Abstract
Researchers struggle to efficiently locate and manage relevant literature within the exponentially growing body of scientific publications. We present WisPaper, an intelligent academic retrieval and literature management platform that addresses this challenge through three integrated capabilities: (1) Scholar Search, featuring both quick keyword-based and deep agentic search modes for efficient paper discovery; (2) Library, a customizable knowledge base for systematic literature organization; and (3) AI Feeds, an intelligent recommendation system that automatically delivers relevant new publications based on user interests. Unlike existing academic tools, WisPaper provides a closed-loop workflow that seamlessly connects literature discovery, management, and continuous tracking of research frontiers. Our multilingual and multidisciplinary system significantly reduces the time researchers from diverse backgrounds spend on paper screening and management, enabling them to focus on their core research activities. The platform is publicly accessible and serves researchers across academia and industry.
Introduction
Modern researchers face an overwhelming volume of scientific literature. Traditional academic search engines, while powerful, often fall short in two critical areas: semantic precision and workflow integration. Keyword-based systems frequently return papers that contain the right terms but fail to address the specific intersection of concepts a researcher is investigating. Furthermore, the research process remains fragmented, requiring users to manually bridge the gap between discovery tools, reference managers, and alert systems.
WisPaper is an end-to-end AI agent system designed to unify these disparate tasks. By leveraging a specialized model (WisModel) trained via reinforcement learning, WisPaper moves beyond simple retrieval to perform "Deep Search"—a process that verifies whether candidate papers actually answer complex research questions using structured reasoning and verbatim evidence.
Methodology: Intent Decomposition and Verification
The core of WisPaper is WisModel, a Large Language Model (LLM) fine-tuned for academic rigor. Unlike general-purpose assistants, WisModel is optimized for two primary tasks: intent decomposition and evidence-backed verification.
When a user enters a natural language query, the system decomposes it into executable Boolean search terms and a set of specific "verification criteria." For instance, a query about psychologist integration in medical teams is broken down into mandatory components that a paper must satisfy to be considered relevant.
To ensure accuracy, WisModel acts as a virtual reviewer, checking candidate papers against these criteria. It provides a verdict (e.g., Support, Reject, or Somewhat Support) backed by direct quotes from the text, effectively mitigating the "hallucination" problem common in standard AI models. This training was achieved using Group Relative Policy Optimization (GRPO), a reinforcement learning technique with a multi-dimensional reward system that penalizes logic errors while rewarding adherence to strict formatting and human-expert reasoning.
The WisPaper Knowledge Loop
WisPaper is structured as a continuous "Knowledge Loop" rather than a static search tool. This ecosystem seamlessly connects literature discovery, management, and tracking.
The loop consists of three integrated components:
- Scholar Search: Offers both quick keyword searches and the agentic Deep Search mode.
- Library: A personalized repository where saved papers are automatically processed for metadata.
- AI Feeds: An intelligent recommendation engine that uses a "User Interest Profile" derived from the Library to push relevant new publications to the researcher daily.
This closed-loop design ensures that the system grows more attuned to the user's specific research frontier over time.
Evaluation and Results
The researchers validated WisPaper using both automated benchmarks and human expert evaluation. On TaxoBench, a benchmark for expert-level literature surveys, WisPaper achieved a recall of 22.26%, outperforming OpenAI’s O3 (20.92%).
In a custom evaluation involving PhD students across 10 disciplines (including Medicine, Physics, and Engineering), WisModel demonstrated exceptional precision. It achieved 93.70% accuracy in matching papers to criteria, significantly outperforming general-purpose models like GPT-4o (70.81%) and Qwen-Max (72.82%). By requiring verbatim evidence snippets for every claim, the system provides a level of transparency and reliability essential for academic work.
Conclusion and Future Work
WisPaper represents a significant step toward an autonomous research assistant. By automating "information logistics," it allows researchers to focus on intellectual synthesis. While the current system excels at discovery and organization, future iterations aim to address the "synthesis" stage—helping researchers write literature reviews or generate new hypotheses. Additionally, the authors plan to expand the database scope and integrate more deeply with "AI Scientist" frameworks to further close the gap between reading and experimentation.
Summary updated