KnowMe Paper Radar: Lightweight Discovery of Promising Papers in the Age of Autoresearch

Zhuo Chen, Jian Wang, Jinjie Gu, Lijun WuarXiv 2026 · arXiv

In the fast-moving world of AI research, the "human year" has been compressed into an "AI day." As the volume of publications explodes, researchers and AI agents alike face a daunting challenge: how do you find the truly promising work before it becomes a citation classic?

A new paper from researchers at Ant Group introduces KnowMe Paper Radar, a lightweight discovery tool designed to prioritize your reading list by looking beyond simple keyword matches and into the "who" and "where" of research.

TL;DR

KnowMe Paper Radar is a lightweight discovery tool that prioritizes new research by combining topical relevance with deep contextual signals like author history, collaboration networks, and institutional track records. It significantly improves retrieval quality (boosting NDCG@10 from 0.412 to 0.598) while remaining efficient enough to run on a standard workstation without needing constant LLM inference.

What Problem Does This Paper Address?

We are currently in the "Age of Autoresearch," where the sheer scale of literature is compounding. As shown in the authors' analysis of arXiv trends, primary computer science submissions increased by a staggering 36.5% in just one year (2025 to 2026).

Figure 1: Monthly observations and a fitted trend continuation. (a) Raw monthly all-arXiv and primary computer-science (CS) submissions through October 2026 [1, 2]. From November 2026, the dashed segment shows the saved model’s trailing twelve-month mean; shading spans three fitted models. January 2023 marks the fitted period boundary. (b) Fitted annualized rates with 95% intervals and a separate observed January–October comparison. Paired bars compare raw June–October CS counts in 2025 and 2026, following the same left-to-right time order. The cumulative CS share uses the matched January–October totals. Assumptions and validation appear in Appendix D.

The problem isn't just finding a paper that matches a query; it's distinguishing between "topical fluency" (papers that sound right) and "methodological substance" (papers that actually work). Traditional citation-based metrics are "lagging indicators"—they only help months or years after a paper is published. KnowMe aims for early discovery, using evidence available at the moment of publication to identify work worth building upon.

Methods

The core of KnowMe is a decomposed scoring model that breaks down a paper’s priority into four interpretable components:

  1. Base Score ($s_{base}$): Content relevance (title/abstract) and recency.
  2. Author Adjustment ($\Delta_a$): Based on the author’s specific research history in that topic.
  3. Institutional Context ($G$): The track record of the lab or company in that specific research direction.
  4. Manual Preferences ($\Delta_m$): Explicit user-defined priorities or exclusions.

To power this, the system builds a Source Graph (Figure 3) that maps "author-paper-author" paths. This allows the system to discover "non-seed" authors—emerging researchers who are collaborating with established experts you already follow.

Figure 2: Architecture of daily discovery and source reuse.

Crucially, KnowMe is designed for efficiency. It uses automatic acquisition and graph-cache reuse to ensure that most daily triage tasks don't require expensive Large Language Model (LLM) inference or high-end GPUs.

Datasets

The authors evaluated KnowMe using a massive collection of scholarly data:

  • arXiv Records: 174,640 acquired records, including a 13-month cohort focused on LLM-related research.
  • AMiner: Used as a source for historical author catalogs and institutional priors.
  • CSFCube: An expert relevance benchmark used to test how well the tool's retrieval matches human expert judgment.
  • OpenAlex: Used to provide citation edges for the graph.

Key Insights and Ideas

  • Abstracts are Essential: One of the most significant "aha" moments in the paper is the quantitative proof that titles alone are insufficient for discovery. Moving from title-only to title+abstract matching dramatically improved retrieval performance across all tested models.
  • The "Who" Matters for "What": By incorporating author and institutional history, the tool can surface high-quality work from unfamiliar names. If a new author publishes a paper with a lab known for rigorous evaluation in "Multimodal Fusion," that paper receives a contextual boost.
  • Agent-Centric Design: KnowMe isn't just for humans. It includes a structured Command-Line Interface (CLI) that outputs JSONL, allowing it to serve as an "upstream" tool for Recursive Self-Improvement (RSI) workflows, where AI agents autonomously find and learn from new literature.
  • Bounded Influence: To prevent "authority bias" (only seeing famous authors), the scoring model caps the influence of historical reputation. This ensures that even a Nobel laureate's paper must still be topically relevant to rank highly.

Key Results

The researchers demonstrated that KnowMe is both more accurate and more efficient than standard baselines:

  • Retrieval Accuracy: On expert-judged tasks, using cached abstracts improved the Normalized Discounted Cumulative Gain (NDCG@10) from 0.412 to 0.598.
  • Efficiency: By reusing the source graph and context caches, the system achieved a threefold speedup in preparing research context for new queries.
  • Scale: The system successfully managed a library of over 174k records, identifying 1,471 unique "institution-direction" contexts (e.g., "Stanford's expertise in RL reasoning").
Figure 3: An overview and an inspectable local author path.

Limitations and Future Work

While powerful, KnowMe currently relies on the accuracy of external catalogs like AMiner for its "cold start" author data. If an author isn't in the catalog, the system has to "discover" them through collaboration paths, which takes time.

The authors suggest that the next logical step is to integrate outcome-driven updates. By connecting a reading decision to a later research outcome (e.g., "I read this paper and it successfully improved my model"), the system could autonomously refine its institutional and author priors over time, creating a truly self-evolving research radar.

Summary updated

Abstract

The growing volume of topically similar papers makes it difficult to identify new work that can offer useful research insights or methodological guidance. We present KnowMe Paper Radar, a lightweight discovery tool that combines multiple sources of evidence into personalized reading priorities. Users specify topics, priorities, exclusions and dates; the system brings together topical relevance, recency, author research histories, collaboration paths and direction-specific institutional context. By exposing each score component and its supporting sources, KnowMe helps readers select promising papers for their current questions and discover relevant work beyond familiar authors. Automatic acquisition, version merging and graph-cache reuse keep this workflow convenient without large language model (LLM) inference during ordinary use. Experiments on 174,640 acquired records show more than threefold faster repeated context preparation, while cached abstracts improve normalized discounted cumulative gain at rank 10 (NDCG@10) from 0.412 to 0.598 on expert-judged retrieval tasks. Complementary analyses examine how the same accumulated records support semantic retrieval, research-history modeling and literature-grounded planning. A browser interface and a structured command-line interface (CLI) make these capabilities available to both researchers and research agents, positioning KnowMe as a practical upstream tool for literature-informed recursive self-improvement (RSI) workflows.
BibTeX
@misc{chen2026knowme,
  author = {Zhuo Chen and Jian Wang and Jinjie Gu and Lijun Wu},
  title = {{KnowMe Paper Radar: Lightweight Discovery of Promising Papers in the Age of Autoresearch}},
  year = {2026},
  eprint = {2611.23702},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.23702},
}
Figure