cs.DLApr 6, 2026
SavePaper Espresso: From Paper Overload to Research Insight
Organizations: National University of Singapore Singapore, Singapore · Nanyang Technological University Singapore, Singapore
Abstract
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.
Figures & tables
Figure 1. Monthly paper volume: arXiv total (red, left axis) vs. Paper Espresso (blue, right axis). Although Paper Espresso selects only community-trending papers ( 2–3% of arXiv), the two curves exhibit a consistent co-trend, confirming that the curated subset tracks the broader publishing rhythm. A dual-axis line chart showing monthly submission counts from April 2024 to March 2026. The red line (left axis) represents total arXiv submissions ranging from 17,000 to 30,000. The blue line (right axis) represents HF Daily Papers curated by Paper Espresso ranging from 150 to 920. Both lines show an upward trend over time.
Figure 2. System architecture of Paper Espresso . The data ingestion layer fetches papers from the Hugging Face Daily Papers API and arXiv. The AI processing layer uses Google Gemini to generate structured summaries and trend analyses. The presentation layer provides an interactive Streamlit interface with multi-granularity browsing. A system architecture diagram showing three layers: data ingestion from Hugging Face API and arXiv on the left, AI processing with Google Gemini in the middle, and the Streamlit web interface on the right. Arrows show data flow between components, with Hugging Face Hub datasets used for persistent storage.
| Dataset | Records | Splits |
| hf_paper_summary | 13,388 | 733 days |
| hf_paper_daily_trending | 733 | 733 days |
| hf_paper_monthly_trending | 34 | 34 months |
| hf_paper_lifecycle | 18 | 18 bi-months |
| Aggregate Statistics | Count | |
| Unique papers | 13,388 | |
Table 1. Dataset statistics (May 2023 – April 2026).
| Field | Type | Description |
| Paper Summaries ( hf_paper_summary ) | ||
| paper_id | str | arXiv identifier |
| title | str | Paper title |
| authors | list | List of author names |
| abstract | str | Original abstract |
| upvotes | int | Community vote count |
Table 2. Field schema of the four released datasets.
Figure 3. Bimonthly proportion (%) of the top-10 research topics from May 2023 to March 2026, smoothed with a Gaussian kernel ( ) for visual clarity. Trend arrows in the legend indicate each topic’s recent trajectory. A line chart showing bimonthly proportion trends for the top-10 topics. Each topic in the legend is annotated with a colored trend arrow: green for rising, gray for stable, red for declining.
Figure 4. Community engagement distribution. The histogram (red, left axis) shows a heavily right-skewed upvote distribution; the CDF (blue, right axis) confirms that 50% of papers receive 13 upvotes and 90% receive 52. A dual-axis chart with a red histogram of upvotes on the left axis and a blue CDF curve on the right axis. Vertical dashed lines mark P50=13 and P90=52.
| Topic | Count | % | Cum.% |
|---|---|---|---|
| Large Language Models | 1,819 | 13.6 | 13.6 |
| Vision-Language Models | 1,598 | 11.9 | 25.5 |
| Diffusion Models | 1,514 | 11.3 | 36.8 |
| Multimodal LLMs | 1,345 | 10.0 | 46.8 |
| Reinforcement Learning | 1,268 | 9.5 | 56.3 |
Table 3. Top-5 consolidated research topics by paper count.
Figure 5. Topic emergence and diversity. Red bars show the number of new topics each month; the blue line tracks Shannon entropy of the monthly topic distribution, which remains flat around 7.9 bits, confirming sustained diversity. A dual-axis chart with red bars showing 19–408 new topics per month on the left axis, and a blue line showing Shannon entropy fluctuating between 6.9 and 8.6 bits on the right axis with a dashed mean line at 7.9.
Figure 6. Co-occurrence heatmap for the top-20 topics. The lower triangle shows raw co-occurrence counts (warm colors); the upper triangle shows Jaccard similarity (cool colors), highlighting topic pairs that co-occur more than base rates. A 20x20 split heatmap. The lower triangle uses a yellow-to-red colormap for co-occurrence counts (0–200). The upper triangle uses a yellow-to-blue colormap for Jaccard similarity (0–0.14). Two horizontal colorbars at the bottom label each scale.
Figure 7. Keyword evolution within three major topics. Each line shows the percentage of papers (within that topic) mentioning a given keyword per month. Top: Reinforcement Learning shows a clear RLHF GRPO/RLVR transition. Middle: Large Language Models mirrors this shift. Bottom: Diffusion Models shows the UNet Transformer architectural migration. Three vertically stacked line charts. Each panel tracks the top-8 keywords within one topic over time. The Reinforcement Learning panel shows GRPO surging to 65% while RLHF declines. The LLM panel shows Chain-of-Thought and GRPO rising. The Diffusion Models panel shows DiT and Flow Matching replacing Stable Diffusion.
Figure 8. AI research hype cycle derived from 35 months of topic proportion time series. Topics are classified into five lifecycle phases based on peak timing, decline ratio, and recent trend slope. Dot size is proportional to total paper count. A Gartner-style hype cycle curve with research topics positioned along it. Topics at the Peak include Reinforcement Learning and Efficient Inference; the Trough contains LLMs, VLMs, and Diffusion Models; the Slope of Enlightenment includes Knowledge Distillation and Prompt Engineering; the Plateau includes Mechanistic Interpretability.
Figure 9. Novelty vs. engagement. Papers with more novel topic combinations (higher scores) receive more upvotes. Scatter plot of novelty score vs log-upvotes with a positive-slope OLS linear fit and 95% confidence band.
Figure 10. Topic velocity. Topics take 8 months to peak (red) yet lose half their prominence within a single month (blue). Side-by-side box-and-strip plots. Time to peak: median 8 months, wide spread. Half-life: median 1 month, tightly clustered.
More in this topic
AI-Assisted Scientific Research
AI-assisted scientific research uses machine learning and AI agents to support discovery, analysis, experimentation, and scientific communication.
See all 251 papers