR (78.0%).
- On MultiRC, it reached an F1m score of 78.3, surpassing the AIR model (77.2).
Ablation studies confirmed that both reordering and augmentation contribute to these gains. The researchers found that reordering alone improved retrieval recall by simplifying the initial search. Furthermore, increasing the proportion of augmented (noisy) training data—up to 90%—consistently improved the model's ability to complete reasoning paths successfully, suggesting that exposure to errors is a critical component of model robustness.
Conclusion and Future Work
RPA demonstrates that "easier-first" heuristics and intentional exposure to errors during training are powerful tools for stabilizing multi-hop retrieval. While the current implementation relies on simple lexical overlap for reordering, future iterations could explore more nuanced semantic difficulty measures. Additionally, addressing the initial retrieval bottleneck—where the system is limited by the quality of the first heuristic search—remains a key area for further research to ensure that the "gold" facts are always within the retriever's reach.
Summary updated