Paper deep dive
Paper Espresso: From Paper Overload to Research Insight
Mingzhe Du, Luu Anh Tuan, Dong Huang, See-kiong Ng
Intelligence
Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 97%
Last extracted: 4/10/2026, 2:52:09 AM
Summary
Paper Espresso is an open-source platform that automates the discovery, summarization, and trend analysis of trending arXiv papers. By leveraging LLMs for structured analysis and multi-granularity trend reporting, the system provides insights into the AI research landscape, including topic emergence, lifecycle dynamics, and the correlation between research novelty and community engagement.
Entities (5)
Relation Signals (3)
Paper Espresso → analyzes → arXiv
confidence 100% · Paper Espresso, an open-source platform that automatically discovers, summarizes, and analyzes trending arXiv papers.
Paper Espresso → uses → Large Language Models
confidence 100% · The system uses large language models (LLMs) to generate structured summaries
Reinforcement Learning → surgedin → mid-2025
confidence 90% · a mid-2025 surge in reinforcement learning for LLM reasoning
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
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 this https URL.
Tags
Links
- Source: https://arxiv.org/abs/2604.04562v1
- Canonical: https://arxiv.org/abs/2604.04562v1
Trouble viewing inline? Open PDF directly →
Full Text
38,635 characters extracted from source content.
Expand or collapse full text
Paper Espresso: From Paper Overload to Research Insight Mingzhe Du mingzhe@nus.edu.sg National University of Singapore Singapore, Singapore Anh Tuan Luu anhtuan.luu@ntu.edu.sg Nanyang Technological University Singapore, Singapore Dong Huang dhuang@nus.edu.sg National University of Singapore Singapore, Singapore See-Kiong Ng seekiong@nus.edu.sg National University of Singapore Singapore, Singapore Abstract The accelerating pace of scientific publishing makes it increas- ingly difficult for researchers to stay current. We present Paper Espresso, an open-source platform that automatically discov- ers, summarizes, and analyzes trending arXiv papers. The sys- tem 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 con- tinuous 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 be- tween topic novelty and community engagement (2.0×median upvotes for the most novel papers). A live demo is available at https://huggingface.co/spaces/Elfsong/Paper_Espresso. CCS Concepts • Information systems→Information extraction; Summa- rization; Web applications. Keywords paper summarization, trend analysis, knowledge discovery, large language models, research tools 1 Introduction The pace of scientific publishing now outstrips any individual re- searcher’s capacity to stay informed. As shown in Figure 1, arXiv alone receives nearly 30,000 submissions per month [1], with no sign of deceleration. This creates an acute information asymme- try: the collective frontier advances rapidly, yet each researcher’s awareness lags behind, filtered through keyword alerts and so- cial media curation. The cost is not merely inconvenience but re- dundant efforts, missed cross-pollination, and delayed adoption of methodological advances. Existing platforms such as Semantic Scholar [3], Papers with Code [32], and ArXiv Sanity [22], along with LLM-powered tools like PaSa [15], LitLLM [2], and Scholar- Copilot [35], address fragments of this problem (indexing, retrieval, or writing assistance) but remain fundamentally reactive: they re- quire researchers to already know what to look for. None provides proactive, continuous monitoring that combines structured paper comprehension with temporal trend analysis. 24-0424-0524-0624-0724-0824-0924-1024-1124-1225-0125-0225-0325-0425-0525-0625-0725-0825-0925-1025-1125-1226-0126-0226-03 17500 20000 22500 25000 27500 30000 arXiv (all fields) Paper Espresso 150 300 450 600 750 900 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. We present Paper Espresso, an open-source system that con- tinuously ingests community-validated trending papers, distills each into a structured summary, and proactively surfaces emerg- ing research directions. Instead of indexing the full arXiv firehose, it targets the∼2–3% curated by the Hugging Face Daily Papers community and applies LLM-powered analysis to produce sum- maries, topical labels, keywords, and multi-scale trend reports. Af- ter 35 months of uninterrupted deployment, the system has grown into both a practical daily tool and a longitudinal observatory of the AI research landscape. It makes three contributions: (1)Open structured dataset. We publicly release a structured dataset of LLM-generated paper summaries, topical labels, and keywords on Hugging Face (13,388 papers, 6,673 topics, 51,036 authors), continuously updated via automated pipelines. (2)Multi-granularity trend analysis. The system surfaces trend- ing research directions at daily, monthly, and lifecycle scales through LLM-driven topic consolidation, enabling researchers to track the evolving landscape without manual search. (3)Longitudinal empirical analysis. Over 35 months of de- ployment, we reveal dynamics in the AI research landscape: a mid-2025 surge in reinforcement learning for LLM reasoning, non-saturating topic emergence, a topic co-occurrence map exposing cross-cutting methodologies and emerging niches, and a divergence between topic frequency and engagement. arXiv:2604.04562v1 [cs.DL] 6 Apr 2026 Du et al. HF Datasets Memory · 5 mins TTL Session Cache HF Hub Parquet · Cross Instance Access 4.2) Data Persistence4.1) Data Caching Data Ingestion Paper Processing Insight Presentation arXiv Data Fetcher Gemini / OpenAI / Anthropic LLM APIs Paper Summary Topcis Keywords Pros & Cons Paper Reader Daily Trending Monthly Trending Topic Aggregation Trend Analyzer Raw Paper PDF HF Daily Papers Trending Paper Metadata 1) User Submission 2) Batch Collection 3) Paper Reading 4) Data Upload 5) Data Retrival Research Trending ☕ Paper Espresso EN / CN Paper Summarization Research Gartner Cycle Topic Co-occurrence Data Persistence Disk · Atomic I/O Local Cache Caching and Persistence Summary / Trending / Lifecycle 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. 2 System Architecture The system is organized as modular CLI-driven pipelines (daily, monthly, and lifecycle) backed by a Streamlit 1 web frontend. All data is persisted to four public Hugging Face datasets in date-partitioned Parquet format, ensuring full reproducibility. As shown in Figure 2, the system comprises three layers: data ingestion, AI processing, and interactive presentation. 2.1 Data Ingestion Layer Processing all∼30,000 monthly arXiv submissions is neither feasible nor necessary; most researchers need only the high-impact subset. We therefore source papers from the Hugging Face Daily Papers API 2 , a community-curated feed where users upvote notable arXiv preprints. This yields a focused stream of∼2–3% of arXiv (Figure 1), with upvote counts serving as a lightweight proxy for community attention. For each paper, the system captures the title, authors, abstract, arXiv identifiers, publication date, upvotes, and (when available) the full PDF for multimodal analysis. 2.2 Paper Processing Layer The processing layer invokes LLMs via LiteLLM [5], decoupling the data processing pipeline from any model provider. A two-tier cache (local JSON checkpoints and remote Hub lookups) makes processing idempotent, so the pipeline skips already-summarized papers and resumes cleanly after any interruption. Paper Summarization. Each paper’s title, abstract, and (when available) full PDF are sent as a single multimodal request. PDF grounding enables the model to capture methodological details beyond the abstract. The returned JSON contains: (1) a concise summary (2–4 sentences), (2) a detailed pros/cons analysis, (3) open- vocabulary topic labels (2–3 free-form strings, not from a fixed taxonomy), and (4) technical keywords (4–6 canonical terms, e.g., “LoRA,” “GRPO,” “DiT” ). 1 https://streamlit.io 2 https://huggingface.co/papers Trend Analysis. Daily reports distill the day’s papers into dom- inant themes, a ranked topic list, and trending keywords. Open- vocabulary labeling naturally yields hundreds of fine-grained topics per month, far too many for direct browsing, so monthly reports au- tomatically consolidate them into∼20 coherent clusters (e.g., “Mul- timodal LLMs” and “Vision-Language Models (VLMs),”→“VLMs” ), with an explicit topic mapping back to the original per-paper la- bels. A bimonthly lifecycle pipeline then classifies each topic into Gartner Hype Cycle [16] phases using purely statistical indicators (Section 4), requiring no additional LLM calls. Bilingual Output. To serve both English-speaking and Chinese- speaking research communities, all LLM-generated fields are pro- duced in both languages within a single call, eliminating a separate translation step. Chinese variants are stored alongside their English counterparts with a _zh suffix. 2.3 Presentation Layer The web interface exposes three views. The Daily view lists pa- pers sorted by upvotes, each rendered as a card with topic pills, the author list, and expandable TL;DR and pros/cons panels. The Monthly view deduplicates papers across the month and prepends an LLM-generated trend summary with ranked topics and key- words. The Lifecycle view presents a Gartner Hype Cycle chart alongside per-topic time-series of paper counts and proportions. 3 Datasets Paper Espresso publicly releases three complementary datasets on HF Hub, continuously updated via the automated pipelines described in Section 2. All datasets are stored as date-partitioned Parquet files. Table 1 summarizes key statistics and Table 2 provides the complete field schema. Paper Summaries (hf_paper_summary). Original paper meta- data includes title, authors, abstract, publish date, upvotes, and full PDF. LLM-generated fields include a summary (2–4 sentence Paper Espresso: From Paper Overload to Research Insight Table 1: Dataset statistics (May 2023 – April 2026). DatasetRecordsSplits hf_paper_summary13,388733 days hf_paper_daily_trending733733 days hf_paper_monthly_trending3434 months hf_paper_lifecycle1818 bi-months Aggregate StatisticsCount Unique papers13,388 Unique authors51,036 Unique Fine-grained topics40,565 Unique Coarse-grained topics6,673 Avg. fine-grained topics / paper3.03 Avg. coarse-grained topics / month18.5 Avg. upvotes23.4 TL;DR), a structured detailed analysis, open-vocabulary topics (2–3 labels), and keywords (4–6 terms). Trending Reports (hf_paper_daily/monthly_trending). Each daily or monthly record contains a trending summary, ranked top topics, and trending keywords. Monthly records additionally provide a topic mapping that traces each of the∼20 consolidated clusters back to its constituent per-paper labels, enabling drill-down from coarse themes to individual papers. Lifecycle Snapshots (hf_paper_lifecycle). Bimonthly snapshots store per-topic lifecycle classifications, monthly topic counts, and corpus-level statistics. These snapshots power the Hype Cycle visu- alization in the web interface and the lifecycle analysis in Section 4. 4 Empirical Analysis Our analysis spans 35 months of deployment (May 2023 to April 2026) and covers four dimensions: (1) paper volume growth and community engagement patterns, (2) topic distribution, temporal evolution, and co-occurrence structure dynamics. (3) topic lifecycle classification and velocity, and (4) the relationship between paper novelty and community engagement. 4.1Paper Volume and Community Engagement Monthly intake grew from 259 papers in May 2023 to a peak of 923 in October 2025 (Figure 1), averaging 18.8 papers on weekdays versus 3.3 on weekends, consistent with the academic publishing cycle. As shown in Figure 4, community upvotes are heavily right-skewed (skewness=5.28): the median paper receives 13 upvotes, yet the 90th percentile reaches 52 and the maximum upvote is 664. This long tail means that upvotes carry genuine discriminative power: a uniformly distributed signal would make ranking meaningless, but the concentration of attention on the top 10% of papers creates a clear separation between high-impact work and the majority, validating upvote-based ranking as a practical curation signal. 4.2 Topic Landscape and Dynamics Topic Distribution. With an average of 3.03 topic labels per paper, the system produces 6,673 unique fine-grained topics across Table 2: Field schema of the four released datasets. FieldType Description Paper Summaries (hf_paper_summary) paper_idstrarXiv identifier titlestrPaper title authorslistList of author names abstractstrOriginal abstract upvotesintCommunity vote count published_atdatePublication timestamp concise_summarystrTL;DR (avg. 551 chars) detailed_analysisstrPros/cons analysis (avg. 1,827 chars) topicslistFine-grained topic labels (avg. 3.03) keywordslistExtracted keywords Daily Trends (hf_paper_daily_trending) trending_summarystrNarrative overview of daily themes top_topicslistRanked dominant topics keywordslistTrending keywords of the day daily_reportstrHuman-readable daily report Monthly Trends (hf_paper_monthly_trending) trending_summarystrMonthly trend narrative top_topicslistConsolidated topic clusters (15–20) topic_mappingdictMaps consolidated labels to originals monthly_reportstrDetailed monthly analysis Lifecycle Snapshots (hf_paper_lifecycle) lifecycle_datadictPer-topic phase, peak, slope, counts sorted_monthslistOrdered month labels in snapshot topics_by_monthdictTopic counts per month total_by_monthdictTotal topic mentions per month n_papersintCumulative paper count at snapshot n_monthsintNumber of months in snapshot 13,388 papers (Table 1). Because labels are open-vocabulary (Sec- tion 2), lexically distinct but semantically equivalent labels (e.g., “VLMs” vs. “Vision-Language Models” ) are counted separately; the monthly consolidation step merges such variants, reducing hun- dreds of labels to 15–20 coherent clusters (∼50:1 compression). Table 3 lists the five most frequent consolidated topics, which col- lectively cover over 56% of all papers. Table 3: Top-5 consolidated research topics by paper count. TopicCount% Cum.% Large Language Models1,81913.613.6 Vision-Language Models1,59811.925.5 Diffusion Models1,51411.336.8 Multimodal LLMs1,34510.046.8 Reinforcement Learning1,2689.556.3 Topic Temporal Evolution. Figure 3 shows how topic domi- nance shifts over time. In early 2025, Large Language Models and Diffusion Models led the landscape. By mid-2025, Reinforcement Learning surged to the top, driven by rapid adoption of Group Rela- tive Policy Optimization (GRPO) and Reinforcement Learning with Du et al. 23-05/0623-07/0823-09/1023-11/1224-01/0224-03/0424-05/0624-07/0824-09/1024-11/1225-01/0225-03/0425-05/0625-07/0825-09/1025-11/1226-01/0226-03/04 0.0 1.5 3.0 4.5 6.0 7.5 % of papers Efficient Fine-tuning Model Evaluation Code Generation Video Generation Efficient Inference Diffusion Models Large Language Models Multimodal LLMs Vision-Language Models Reinforcement Learning Rising Stable Declining Figure 3: Bimonthly proportion (%) of the top-10 research topics from May 2023 to March 2026, smoothed with a Gaussian kernel (휎= 0.8) for visual clarity. Trend arrows in the legend indicate each topic’s recent trajectory. 050100150 0 1000 2000 Number of Papers Histogram CDF 0.00 0.25 0.50 0.75 1.00 CDF P50=13 P90=52 Figure 4: Community engagement distribution. The his- togram (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. Verifiable Rewards (RLVR) for LLM reasoning. VLMs remain con- sistently prominent, while Efficient Inference gains steady traction as deployment-oriented research matures. Topic Emergence and Diversity. As shown in Figure 5, new topics appear at a rate of 19–408 per month with no sign of satura- tion, while Shannon entropy퐻=− Í 푖 푝 푖 log 2 푝 푖 over the monthly topic-frequency distribution remains stable around 7.9 bits (range 6.9–8.6). Together these indicate that the research frontier continues to diversify rather than collapsing toward a few dominant themes. Topic Co-occurrence. Figure 6 shows raw co-occurrence counts (lower triangle) and Jaccard similarity퐽= |퐴∩ 퐵|/|퐴∪ 퐵|(up- per triangle) for the top-20 topics. Raw counts reflect absolute volume but are biased toward frequent topics; Jaccard normalizes by union size, revealing whether two topics co-occur more than their individual base rates would predict. Three patterns emerge: 23-0523-0924-0124-0624-1025-0225-0625-1026-02 0 100 200 300 400 New Topics / Month New topics / month Shannon entropy 7.00 7.25 7.50 7.75 8.00 8.25 8.50 Entropy (bits) mean=7.9 Figure 5: Topic emergence and diversity. Red bars show the number of new topics each month; the blue line tracks Shan- non entropy of the monthly topic distribution, which re- mains flat around 7.9 bits, confirming sustained diversity. (1) RL as cross-cutting methodology: Reinforcement Learning has the highest co-occurrence with LLMs (215), VLMs (152), Multi- modal LLMs (132), and Mathematical Reasoning (123), permeating nearly every major direction. (2) Generative-vision cluster: Diffusion Models pairs strongly with Video Generation (197) and Text-to- Image (71), with the Diffusion–Video pair also showing the second- highest Jaccard (0.13), reflecting genuine technical coupling. (3) Fre- quency is not affinity: the top-count pair (RL + LLMs, 215) has only moderate Jaccard (0.09) because both topics are individually com- mon, whereas Embodied AI and Vision-Language-Action Models share the highest Jaccard (0.14) from just 50 papers, exposing a tightly coupled niche invisible to raw counts alone. Keyword Evolution. Tracking keywords within a topic reveals which specific methods drive its rise or fall. Figure 7 traces the top-8 keywords for three major topics. In Reinforcement Learning, Paper Espresso: From Paper Overload to Research Insight Large Language Models Vision-Language Models Diffusion Models Multimodal LLMs Reinforcement Learning Efficient Inference Video Generation Model Evaluation Efficient Fine-tuning Code Generation Mathematical Reasoning LLM Agents LLM Evaluation Retrieval-Augmented Generation Multi-Agent Systems Text-to-Image Generation Embodied AI Mechanistic Interpretability Image Generation Novel View Synthesis Large Language Models Vision-Language Models Diffusion Models Multimodal LLMs Reinforcement Learning Efficient Inference Video Generation Model Evaluation Efficient Fine-tuning Code Generation Mathematical Reasoning LLM Agents LLM Evaluation Retrieval-Augmented Generation Multi-Agent Systems Text-to-Image Generation Embodied AI Mechanistic Interpretability Image Generation Novel View Synthesis 3 1375 216927 22215725132 11234783124 120215261029 63811909314 6757394645103 7026220521117 101712712355106 336334421 28122516 36147191572143386 361718265611866108 233981913101416121 96123017213412325 501375561423532 52732871215632 1451131 0.000.000.000.080.050.000.030.030.030.050.020.020.000.000.03 0.030.060.060.020.010.040.030.010.000.000.010.010.020.030.010.030.00 0.010.010.040.120.000.020.000.000.000.060.000.000.050.03 0.050.020.010.050.030.010.020.000.000.010.010.010.020.000.020.00 0.010.010.010.030.030.080.040.010.010.020.010.010.000.00 0.030.000.000.010.000.000.010.010.010.000.010.02 0.020.010.000.010.020.000.000.02 0.000.010.010.010.020.020.010.010.01 0.010.010.010.000.000.000.020.000.000.01 0.010.030.040.000.030.000.01 0.030.010.01 0.010.010.000.01 0.010.02 0.010.000.000.010.01 0.000.010.00 0.000.00 0255075100125150175200 Co-occurrence count (lower) 0.000.020.040.060.080.100.12 Jaccard similarity (upper) Figure 6: Co-occurrence heatmap for the top-20 topics. The lower triangle shows raw co-occurrence counts (warm col- ors); the upper triangle shows Jaccard similarity (cool colors), highlighting topic pairs that co-occur more than base rates. RLHF [27] (∼25% of RL papers in mid-2024) was rapidly displaced by GRPO [31] (∼65% by early 2025) and RLVR [24], marking a clear pivot from preference-based to verifiable-reward training. Large Language Models mirrors this shift: RLHF and DPO [29] declined while Chain-of-Thought [36], GRPO, and RLVR rose, signaling reasoning-oriented techniques as the new dominant paradigm. In Diffusion Models, the UNet-to-Transformer architectural migration is evident: Stable Diffusion [30] and ControlNet [42] faded while DiT [28] and Flow Matching [26] gained steady traction. 4.3 Topic Lifecycle We adapt the Gartner Hype Cycle [16] to bibliometric data in order to characterize how research topics mature. For every topic with at least 15 papers, we first compute its monthly proportion푝 푡 = 푐 푡 /푁 푡 , where푐 푡 is the number of papers assigned to the topic in month푡and푁 푡 is the total number of topic assignments that month. We then summarize each trajectory with five indicators: the peak proportion푝 ∗ and the month at which it occurs; the current level ̄ 푝 cur , averaged over the most recent 3 months; the decline ratio훿= ̄ 푝 cur /푝 ∗ , capturing how far the topic has fallen from its peak; the trend slope훽, fit by Ordinary Least Squares (OLS) over the last 6 months; and the recent fraction휌, the share of a topic’s papers published in the last 8 months. Based on these indicators, each topic is assigned to one of five lifecycle phases: (1) Innovation Trigger. Newly emerging topics: active for≤8 months, or surging niches with 휌> 0.60 and<200 papers. (2)Peak of Inflated Expectations. Topics near their all-time high (훿>0.70, peak within 6 months) or still rising strongly (훽> 0.001, 훿> 0.65). (3)Trough of Disillusionment. Topics well below peak with no sign of recovery (훿<0.65,훽 ≤0.0003), or actively declining (훽<−0.001, 훿< 0.75). (4) Slope of Enlightenment. Topics that have declined from peak but show renewed growth (훿< 0.65, 훽> 0.0003). (5) Plateau of Productivity. Mature, stable topics that match none of the above conditions. Figure 8 maps notable topics to the lifecycle. Reinforcement Learn- ing [14,33], Efficient Inference [12,44], and LLM Agents [19,20,39] sit at the Peak, consistent with the mid-2025 surge in Figure 3. LLMs [21,43], VLMs [41], and Diffusion Models [38] have entered the Trough, their proportional share declining even as absolute counts grow. Knowledge Distillation [18] and Code Generation [13] occupy the Slope of Enlightenment, finding renewed applica- tions after earlier decline, while Mechanistic Interpretability [4,37] has reached a stable Plateau. Vision-Language-Action Models [23] and World Models [11] appear at the Innovation Trigger, marking nascent research fronts. Topic Velocity. For each topic with≥15 papers and≥4 active months, we measure time to peak (months from first appearance to maximum proportion) and half-life (months from peak to 50% of peak). As shown in Figure 10, the contrast is stark: the median time to peak is 8 months, but the median half-life is just 1 month. AI research topics rise gradually yet decline abruptly, losing half their prominence within a single month of peaking. A few practically grounded topics resist this pattern, notably Instruction Tuning (7- month half-life), 3D Reconstruction (6), and Efficient Inference (4). 4.4Paper Novelty and Community Engagement We investigate whether papers with unusual topic combinations attract more community attention. For each paper with at least two topic labels, we define a novelty score as the negated mean Point- wise Mutual Information (PMI) across all co-assigned topic pairs: PMI(푡 푖 ,푡 푗 )= log 2 푃(푡 푖 ,푡 푗 ) 푃(푡 푖 ) 푃(푡 푗 ) , where co-occurrence probabilities are estimated from the full corpus with Laplace smoothing (훼=0.5) for unseen pairs. Papers combining commonly co-occurring topics score low; those with unexpected pairings score high. As shown in Figure 9, novelty correlates positively with engage- ment. Frequency and engagement also diverge: Large Language Models is the most common topic, yet niche topics like Pre-training Strategies (55), Computer Use Agents (38), and Agentic Reason- ing (36) far exceed the global median of 14. Novelty and popularity thus carry complementary signals for paper recommendation. 4.5 Takeaways (1) The AI research frontier is broadening, not converging. New topics emerge at an undiminished rate (up to 408/month) while Shannon entropy remains stable (∼7.9 bits), indicating sustained diversification rather than consolidation around a few dominant themes. Researchers should actively monitor peripheral topics to avoid tunnel vision. Du et al. 0 50 % of papers Reinforcement Learning GRPO Chain-of-Thought RLVR RLHF PPO Reinforcement Learning (RL) Curriculum Learning Vision-Language Models 0 10 20 % of papers Large Language Models Chain-of-Thought GRPO LoRA RLHF RLVR DPO Instruction Tuning RAG 23-0523-0623-0723-0823-0923-1023-1123-1224-0124-0224-0424-0524-0624-0724-0824-0924-1024-1124-1225-0125-0225-0325-0425-0525-0625-0725-0825-0925-1025-1125-1226-0126-0226-03 0 20 % of papers Diffusion Models Diffusion Transformer (DiT) LoRA Stable Diffusion ControlNet Flow Matching Diffusion Transformer Video Diffusion Models Latent Diffusion Models 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. (2) Topics peak slowly but fade fast. The median topic takes 8 months to reach peak prominence yet loses half of it within a single month, making timely awareness critical. Systems that report trends only retrospectively (e.g., annual surveys) risk delivering insights after the window of opportunity has closed. (3)Novelty attracts attention. Papers combining unexpected topic pairs receive 2.0×the upvotes of those with conventional combinations. This suggests that the community rewards cross- pollination, and that recommendation systems should surface surprising intersections, not just popular categories. (4)Popularity and engagement are distinct signals. The most frequent topic (LLMs, 13.6% of papers) is far from the most engaging per paper; niche topics such as Pre-training Strategies and GUI Agents draw 2–4×higher median upvotes. Effective curation must weigh both volume and per-paper impact. 5 Related Work Academic Paper Discovery. Semantic Scholar [3] offers large- scale indexing with AI-generated TLDRs [8], Papers with Code [32] links papers to implementations, and ArXiv Sanity [22] pioneered SVM-based personalized recommendation. LLM-era tools extend this landscape: PaSa [15] navigates citation graphs, LitLLM [2] ap- plies RAG to literature reviews, and ScholarCopilot [35] fine-tunes a 7B model for citation-grounded writing. These systems are funda- mentally reactive, requiring users to know what to search for. Paper Espresso fills a different niche: proactive daily monitoring that com- bines structured summarization with temporal trend analysis, so researchers discover what matters without issuing a query. Scientific Document Summarization. Prior work ranges from discourse-aware attention models [10] and extreme summariza- tion [8] to LLM-based scholarly review [25], with recent surveys charting this evolution [40]. Unlike free-form summarizers, Paper Espresso produces structured JSON output (summaries, pros/cons, topics), enabling programmatic filtering and aggregation. Research Trend Analysis. Classical approaches include LDA [6] for topic modeling, VOSviewer [34] for bibliometric mapping, and CiteSpace [9] for citation burst detection. Neural topic models such as BERTopic [17] and its temporal extension BERTrend [7] offer embedding-based alternatives. Our system takes an orthogonal approach: instead of post-hoc analysis, it uses LLMs for real-time topic labeling and consolidation as papers are published, producing human-readable trend reports within hours. Paper Espresso: From Paper Overload to Research Insight Visibility Innovation Trigger Peak of Inflated Expectations Trough of Disillusionment Slope of Enlightenment Plateau of Productivity Maturity 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. Novelty Upvotes Spearman = 0.185 p < 10 98 n = 13,013 95% CI OLS fit Figure 9: Novelty vs. engagement. Papers with more novel topic combinations (higher scores) receive more upvotes. 6 Conclusion Paper Espresso is an open-source system that converts the daily stream of AI papers into structured summaries and multiple granu- larity trend reports. Analysis over 35 months reveals non-saturating topic emergence (6,673 unique labels), rapid topic decay (median half-life of one month), and a positive novelty-engagement effect (2.0×median upvotes for unconventional topic combinations). All code, data, and a live demo are publicly available. Time to PeakHalf-life 0 5 10 15 20 25 Months median=8 mo median=1 mo Figure 10: Topic velocity. Topics take 8 months to peak (red) yet lose half their prominence within a single month (blue). References [1][n. d.]. arXiv Monthly Submission Statistics. https://arxiv.org/stats/monthly_ submissions. Accessed: 2026-04-02. [2]Shubham Agarwal, Issam H. Laradji, Laurent Charlin, and Christopher Pal. 2024. LitLLM: A Toolkit for Scientific Literature Review. arXiv preprint arXiv:2402.01788 (2024). [3]Waleed Ammar, Dirk Groeneveld, Chandra Bhagavatula, Iz Beltagy, Miles Craw- ford, Doug Downey, Jason Dunkelberger, Ahmed Elgohary, Sergey Feldman, Vu Ha, Rodney Kinney, Sebastian Kohlmeier, Kyle Lo, Tyler Murray, Hsu-Han Ooi, Matthew Peters, Joanna Power, Sam Skjonsberg, Lucy Lu Wang, Chris Wilhelm, Zheng Yuan, Madeleine van Zuylen, and Oren Etzioni. 2018. Construction of the Du et al. Literature Graph in Semantic Scholar. arXiv preprint arXiv:1805.02262 (2018). [4]Leonard Bereska and Efstratios Gavves. 2024. Mechanistic Interpretability for AI Safety – A Review. Transactions on Machine Learning Research (2024). [5] BerriAI. 2025. LiteLLM: A Unified Interface for LLM APIs. https://github.com/ BerriAI/litellm. [6]David M. Blei, Andrew Y. Ng, and Michael I. Jordan. 2003. Latent Dirichlet Allocation. Journal of Machine Learning Research 3 (2003), 993–1022. [7]Allaa Boutaleb, Jerome Picault, and Guillaume Grosjean. 2024. BERTrend: Neural Topic Modeling for Emerging Trends Detection. In Proceedings of the Workshop on Future Directions in Event Detection (FuturED). [8]Isabel Cachola, Kyle Lo, Arman Cohan, and Daniel Weld. 2020. TLDR: Extreme Summarization of Scientific Documents. In Findings of the Association for Com- putational Linguistics: EMNLP 2020. 4766–4777. [9]Chaomei Chen. 2006. CiteSpace I: Detecting and Visualizing Emerging Trends and Transient Patterns in Scientific Literature. Journal of the American Society for Information Science and Technology 57, 3 (2006), 359–377. [10]Arman Cohan, Franck Dernoncourt, Doo Soon Kim, Trung Bui, Seokhwan Kim, Walter Chang, and Nazli Goharian. 2018. A Discourse-Aware Attention Model for Abstractive Summarization of Long Documents. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 615–621. [11] Jingtao Ding, Yunke Zhang, Yu Shang, Jie Feng, Yuheng Zhang, Zefang Zong, Yuan Yuan, Hongyuan Su, Nian Li, Jinghua Piao, Yucheng Deng, Nicholas Sukien- nik, Chen Gao, Fengli Xu, and Yong Li. 2025. Understanding World or Predicting Future? A Comprehensive Survey of World Models. Comput. Surveys (2025). [12] Mingzhe Du, Anh Tuan Luu, Bin Ji, Qian Liu, and See-Kiong Ng. 2024. Mercury: A Code Efficiency Benchmark for Code Large Language Models. In Advances in Neural Information Processing Systems, Vol. 37. [13]Mingzhe Du, Anh Tuan Luu, Bin Ji, Xiaobao Wu, Yuhao Qing, Dong Huang, Terry Yue Zhuo, Qian Liu, and See-Kiong Ng. 2025. CodeArena: A Collective Evaluation Platform for LLM Code Generation. In Proceedings of the 63rd An- nual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations). Association for Computational Linguistics, Vienna, Austria, 502–512. doi:10.18653/v1/2025.acl-demo.48 [14] Mingzhe Du, Anh Tuan Luu, Yue Liu, Yuhao Qing, Dong Huang, Xinyi He, Qian Liu, Zejun Ma, and See-Kiong Ng. 2025. Afterburner: Reinforcement Learning Facilitates Self-Improving Code Efficiency Optimization. arXiv preprint arXiv:2505.23387 (2025). [15] Zhangyin Feng, Yuqi Huo, Teng Fei, Jiawei Zhang, et al.2025. PaSa: An LLM Agent for Comprehensive Academic Paper Search. arXiv preprint arXiv:2501.10120 (2025). [16]Jackie Fenn and Mark Raskino. 2008. Mastering the Hype Cycle: How to Choose the Right Innovation at the Right Time. Harvard Business Press. [17] Maarten Grootendorst. 2022. BERTopic: Neural Topic Modeling with a Class- Based TF-IDF Procedure. arXiv preprint arXiv:2203.05794 (2022). [18] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. 2015. Distilling the Knowledge in a Neural Network. arXiv preprint arXiv:1503.02531 (2015). [19]Dong Huang, Mingzhe Du, Jie M. Zhang, Zheng Lin, Meng Luo, Qianru Zhang, and See-Kiong Ng. 2025. Nexus: Execution-Grounded Multi-Agent Test Oracle Synthesis. arXiv preprint arXiv:2510.26423 (2025). [20] Bin Ji, Huijun Liu, Mingzhe Du, Shasha Li, Xiaodong Liu, Jun Ma, Jie Yu, and See-Kiong Ng. 2025. Towards Verifiable Text Generation with Generative Agent. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 39. [21] Bin Ji, Huijun Liu, Mingzhe Du, and See-Kiong Ng. 2024. Chain-of-Thought Im- proves Text Generation with Citations in Large Language Models. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 38. 18345–18353. [22]Andrej Karpathy. 2021. arxiv-sanity-lite: Tag arxiv Papers of Interest and Get Recommendations. https://github.com/karpathy/arxiv-sanity-lite. [23]Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakr- ishna, Suraj Nair, Rafael Rafailov, Ethan P. Foster, Pannag R. Sanketi, Quan Vuong, Thomas Kollar, Benjamin Burchfiel, Russ Tedrake, Dorsa Sadigh, Sergey Levine, Percy Liang, and Chelsea Finn. 2025. OpenVLA: An Open-Source Vision- Language-Action Model. In Proceedings of The 8th Conference on Robot Learning. [24]Nathan Lambert et al.2024. Reinforcement Learning with Verifiable Rewards. arXiv preprint (2024). [25]Weixin Liang, Yuhui Zhang, Hancheng Cao, Binglu Wang, Daisy Ding, Xinyu Yang, Kailas Vodrahalli, Siber He, Daniel Smith, Yian Yin, Daniel McFarland, and James Zou. 2024. Can Large Language Models Provide Useful Feedback on Research Papers? A Large-Scale Empirical Analysis. NEJM AI 1, 8 (2024). [26] Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. 2023. Flow Matching for Generative Modeling. In International Conference on Learning Representations. [27] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. 2022. Training Language Models to Follow Instructions with Human Feedback. Advances in Neural Information Processing Systems 35 (2022), 27730–27744. [28]William Peebles and Saining Xie. 2023. Scalable Diffusion Models with Trans- formers. In Proceedings of the IEEE/CVF International Conference on Computer Vision. 4195–4205. [29]Rafael Rafailov, Archit Sharma, Eric Mitchell, Stefano Ermon, Christopher D. Manning, and Chelsea Finn. 2023. Direct Preference Optimization: Your Language Model Is Secretly a Reward Model. Advances in Neural Information Processing Systems 36 (2023). [30]Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. 2022. High-Resolution Image Synthesis with Latent Diffusion Mod- els. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 10684–10695. [31]Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Mingchuan Zhang, Y.K. Li, Y. Wu, and Daya Guo. 2024. DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models. arXiv preprint arXiv:2402.03300 (2024). [32]Robert Stojnic, Ross Taylor, Ilia Sucholutsky, Douwe Kiela, et al.2019. Papers with Code. (2019). https://paperswithcode.com. [33]Richard S. Sutton and Andrew G. Barto. 2018. Reinforcement Learning: An Intro- duction (2nd ed.). MIT Press. [34] Nees Jan van Eck and Ludo Waltman. 2010. Software Survey: VOSviewer, a Computer Program for Bibliometric Mapping. Scientometrics 84, 2 (2010), 523– 538. [35]Yubo Wang, Xueguang Ma, Ping Nie, et al.2025. ScholarCopilot: Training Large Language Models for Academic Writing with Accurate Citations. arXiv preprint arXiv:2504.00824 (2025). [36]Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc V. Le, and Denny Zhou. 2022. Chain-of-Thought Prompting Elicits Reasoning in Large Language Models. Advances in Neural Information Processing Systems 35 (2022), 24824–24837. [37] Zhaomin Wu, Mingzhe Du, See-Kiong Ng, and Bingsheng He. 2026. Beyond Prompt-Induced Lies: Investigating LLM Deception on Benign Prompts. In Inter- national Conference on Learning Representations. [38]Ling Yang, Zhilong Zhang, Yang Song, Shenda Hong, Runsheng Xu, Yue Zhao, Wentao Zhang, Bin Cui, and Ming-Hsuan Yang. 2024. Diffusion Models: A Comprehensive Survey of Methods and Applications. Comput. Surveys 56, 4 (2024), 1–39. [39] Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. 2023. ReAct: Synergizing Reasoning and Acting in Language Models. In International Conference on Learning Representations. [40]Haopeng Zhang, Philip S. Yu, and Jiawei Zhang. 2025. A Systematic Survey of Text Summarization: From Statistical Methods to Large Language Models. Comput. Surveys 57, 11 (2025), 1–55. [41] Jingyi Zhang, Jiaxing Huang, Sheng Jin, and Shijian Lu. 2024. Vision-Language Models for Vision Tasks: A Survey. IEEE Transactions on Pattern Analysis and Machine Intelligence 46, 8 (2024), 5625–5644. [42] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. 2023. Adding Conditional Control to Text-to-Image Diffusion Models. In Proceedings of the IEEE/CVF Inter- national Conference on Computer Vision. 3836–3847. [43]Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, Yifan Du, Chen Yang, Yushuo Chen, Zhipeng Chen, Jinhao Jiang, Ruiyang Ren, Yifan Li, Xinyu Tang, Zikang Liu, Peiyu Liu, Jian-Yun Nie, and Ji-Rong Wen. 2023. A Survey of Large Language Models. arXiv preprint arXiv:2303.18223 (2023). [44]Zixuan Zhou, Xuefei Ning, Ke Hong, Tianyu Fu, Jiaming Xu, Shiyao Li, Yuming Lou, Luning Wang, Zhihang Yuan, Xiuhong Li, Shengen Yan, Guohao Dai, Xiao- Ping Zhang, Yuhan Dong, and Yu Wang. 2024. A Survey on Efficient Inference for Large Language Models. arXiv preprint arXiv:2404.14294 (2024).