Wyvern is a multi-agent framework designed to automate the creation of technical reports that integrate text, images, and tables while ensuring the content is grounded in verifiable evidence. The researchers, including Beatrice Alessandra Motetti and colleagues from Politecnico di Torino and Huawei Research, developed this system to help users manage the rapid growth of technical information by filtering, aggregating, and synthesizing data from the web into reliable, multimodal documents.
How Wyvern Works
The framework operates through a structured workflow of specialized Large Language Model (LLM) agents, each assigned to a specific task:
Search and Retrieval: The system retrieves information from the web, parses it into Markdown, and uses a series of agents to score documents for completeness and relevance. It includes a deduplication process and a final filtering step to remove semantic outliers, ensuring the knowledge base remains focused on the user's topic.
Report Generation: After creating an outline, agents draft sections and integrate visual content. A dedicated pipeline selects relevant figures from source documents, generates descriptions for them, and places them into the report. Another set of agents builds and inserts comparative overview tables to summarize key concepts.
Grounding and Revision: To address the common issue of hallucinations in generative models, Wyvern includes a claims auto-revision stage. The system decomposes the report into atomic claims and verifies each against the cited references. If a claim is unsupported, the system revises or removes it, and subsequently updates headings to maintain consistency.
Performance and Evaluation
The researchers conducted a human evaluation study comparing Wyvern against three existing methods: STORM, WebThinker, and WikiAutoGen. Participants, including researchers and engineers, rated Wyvern’s reports as more useful than those produced by the alternative methods in 63% to 100% of instances. Additionally, users perceived the informativeness of figures in Wyvern’s reports as superior to a recent baseline in 87% of cases.
Automatic evaluations also indicated improvements in citation accuracy. The study reports that Wyvern achieved up to 2.3 times higher citation recall and 1.6 times higher citation precision compared to the baselines.
Franklin Analysis
The evidence suggests that Wyvern’s primary contribution is the integration of an explicit claims-verification stage into a multimodal report-generation pipeline. While many existing tools focus on text generation or basic image inclusion, Wyvern’s use of atomic claim decomposition and NLI-based (Natural Language Inference) verification provides a systematic approach to grounding. The performance gains in citation recall and precision, as reported by the authors, support the conclusion that this multi-agent, modular architecture effectively reduces unsupported statements compared to the tested baselines.
Considerations
The framework relies on the quality of the retrieved information and the capabilities of the underlying LLMs, such as DeepSeek-R1 and DeepSeek-V3, which were used in the implementation. The authors note that while automated tools can assist in managing information, they are not intended to replace human judgment. Furthermore, the system’s performance is dependent on the effectiveness of the parsers used for different document types, such as PDFs and web pages, to extract meaningful content for the agents to process.
Comments (0)
to join the discussion
No comments yet
Be the first to share your thoughts!