Source-linked AI summary
Paper Espresso: From Paper Overload to Research Insight
Mingzhe Du, Luu Anh Tuan, Dong Huang, See-kiong Ng
TL;DR
Scientific publishing has outpaced individual researchers’ ability to stay informed. Paper Espresso addresses this gap by continuously curating trending papers, generating structured LLM summaries, and analyzing trends across multiple timescales; over 35 months, it reveals broadening topics, rapid decay, and higher engagement for novel combinations.
Problem
The pace of scientific publishing exceeds individual researchers’ capacity to stay informed, creating information asymmetry and risking redundant efforts, missed cross-pollination, and delayed methodological adoption.
Method
Paper Espresso continuously ingests community-validated trending papers and uses LLMs to produce structured summaries, topical labels, keywords, and multi-scale trend reports.
Results
Over 35 months, Paper Espresso processed 13,388 papers across 6,673 topics and found non-saturating topic emergence, a median topic half-life of 1 month, and positive novelty–engagement correlation.
Takeaways & Limitations
The AI research frontier is broadening rather than converging, while novelty and popularity provide complementary signals for paper recommendation.
Abstract
from arXiv · showhide
The accelerating pace of scientific publishing makes it increasingly difficult for researchers to stay current. We present Paper Espresso, an open-source platform that automatically discovers, summarizes, and analyzes trending arXiv papers. The system uses large language models (LLMs) to generate structured summaries with topical labels and keywords, and provides multi-granularity trend analysis at daily, weekly, and monthly scales through LLM-driven topic consolidation. Over 35 months of continuous deployment, Paper Espresso has processed over 13,300 papers and publicly released all structured metadata, revealing rich dynamics in the AI research landscape: a mid-2025 surge in reinforcement learning for LLM reasoning, non-saturating topic emergence (6,673 unique topics), and a positive correlation between topic novelty and community engagement (2.0x median upvotes for the most novel papers). A live demo is available at https://huggingface.co/spaces/Elfsong/Paper_Espresso.
1 Introduction
Paper Espresso addresses the information overload created by rapidly increasing scientific publishing by continuously analyzing a focused stream of community-trending arXiv papers. Its open structured metadata, multi-scale trend reports, and 35-month deployment reveal changing AI research dynamics.
- Nearly 30,000 arXiv submissions arrive monthly, limiting individual researchers’ ability to stay informed and contributing to redundant effort, missed cross-pollination, and delayed methodological adoption.
- Paper Espresso continuously ingests community-validated trending papers and uses LLMs to generate structured summaries, topical labels, keywords, and emerging-direction reports.
- 13,388 papers, 6,673 topics, and 51,036 authors are represented in the publicly released, continuously updated structured dataset.
- Daily, monthly, and lifecycle analyses consolidate topics at multiple granularities so researchers can track evolving directions without manual search.
- Over 35 months, the system identified a mid-2025 reinforcement-learning surge for LLM reasoning, non-saturating topic emergence, cross-cutting methodologies, and frequency–engagement divergence.
2 System Architecture
Paper Espresso uses modular ingestion, LLM processing, and interactive presentation layers to turn a curated paper stream into structured summaries and trend views. Its processing supports multimodal paper analysis, topic consolidation, and multi-view browsing.
- Architecture: The architecture has data ingestion, AI processing, and interactive presentation layers backed by CLI pipelines and a Streamlit frontend.
- Data ingestion: The system samples roughly 2–3% of arXiv through the community-curated Hugging Face Daily Papers API, using upvotes as a lightweight attention proxy.
- Paper processing: Each paper’s title, abstract, and available full PDF are sent in one multimodal request that returns summaries, pros/cons, open-vocabulary topics, and technical keywords.
- Trend analysis: Monthly reports consolidate hundreds of fine-grained topics into roughly 20 coherent clusters while preserving mappings to the original per-paper labels.
- Presentation: Daily, monthly, and lifecycle views present ranked papers, trend summaries, topic keywords, Hype Cycle charts, and topic time series.
3 Datasets
Paper Espresso releases complementary, continuously updated datasets that preserve paper summaries, trend reports, and lifecycle snapshots in reproducible date-partitioned formats. The datasets connect coarse trends to individual paper labels and lifecycle statistics.
- Dataset organization: Three complementary datasets are continuously updated through automated pipelines and stored as date-partitioned Parquet files.
- Paper summaries: The paper-summary dataset combines original metadata with LLM-generated TL;DRs, structured analyses, open-vocabulary topics, and technical keywords.
- Dataset statistics: Table 1 reports dataset statistics for the May 2023–April 2026 collection period.
- Trending reports: Monthly trending records provide consolidated topic clusters and mappings back to constituent per-paper labels for drill-down.
- Lifecycle snapshots: Lifecycle snapshots store per-topic classifications, monthly counts, and corpus-level statistics that support the Hype Cycle visualization and lifecycle analysis.
4 Empirical Analysis
Across 35 months, Paper Espresso reveals a rapidly evolving and diversifying AI research landscape, with shifting topic prominence, fast topic decay, and engagement patterns that reward novelty beyond raw frequency.
- Scope: 35 months of deployment cover paper volume, engagement, topic evolution, co-occurrence, lifecycle, velocity, and novelty.The analysis spans May 2023 to April 2026 across four stated dimensions.
- Topic landscape: 13,388 papers generate 6,673 fine-grained topics, later consolidated into 15–20 coherent clusters representing roughly 50:1 compression.Open-vocabulary labels count semantically equivalent variants separately before monthly consolidation.
- Temporal evolution: By mid-2025, Reinforcement Learning surged to the top as GRPO and RLVR gained adoption for LLM reasoning, while Efficient Inference steadily gained traction.VLMs remained consistently prominent during this shift.
- Topic emergence and diversity: 19–408 new topics emerge monthly while Shannon entropy remains around 7.9 bits, indicating continued diversification rather than consolidation.The reported entropy range is 6.9–8.6 bits.
- Topic co-occurrence: Reinforcement Learning co-occurs most with LLMs, VLMs, Multimodal LLMs, and Mathematical Reasoning, while Diffusion Models strongly pairs with Video Generation.The Diffusion–Video pair has the second-highest reported Jaccard similarity, 0.13.
- Topic lifecycle and velocity: Topics reach peak prominence after a median of 8 months but lose half their prominence within 1 month, although Instruction Tuning, 3D Reconstruction, and Efficient Inference persist longer.Their reported half-lives are 7, 6, and 4 months respectively.
- Takeaways: The AI research frontier is broadening rather than converging, so timely monitoring and recommendations should consider peripheral topics, novelty, volume, and per-paper engagement.The paper explicitly connects rapid topic decay to the need for timely awareness and distinguishes popularity from engagement.
- Novelty and engagement: Novelty correlates positively with engagement: unexpected topic combinations receive 2.0× the upvotes of conventional combinations, while niche topics can outperform the most frequent topic.The most frequent topic, LLMs, accounts for 13.6% of papers; Pre-training Strategies and GUI Agents receive 2–4× higher median upvotes.
5 Related Work
Prior systems support paper indexing, implementation linking, recommendation, summarization, and trend analysis through reactive or post-hoc workflows. Paper Espresso instead combines structured summarization with proactive monitoring and temporal trend reporting.
- Existing scholarly tools provide large-scale indexing, AI-generated TLDRs, implementation links, personalized recommendation, citation-graph navigation, literature-review retrieval, and citation-grounded writing.
- Unlike free-form summarizers, Paper Espresso produces structured JSON summaries, pros/cons, and topics for programmatic filtering and aggregation.
- Classical and neural trend-analysis methods rely on topic modeling, bibliometric mapping, citation bursts, or embedding-based temporal models.
- Paper Espresso uses LLMs for real-time topic labeling and consolidation, generating human-readable trend reports within hours rather than relying on post-hoc analysis.
- Its lifecycle view classifies topics into five phases using peak timing, decline ratio, recent trend slope, and paper-count-scaled dots.
- Novelty-versus-engagement analysis compares papers with more novel topic combinations against their community upvotes.
6 Conclusion
Paper Espresso converts AI papers into structured summaries and multi-granularity trend reports, while its 35-month deployment reveals rapid topic decay and sustained topic emergence. Unconventional topic combinations receive higher median engagement, and the system releases its code and data publicly.
- Paper Espresso is an open-source system for structured summaries and multiple-granularity trend reports from the daily AI-paper stream.
- 6,673 unique labels indicate non-saturating topic emergence across 35 months of deployment.
- A one-month median topic half-life indicates rapid topic decay.
- 2.0× median upvotes for unconventional topic combinations show a positive novelty-engagement effect.
- All code, data, and a live demo are publicly available.