TLDR: ReportBench is a new benchmark designed to evaluate the quality of research reports generated by Deep Research agents and large language models. It assesses the relevance of cited literature and the factual accuracy of statements by comparing AI-generated reports against expert-authored survey papers from arXiv. The benchmark uses reverse prompt engineering to create tasks and employs a dual validation process for cited and non-cited statements. Initial evaluations show commercial Deep Research agents outperform base models, but challenges like hallucination and over-citation persist, highlighting areas for future improvement in AI-driven knowledge synthesis.
The world of artificial intelligence is rapidly advancing, with new ‘Deep Research agents’ emerging that can perform extensive research tasks in a fraction of the time it used to take humans. These AI systems are designed to synthesize vast amounts of information, conduct academic literature surveys, and even perform industry analyses. However, with such powerful capabilities comes the critical need for rigorous evaluation to ensure their outputs are factually accurate and comprehensive before they are widely adopted.
A new benchmark called ReportBench has been introduced to address this very challenge. Developed by researchers at ByteDance BandAI, ReportBench aims to systematically evaluate the quality of research reports generated by large language models (LLMs) and Deep Research agents. The evaluation focuses on two key areas: the quality and relevance of the literature cited in the reports, and the faithfulness and truthfulness of the statements made within those reports.
ReportBench uses high-quality, published survey papers from arXiv as its ‘gold-standard’ references. The researchers employ a clever technique called reverse prompt engineering. This involves taking these expert-authored survey papers and generating domain-specific prompts from them. These prompts then serve as the tasks for the AI agents, which are asked to generate their own research reports. The generated reports are then compared against the original expert papers.
The evaluation framework within ReportBench is quite sophisticated. It automatically analyzes the AI-generated reports by extracting citations and statements. For statements that include citations, the system checks their faithfulness against the original source documents using semantic matching. For statements that are not cited, the framework uses a voting mechanism across multiple web-connected AI models to verify their factual accuracy. This dual approach ensures a thorough assessment of both cited and non-cited information.
The process of creating the ReportBench dataset involves three main phases. First, high-quality survey papers are identified from arXiv, ensuring they have undergone peer review. Second, prompts are generated from these papers at different levels of detail (sentence-level, paragraph-level, and detail-rich), along with specific constraints like publication cut-off dates. Finally, these prompts are classified into distinct application domains to ensure a balanced and diverse test set.
When testing various Deep Research agents and standalone LLMs, the findings were quite insightful. Commercial Deep Research agents, such as those from OpenAI and Google, consistently generated more comprehensive and reliable reports compared to standalone LLMs that were simply augmented with search or browsing tools. This suggests that the specialized design and fine-tuning of these Deep Research products offer a significant advantage in complex research tasks.
However, the evaluations also highlighted areas where there is still considerable room for improvement. Issues like ‘statement hallucination’ (where the AI makes claims that deviate from the cited source) and ‘citation hallucination’ (where the AI fabricates non-existent references) were observed. For example, one instance showed an AI attributing an author to a paper they didn’t write, and another created a plausible-looking but entirely fake URL for a citation. These findings underscore the ongoing challenges in ensuring complete factual consistency and preventing AI-generated misinformation.
Also Read:
- Benchmarking AI’s Tool-Using Abilities in the Real World
- Measuring AI Agent Performance and Safety in Online Shopping
Despite these challenges, ReportBench provides a valuable tool for the research community. It offers a systematic way to monitor, compare, and improve the reliability of AI systems designed for academic survey tasks. The complete code and data for ReportBench will be released to support reproducible research and community-driven progress in evaluating LLM-based knowledge synthesis. You can find more details about this research paper here: ReportBench: Evaluating Deep Research Agents via Academic Survey Tasks.
