AI Research Trend Atlas

An open dashboard and data pipeline for tracking AI research trends across major conference accepted-paper lists.

AI Research Trend Atlas collects accepted papers from ACL, EMNLP, ICLR, ICML, NeurIPS, and NAACL from 2023 onward, classifies them into research-topic categories, and serves a searchable dashboard for exploring what the field is paying attention to.

This is an independent research utility. It is not affiliated with ACL Anthology, OpenReview, or any listed conference.

Live Demo

Open the dashboard here:

https://likeacloud7.github.io/ai-trend/

The app is designed for quick answers to questions like:

• Which topic categories dominate accepted papers this year?
• How are LLMs, agents, multimodal models, safety, evaluation, and efficiency changing over time?
• How do topic distributions differ across ACL, EMNLP, ICLR, ICML, NeurIPS, and NAACL?
• Which papers sit behind each aggregate trend?

Current Snapshot

The committed dataset currently contains:

Metric	Value
Papers	51,960
Conferences	6
Years	2023-2026
Collected source-years	18
Dashboard dataset	public/data/dashboard.json

Coverage depends on whether a conference has published accepted-paper lists for a given year. Missing or unpublished source-years are recorded in data/source-status.json instead of being fabricated.

Features

• Accepted-paper collection from ACL Anthology event pages and OpenReview venue IDs.
• Main / Findings split for ACL-family venues where Findings volumes are available.
• Topic classification using an explicit, editable keyword taxonomy.
• Conference-year matrix for comparing coverage and paper counts.
• Topic trend chart for seeing share changes over time.
• Paper-level search over titles, authors, keywords, conferences, tracks, and categories.
• Reproducible JSON outputs for dashboard use, audits, and downstream analysis.
• Automation-ready refresh flow for scheduled GitHub Actions or local cron jobs.

Data Sources

Conference	Source	Tracks
ACL	ACL Anthology event pages	Main, Findings
EMNLP	ACL Anthology event pages	Main, Findings
NAACL	ACL Anthology event pages	Main, Findings
ICLR	OpenReview venue IDs	Main
ICML	OpenReview venue IDs	Main
NeurIPS	OpenReview venue IDs	Main

ACL-family pages use event URLs such as:

https://aclanthology.org/events/acl-2025/

OpenReview venues use IDs such as:

ICLR.cc/2026/Conference

Topic Taxonomy

Papers are assigned one primary category and may also carry secondary categories. The taxonomy lives in scripts/taxonomy.mjs, so it can be reviewed, edited, and versioned like code.

Current categories include:

Category	Examples of signals
LLMs & Foundation Models	language models, instruction tuning, scaling, LoRA
Agents & Tool Use	tool use, autonomous agents, workflows, multi-agent systems
Retrieval & Knowledge	RAG, retrieval, knowledge graphs, search, reranking
Multimodal & Vision-Language	VLMs, image-text, video-language, VQA
Reasoning & Planning	chain-of-thought, math reasoning, planning, code generation
Evaluation & Benchmarks	benchmarks, metrics, leaderboards, robustness
Alignment, Safety & Trust	hallucination, bias, privacy, jailbreaks, RLHF
Efficient AI & Systems	quantization, distillation, serving, latency, edge
Generative Models	diffusion, GANs, video generation, audio generation
RL, Robotics & Control	reinforcement learning, robotics, policies, control
Core NLP Tasks	translation, summarization, parsing, dialogue, NLI
Multilingual & Low-Resource	cross-lingual, dialects, typology, low-resource languages
Speech & Audio	ASR, TTS, speech translation, music, acoustic modeling
Data, Annotation & Synthetic Data	data curation, weak supervision, active learning
Interpretability & Analysis	probing, attribution, mechanistic interpretability
Optimization & Theory	convergence, learning theory, Bayesian methods
Domain AI & Science	medicine, biology, chemistry, climate, education
Other / General Methods	fallback when no stronger category matches

The classifier is intentionally transparent. It is useful for trend exploration, not as a definitive scientific taxonomy.

Quick Start

npm install
npm run data:update
npm run data:verify
npm run dev

The Vite dev server runs locally at:

http://127.0.0.1:5173/

Build the static app:

npm run build

Preview the production build:

npm run preview

Useful Commands

Command	Description
npm run dev	Start the local Vite app
npm run build	Build the production dashboard
npm run preview	Preview the built dashboard
npm run data:update	Re-fetch sources, classify papers, and write JSON outputs
npm run data:verify	Validate the generated dataset for consistency

You can override the year range when updating data:

YEAR_START=2023 YEAR_END=2026 npm run data:update

Generated Files

The data updater writes several artifacts:

Path	Purpose
data/papers.json	Compact paper corpus for reviews, diffs, and audits
data/full/*.json	Full enriched source shards split by conference-year
data/source-status.json	Collection status for each conference-year
data/run-summary.json	Small summary of the latest collection run
public/data/dashboard.json	UI-optimized dataset consumed by the React dashboard

public/data/dashboard.json is the only file the browser app needs at runtime.

Repository Structure

.
|-- data/
|   |-- full/                 # Full per conference-year shards
|   |-- papers.json           # Compact paper corpus
|   |-- run-summary.json      # Latest run metadata
|   `-- source-status.json    # Source-year collection status
|-- public/
|   `-- data/dashboard.json   # Dataset loaded by the dashboard
|-- scripts/
|   |-- taxonomy.mjs          # Topic taxonomy and classifier
|   |-- update-data.mjs       # Source collectors and dataset builder
|   `-- verify-data.mjs       # Dataset integrity checks
|-- src/
|   |-- App.jsx               # Dashboard app shell
|   |-- fallbackDashboard.js  # Minimal fallback dataset
|   |-- main.jsx              # React entrypoint
|   `-- styles.css            # Dashboard styles
`-- vite.config.js

Automation

The refresh flow is intentionally simple:

1. Run npm run data:update.
2. Run npm run data:verify.
3. Build the static dashboard with npm run build.
4. Commit changed JSON and static output.

This makes the project easy to run from GitHub Actions, cron, or another scheduler. The same pipeline can be mirrored into a GitHub Pages repository or hosted as a standalone static Vite app.

Interpreting the Data

This project optimizes for transparent trend analysis, not perfect bibliographic authority.

• Always follow each paper's source URL for canonical metadata.
• Some future conference-year pages may be missing because accepted lists have not been published yet.
• OpenReview venue metadata can vary by year and conference.
• Topic classification is keyword-based and intentionally inspectable.
• Counts can change when conferences update proceedings pages or OpenReview metadata.

If you use this dataset for a report or blog post, cite the original paper pages alongside this project.

Contributing

Contributions are welcome, especially:

• better topic taxonomy keywords,
• new conference collectors,
• source parsing fixes,
• dashboard UX improvements,
• validation checks,
• documentation and examples.

Suggested workflow:

git checkout -b feature/your-change
npm install
npm run data:verify
npm run build

For taxonomy changes, include a short note explaining what changed and why.

Roadmap

• Add more conferences and workshops.
• Add per-conference topic trend pages.
• Add exportable CSV and Parquet artifacts.
• Add topic co-occurrence views.
• Add manual review tooling for taxonomy calibration.
• Add stable citation metadata for dataset releases.

Citation

If this project helps your work, please cite the repository and link to the live dashboard:

@software{ai_research_trend_atlas,
  title = {AI Research Trend Atlas},
  author = {LikeACloud7},
  url = {https://github.com/LikeACloud7/ai-trend},
  year = {2026}
}

License

Released under the MIT License.