Hurdles to Progress in Long-form Question Answering

Long-form question answering has evolved from early retrieval-based systems like DPR and REALM to more sophisticated models leveraging large pre-trained language models such as BERT, GPT, and T5. The emergence of datasets like Natural Questions and ELI5 has driven progress, emphasizing the importance of generating detailed, contextually relevant answers. Recent benchmarks like KILT integrate retrieval and generation, aiming to evaluate models holistically. Despite these advances, challenges persist in ensuring models genuinely utilize external knowledge, interpretability, and evaluation reliability. Many systems achieve high scores by exploiting dataset biases or optimizing metrics without true knowledge grounding, raising concerns about their practical deployment.

The core challenge lies in the disconnect between model performance metrics and actual knowledge utilization. Many models, despite high scores, generate answers that are not based on retrieved documents, partly due to dataset overlaps and the limitations of automatic evaluation metrics like ROUGE-L. Additionally, the heavy overlap between training and validation sets in datasets like ELI5 diminishes the necessity for external retrieval, further complicating the assessment of true retrieval dependence. Human evaluation methods are also unreliable, as they often fail to distinguish whether answers are grounded in external knowledge. These issues hinder the development of trustworthy, explainable LFQA systems suitable for real-world applications.

The paper introduces a novel framework combining a sparse Routing Transformer for long sequence modeling with a contrastively trained retriever (C-REALM) that leverages large-batch contrastive learning. It emphasizes the importance of control experiments—specifically, replacing retrieved documents with random samples—to assess whether models genuinely depend on external knowledge. This approach is innovative in systematically exposing the gap between performance metrics and actual knowledge grounding, setting a new standard for evaluating retrieval-augmented systems. Additionally, the work highlights the significance of dataset quality, advocating for rigorous data cleaning to reduce overlaps, and calls for improved evaluation metrics beyond ROUGE-L.

• �� Develop a C-REALM retriever using large-batch contrastive learning (batch size 12,288) to embed questions and documents into a shared 128-dimensional space, optimizing the similarity via a contrastive loss.
• �� Use a sparse Routing Transformer (RT) with 22 layers, 1032 hidden units, and a maximum sequence length of 8192 tokens to encode long Wikipedia paragraphs.
• �� Pretrain RT on PG-19, a long-form dataset from Gutenberg books, to encourage fluent, long outputs.
• �� Fine-tune the combined system by concatenating retrieved documents with questions, training the RT to generate answers without backpropagating through the retriever.
• �� Conduct control experiments by replacing retrieved documents with random samples during inference, to test reliance on retrieval.
• �� Evaluate using ROUGE-L, R-precision, Recall@5, and human AB tests, analyzing dataset overlaps and retrieval dependence.

The experiments utilize the ELI5 dataset's KILT subset, with 272,634 training, 1,507 validation, and 600 test examples. Multiple retrieval strategies (REALM, C-REALM) and generation sampling p-values (0.6, 0.9) are tested. Metrics include ROUGE-L, R-precision, and Recall@5, complemented by human AB testing and question paraphrase analysis. The study also measures the overlap between validation questions and training data, revealing significant dataset leakage. Ablation studies compare the impact of different retrievers and sampling strategies, demonstrating that models perform similarly whether conditioned on predicted or random retrievals, exposing the weak reliance on external knowledge.

Despite high ROUGE-L scores surpassing human answers, AB tests and random retrieval experiments show answers are not grounded in retrieved documents. Over 81% of validation questions have paraphrased counterparts in training, indicating dataset leakage. Random retrievals have negligible impact on answer quality, exposing the superficial reliance on retrievals. Human evaluations reveal no significant preference between models conditioned on predicted or random retrievals, emphasizing the disconnect between metrics and actual knowledge use. These findings challenge the validity of current evaluation practices and highlight the need for more robust grounding verification.

The insights from this work inform the development of more trustworthy LFQA systems in domains like healthcare, legal advice, and scientific research, where factual accuracy and interpretability are critical. The methodology encourages designing datasets and evaluation protocols that genuinely measure knowledge reliance, fostering AI systems capable of transparent reasoning and reliable information dissemination. Long-term, these advancements could lead to AI assistants capable of autonomous knowledge reasoning, supporting complex decision-making processes across industries.

The primary limitation is that current datasets and training strategies do not enforce genuine reliance on external knowledge, leading to superficial performance. Dataset overlaps diminish the necessity for retrieval, skewing evaluation results. Metrics like ROUGE-L are insufficient for capturing factual correctness in long, open-ended answers. Human evaluation remains unreliable due to subjective judgments and inconsistency. Future work must address dataset quality, develop better grounding metrics, and integrate explicit reasoning modules to ensure models produce truly knowledge-grounded answers.