arxgraph/arxiv-ai-entity-graph

arXiv AI Entity Graph — Models, Frameworks and Benchmarks in CS Research

Structured entity extraction from arXiv CS papers — every AI model, framework, benchmark, and dataset mentioned in paper titles and abstracts, normalized to canonical names with provider, family, and confidence metadata.

Coverage: January 2024 – present · Daily updates · 23 CS categories
Sample: January 2024 + January 2025 + June 2026 (10 days each)

What makes this different

• arXiv has 2 million papers. Anyone can read them. Nobody has turned them into a structured, normalized, daily-updating database of who's using what — until now.
• Every entity mention is normalized: GPT-4, gpt4, GPT-4-turbo all resolve to the same canonical row with family=GPT, provider=OpenAI, model_type=llm. Your queries don't need to handle aliases.
• This is not a retrieval benchmark or a paper corpus. It is an analytics database — built to answer questions like "which models are researchers actually using?", "when did DeepSeek R1 go mainstream?", and "what benchmarks are used to evaluate GPT-4o?"

Key findings (Jan 2024 – Jun 2026)

• GPT dominates at 5,652 papers citing it across the full window — but Alibaba's Qwen family (Qwen3 + Qwen2.5 + Qwen combined = 4,723 papers) rivals OpenAI's total footprint in 2026
• DeepSeek R1: fastest adoption on record — 3 papers in January 2025, 48 in February (16x spike in one month), sustained at 50–90/month since
• o1 enters the top 20 at 620 papers — reasoning models are now mainstream research infrastructure, not a curiosity
• cs.CL overtook cs.LG as the dominant category — language research (27,452 entity rows) now exceeds pure ML (20,040)
• Stable Diffusion at 884 papers — image generation is a serious parallel research track to LLMs

Schema

entities — the core table

Column	Type	Description
paper_id	varchar	arXiv paper ID
version	int	Paper version
paper_year	int	Year derived from paper ID
paper_month	int	Month derived from paper ID
entity_type	varchar	model, framework, benchmark, dataset
entity	varchar	Canonical entity name
confidence	double	1.0 = confirmed · 0.2 = ambiguous · 0.1 = likely false positive
family	varchar	Model family (e.g. GPT, Llama, Qwen)
provider	varchar	Provider (e.g. OpenAI, Meta, Alibaba)
model_type	varchar	llm, reasoning_llm, vlm, embedding, image_gen, speech, ssm
domain	varchar	Benchmark domain (e.g. math, code, reasoning)
ingest_date	date	Partition column — date paper was ingested

papers

Column	Type	Description
paper_id	varchar	arXiv paper ID
title	varchar	Paper title
abstract	varchar	Full abstract
primary_category	varchar	arXiv primary category (e.g. cs.LG)
paper_year	int	Year
published_at	timestamp	Submission date
doi	varchar	DOI if available
pdf_url	varchar	Direct PDF link
is_ai	boolean	Core AI category (cs.AI, cs.CV, cs.CL, cs.NE)
is_ml	boolean	ML category (cs.LG, stat.ML)
is_llm	boolean	LLM-related paper
is_other_cs	boolean	CS but outside core AI/ML

Additional tables: authors, citations, categories, versions, links

Try it in 30 seconds

import duckdb

con = duckdb.connect("arxiv_ai_graph_sample.duckdb")

# Most cited models in AI/NLP papers — pre-built view
con.execute("SELECT * FROM model_citations LIMIT 20").df()

# Or write your own
con.execute("""
    SELECT provider, entity, model_type,
           COUNT(DISTINCT paper_id) AS papers_citing
    FROM entities
    WHERE entity_type = 'model'
      AND confidence = 1.0
      AND paper_year >= 2025
    GROUP BY 1, 2, 3
    ORDER BY papers_citing DESC
    LIMIT 20
""").df()

Sample output — top models, Jan 2024 – Jun 2026

provider	entity	model_type	papers_citing
OpenAI	GPT	llm	5,652
OpenAI	GPT-4	llm	3,549
OpenAI	GPT-4o	llm	2,967
Meta	Llama	llm	1,822
Alibaba	Qwen3	llm	1,782
Meta	Llama 3	llm	1,727
Alibaba	Qwen2.5	llm	1,668
Anthropic	Claude	llm	1,585
Google	Gemini	llm	1,476
Alibaba	Qwen	llm	1,273
Meta	Llama 2	llm	1,141
Meta	Llama 3.1	llm	1,039
DeepSeek AI	DeepSeek R1	llm	1,007
Stability AI	Stable Diffusion	image_gen	884
LLaVA Team	LLaVA	vlm	867
Google	Gemini 2.5	llm	747
Google	Gemma	llm	731
Mistral AI	Mistral	llm	708
OpenAI	o1	reasoning_llm	620
DeepSeek AI	DeepSeek	llm	613

DeepSeek R1 adoption curve

The sharpest adoption spike in the dataset — from 3 papers in January 2025 to 48 in February, sustained at 50–90/month ever since:

Month	Papers citing DeepSeek R1
Jan 2025	3
Feb 2025	48
Mar 2025	49
Apr 2025	47
May 2025	90
Jun 2025	82
Jul 2025	52
Aug 2025	52
Sep 2025	56
Oct 2025	84
Nov 2025	54
Dec 2025	48
Jan 2026	53
Feb 2026	44
Mar 2026	57
Apr 2026	80
May 2026	67
Jun 2026	38

More example queries

-- Framework adoption across ML papers
SELECT entity, provider, model_type AS framework_type,
       COUNT(DISTINCT paper_id) AS papers
FROM entities
WHERE entity_type = 'framework' AND confidence = 1.0
GROUP BY 1, 2, 3
ORDER BY papers DESC
LIMIT 15;

-- Which benchmarks does GPT-4o get evaluated on?
SELECT e2.entity AS benchmark, COUNT(*) AS papers
FROM entities e1
JOIN entities e2 ON e1.paper_id = e2.paper_id
WHERE e1.entity = 'GPT-4o'
  AND e2.entity_type = 'benchmark'
  AND e1.confidence = 1.0
  AND e2.confidence = 1.0
GROUP BY 1
ORDER BY papers DESC;

-- Provider market share by month (trend analysis)
SELECT DATE_TRUNC('month', ingest_date) AS month,
       provider,
       COUNT(DISTINCT paper_id) AS papers
FROM entities
WHERE entity_type = 'model' AND confidence = 1.0
GROUP BY 1, 2
ORDER BY 1, papers DESC;

-- LLM research bleeding into non-core CS categories
SELECT p.primary_category, COUNT(DISTINCT p.paper_id) AS papers
FROM papers p
WHERE p.is_llm = true AND p.is_other_cs = true
GROUP BY 1
ORDER BY papers DESC;

-- Model adoption velocity — which models grew fastest in 2025?
SELECT entity, provider, COUNT(1) knt
FROM entities
WHERE entity_type = 'model' AND confidence = 1.0
GROUP BY 1, 2
ORDER BY knt DESC
LIMIT 15;

Confidence scoring

Entity extraction uses a deterministic rule-based pipeline with a curated YAML knowledge base — fully auditable, no LLM black box.

Score	Meaning
1.0	Unambiguous alias match, or ambiguous alias confirmed by positive context
0.2	Ambiguous alias — matched but context inconclusive
0.1	Likely false positive — negative context matched

Always filter WHERE confidence = 1.0 for highest-precision results.

Ambiguous tokens (e.g. "Flux" in physics papers, "Ray" in computer vision) are handled with positive/negative context windows — the methodology is documented in the dataset repository.

Coverage

Date range: January 2024 – present (daily updates)

Categories: cs.LG · cs.CL · cs.CV · cs.AI · cs.NE · cs.CR · cs.SE · cs.DC · cs.IR · cs.HC · cs.RO · cs.MA · cs.SD · cs.DB · cs.GT · cs.PL · stat.ML · eess.IV · eess.SP · q-bio.QM

Entity counts (Jan 2024 – Jun 2026):

Type	Total entities	Unique entities	Example
model	~65,000	100+	GPT-4o, Llama 3.1, Qwen3, DeepSeek R1
framework	~55,000	40+	PyTorch, DeepSpeed, vLLM, LoRA
benchmark	~15,000	50+	MMLU, HumanEval, GSM8K, SWE-bench
dataset	~4,000	30+	The Pile, FineWeb, LAION

Category breakdown (entity rows):

Category	Entities	Description
cs.CL	27,452	Computation and Language
cs.LG	20,040	Machine Learning
cs.CV	19,790	Computer Vision
cs.AI	9,167	Artificial Intelligence
cs.CR	3,951	Security and Cryptography
cs.SE	3,763	Software Engineering

Pricing and access

Tier	Coverage	Format	Price
Free sample	3 × 10-day windows (Jan 2024, Jan 2025, Jun 2026)	DuckDB file (this repo)	Free
Snapshot	Jan 2024 – Jun 2026, full	Parquet on S3	$49 one-time
Subscription	Jan 2024 – present + daily updates	Parquet on S3, daily delta	$29/month

What you get with the full dataset

• All 7 tables in Parquet format, partitioned by ingest_date
• aws s3 sync access — pipe directly into your existing data stack
• 2.5 years of history with daily updates going forward
• Schema-stable with a changelog — no breaking changes without notice
• Daily new partition by 9am UTC

How to purchase

→ Snapshot — $49 one-time
Full dataset, Jan 2024 – Jun 2026. Download once, use forever.

→ Subscription — $29/month
Full dataset + daily updates. Cancel anytime.

After purchase you will receive AWS credentials by email within 24 hours. Download with:

```bash aws configure --profile arxiv-ai-graph aws s3 sync s3://arxiv-ai-graph/snapshot/ ./arxiv-data/ --profile arxiv-ai-graph ```

For enterprise pricing, custom date ranges, or questions: [nowudoit@gmail.com]

License

Metadata (titles, abstracts, authors, categories) derived from arXiv under CC0 1.0 Public Domain. Entity extraction, normalization, confidence scoring, and schema are original work licensed under CC BY 4.0.

Attribution: arXiv AI Entity Graph by arxgraph, CC BY 4.0
https://huggingface.co/datasets/arxgraph/arxiv-ai-entity-graph