Source-linked AI summary
WisPaper: Your AI Scholar Search Engine
Li Ju, Jun Zhao, Mingxu Chai, Ziyu Shen, Xiangyang Wang, Yage Geng, Chunchun Ma, Hao Peng, Guangbin Li, Tao Li, Chengyong Liao, Fu Wang, Xiaolong Wang, Junshen Chen, Rui Gong, Shijia Liang, Feiyan Li, Ming Zhang, Kexin Tan, Junjie Ye, Zhiheng Xi, Shihan Dou, Tao Gui, Yuankai Ying, Yang Shi, Yue Zhang, Qi Zhang
TL;DR
Existing academic search relies heavily on keywords and may not verify whether papers address complex research questions, while discovery, organization, and monitoring remain fragmented. WisPaper links these tasks through Scholar Search, Library, and AI Feeds, using WisModel for structured, evidence-backed validation. The system reports 22.26% recall on TaxoBench and 93.70% accuracy in paper validation.
Problem
Academic search engines rely primarily on keywords, and researchers must manually connect separate tools for discovery, organization, and monitoring.
Source: §1 · pp. 1–2Method
WisPaper integrates keyword and Deep Search with a Library and AI Feeds, while WisModel decomposes queries into verification criteria and validates candidate papers through structured reasoning.
Source: §1 · p. 2Results
WisPaper achieves 22.26% recall on TaxoBench and 93.70% accuracy in paper validation, surpassing existing state-of-the-art agents.
Source: §1 · p. 2, §6 · p. 11Takeaways & Limitations
The system connects literature discovery, organization, and monitoring in a knowledge loop whose modules exchange information to enhance one another.
Source: §1 · p. 2, §6 · p. 11
Abstract
from arXiv · showhide
We present \textsc{WisPaper}, an end-to-end agent system that transforms how researchers discover, organize, and track academic literature. The system addresses two fundamental challenges. (1)~\textit{Semantic search limitations}: existing academic search engines match keywords but cannot verify whether papers truly address complex research questions; and (2)~\textit{Workflow fragmentation}: researchers must manually stitch together separate tools for discovery, organization, and monitoring. \textsc{WisPaper} tackles these through three integrated modules. \textbf{Scholar Search} combines rapid keyword retrieval with \textit{Deep Search}, in which an agentic model, \textsc{WisModel}, validates candidate papers against user queries through structured reasoning. Discovered papers flow seamlessly into \textbf{Library} with one click, where systematic organization progressively builds a user profile that sharpens the recommendations of \textbf{AI Feeds}, which continuously surfaces relevant new publications and in turn guides subsequent exploration, closing the loop from discovery to long-term awareness. On TaxoBench, \textsc{WisPaper} achieves 22.26\% recall, surpassing the O3 baseline (20.92\%). Furthermore, \textsc{WisModel} attains 93.70\% validation accuracy, effectively mitigating retrieval hallucinations.
1 Introduction
WisPaper addresses semantic search limitations and fragmented literature workflows with integrated discovery, organization, and monitoring. Its reported results include 22.26% recall on TaxoBench and 93.70% accuracy in criteria matching.
1 Introduction
- Researchers face growing publication volumes and must manually bridge discovery, reference management, and new-publication monitoring, spending time on information logistics.The introduction identifies these as two fundamental challenges: semantic search limitations and workflow fragmentation.Source: §1 · pp. 1–2
- Existing academic search engines rely primarily on keywords, while recent LLM query reformulation approaches still lack deep reasoning to verify papers’ actual relevance.Complex queries may require assessing whether a paper addresses the intersection of concepts rather than merely mentioning terms.Source: §1 · pp. 1–2
- WisPaper integrates Scholar Search, Library, and AI Feeds so papers move from discovery into organization and profile-informed recommendations.The system is presented as a closed loop connecting search, organization, and monitoring.Source: §1 · p. 2
- WisModel decomposes research queries into editable verification criteria and validates candidate papers through structured reasoning.WisPaper’s Scholar Search offers quick keyword retrieval alongside this agent-powered Deep Search mode.Source: §1 · p. 2
- 22.26% recall on TaxoBench and 93.70% accuracy in criteria matching are reported for WisPaper, which the authors say surpasses existing state-of-the-art agents.The introduction presents both as results of the system.Source: §1 · p. 2
2 The WisPaper Knowledge Loop
WisPaper links Scholar Search, Library, and AI Feeds in a shared knowledge loop. Search results populate a personal library and interest profile, while personalized feeds can restart discovery.
2 The WisPaper Knowledge Loop
- WisPaper’s Scholar Search, Library, and AI Feeds share a paper database and exchange information to form a knowledge loop.Papers discovered through search and recommendations can flow into Library, while Library’s profile informs AI Feeds.Source: §2 · p. 2
2.1 Scholar Search
- Deep Search decomposes intent into retrieval queries and verification criteria, then reports per-criterion evidence and ranks candidates by checklist satisfaction.Users can edit the queries and criteria and rerun verification without rephrasing the original request.Source: §2.1 · p. 3
2.2 Library
- Library ingests papers from PDFs, Scholar Search, and AI Feeds, then builds an evolving interest profile from explicit preferences, imported papers, and folder organization.The profile combines topical cues from paper metadata with user-defined folder labels.Source: §2.2 · pp. 3–4
2.3 AI Feeds
- AI Feeds selects broad research sources, then filters new papers against the user’s profile, narrowing typical daily candidates of 50–100 per category to 3–5 matches.The profile is initially seeded by user-defined interests and enriched through Library interactions.Source: §2.3 · p. 4
- A recommended paper can return to Library or prompt a new Scholar Search, restarting the discovery cycle.The search-import-profile-feed cycle is intended to refine subsequent recommendations.Source: §2.3 · p. 4, §2.2 · p. 4
3 WisModel: The Deep Search Agent
WisModel is trained to decompose research queries into verification criteria and assess papers with evidence-backed reports. Its training combines supervised initialization with reinforcement learning rewards for structured, grounded, and logically aligned outputs.
3 WisModel: The Deep Search Agent
- Deep Search is framed as structured reasoning to translate underspecified intents into criteria and validate papers against complex academic texts.The motivation is that standard LLMs may ignore nuanced constraints or fabricate supporting evidence.Source: §3 · p. 4
3.1 Task Formulation & Initialization
- WisModel generates retrieval queries and weighted verification criteria from a raw user query, then produces a report with an assessment, verbatim evidence, and rationale for each criterion.Supervised fine-tuning initializes the model to follow the output schema and map academic discourse to verification criteria.Source: §3.1 · p. 5
3.2 Reinforcement Learning via Multi-Dimensional Shaped Rewards
- GRPO supplements supervised fine-tuning with rewards for format consistency, faithful evidence grounding, logical entailment, and reasoning alignment.The grounding reward compares generated evidence with the source text and penalizes evidence that is not a substring.Source: §3.2 · p. 5
- The evaluation dataset’s query lengths and numbers of criteria are summarized across 10 academic disciplines.Figure 3’s left panel reports average query tokens and its right panel average criteria per query.Source: Figure 3 · p. 6
- Training uses time-dependent reward weights, prioritizing format early and increasing grounding and semantic reasoning weights as KL divergence stabilizes.This curriculum shifts optimization focus from syntax toward deeper semantic reasoning.Source: §3.2 · pp. 5–6
4 Evaluation
The evaluation tests query understanding and criteria generation, paper–criteria matching, and end-to-end paper recall. It uses expert-annotated queries and criteria, then evaluates WisModel and the full Deep Search pipeline.
4 Evaluation
- The experiments assess intent decomposition, evidence-backed validation, and end-to-end retrieval of papers deemed essential by domain experts.TaxoBench’s 3,815 ground-truth papers are mapped to hierarchical taxonomies from 72 surveys across eight computer-science subfields.Source: §4 · p. 6, §4.4 · p. 9
4.1 Evaluation Setup
- The evaluation dataset contains 2,777 authentic Chinese- and English-language queries across 10 disciplines, annotated with search queries and criteria for paper scope.Examples include a computer-science query requiring both in-context learning and inference-time scaling, and a medicine query specifying an ELN guideline and treatment recommendation.Source: §4.1 · p. 6
4.2 Query Understanding and Criteria Generation
- 94.8% Semantic Similarity, 67.7% ROUGE-L, and 39.8% BLEU; WisModel leads all metrics, with reported gains of 4.8%, 15.1%, and 18.3% over the second-best model.DeepSeek-V3.2-Exp reaches 90.2% semantic similarity, while GPT-5.1 combines 87.0% semantic understanding with 13.2% BLEU.Source: §4.2 · p. 8, Table 2 · p. 8
4.3 Paper-Criteria Matching
- 93.70% overall accuracy versus the next-best model’s 73.23%, with WisModel above 90% in all four assessment categories.Baseline performance is especially weak for somewhat-support cases: Gemini-3-Pro reaches 15.9%, compared with 91.1% for support; the authors identify partial-support judgment as a challenge.Source: §4.3 · pp. 8–9, Table 3 · p. 9
4.4 End-to-End Paper Recall
- 22.26% Recall on TaxoBench exceeds O3’s previous-best 20.92% in end-to-end paper retrieval.WisPaper retrieves from a scholar-indexed database, while O3 uses unconstrained web access.Source: Figure 4 · p. 9, §4.4 · pp. 9–10
5 Related Work
Prior work addresses academic retrieval and research assistance through search agents, citation tools, and automated workflows. The related work distinguishes these systems by their support for literature management and researcher control.
5.1 Academic Retrieval and Discovery
- Academic retrieval tools address different parts of discovery: query reformulation and expansion [10] [15], citation-based recommendations, visual citation networks, and agentic search [9].PaSa improves recall over baselines but remains confined to search, without post-retrieval organization or frontier tracking.Source: §5.1 · p. 10
5.2 LLM Agents for Research Assistance
- Research-assistance agents span domain-specific laboratory tools [5] [7] and broader automation systems [13] [3], but the latter prioritize research outputs over literature management or limit user control.Agent Laboratory includes literature review, while emphasizing report writing and new research outputs rather than managing existing literature.Source: §5.2 · p. 10
6 Conclusion
WisPaper closes the loop between search, organization, and monitoring, using structured intent decomposition and evidence-backed verification. It reports 93.70% paper-validation accuracy and 22.26% TaxoBench recall, surpassing existing state-of-the-art agents.
6 Conclusion
- Closing the loop between search, organization, and monitoring is intended to turn passive literature consumption into an active, self-refining workflow.WisPaper is designed to bridge information retrieval and knowledge management.Source: §6 · p. 10
- WisPaper uses structured intent decomposition and evidence-backed verification as central parts of its scholarly-agent approach.The paper describes the agent as explicitly trained for these tasks.Source: §6 · p. 10
- WisPaper’s full pipeline attains a recall of 22.26% on the TaxoBench benchmark, surpassing existing state-of-the-art agents; WisPaper is publicly deployed as a transparent and efficient tool for researchers across disciplines.Source: §6 · p. 11
- WisPaper is publicly deployed as a tool for researchers across disciplines.Source: §6 · p. 11
Appendix
The appendix illustrates WisModel’s query decomposition and paper-criteria validation, and documents related prompts, output format, and decision rules.
Appendix
- For a query about embedding psychologists in medical teams, WisModel generates three Boolean search queries and three weighted validation criteria.The criteria operationalize the information need as verifiable checkpoints.Source: Table 5 · p. 14
- 91.82% accuracy on somewhat_support cases is reported for WisModel, compared with baseline models' 15.9–45.0% accuracy.The example distinguishes clear abstract evidence for one criterion from indirect evidence for another.Source: Table 6 · p. 15
- The appendix includes a paper-validation decision taxonomy, prompts for query decomposition and criteria alignment, and an expected validation-output format.Source: Table 7 · p. 16, Table 8 · p. 17, Table 10 · p. 18