Papers With Code is the closest thing the AI research community has to a canonical reference book. Started by Robert Stojnic in 2018, acquired by Meta a couple of years later, and now hosting leaderboards for thousands of distinct tasks, it does one thing exceptionally well: pair a paper with its code, its dataset, and its benchmark numbers in a single page.
If you've ever read an arXiv paper that claimed SOTA on some benchmark and wondered "compared to what, exactly?", PwC is where you go to find out.
- Comprehensive coverage of academic AI tasks and benchmarks
- Each leaderboard links to paper, code repo, and dataset
- Open data — you can pull any leaderboard via the API
- Free with no signup or paywall
- Updated by the research community as papers get published
- Dataset section is a useful catalog in its own right
- Indexes papers, not commercial products — frontier closed models often missing
- Coverage of newer tasks (multimodal, long-context) lags older ones
- Listings depend on community submission; gaps in less-active subfields
- Doesn't capture deployment metrics like price or latency
- Some leaderboards haven't been updated in 18+ months
- Researchers tracking SOTA on a specific task
- Engineers picking a baseline model from open research
- Students learning what benchmarks exist for an area
- Anyone who needs the paper-code-dataset triad in one click
- You're picking a closed commercial model (use LMArena or Artificial Analysis)
- You need real-world deployment numbers (use Artificial Analysis)
- You're looking for product-quality reviews (use this site)
Pricing
Full site access. All leaderboards, papers, datasets, and code links. No account needed.
What it is
Each task on Papers With Code has its own page with a leaderboard at the top. The leaderboard is sorted by published score, with rows linking out to:
- The paper that reported the result (arXiv link, PDF, abstract)
- The code repository (GitHub, usually)
- The dataset (with license and download instructions)
- A "compare" view that lets you sort, filter, and chart against alternatives
Multiply that by thousands of tasks — image classification, machine translation, abstractive summarisation, code generation, math reasoning, video question answering — and you get a research-grade map of "what's the best published result on X."
Where it shines
For researchers, it answers a specific question better than any other site: what is the current state of the art on this task, and what code reproduces it? Not "what does a vendor claim is best." Not "what's trending on Twitter." The published-result-with-code-and-dataset triad.
This is irreplaceable for two cases:
Picking a baseline. If you're starting research on a new task, PwC tells you which prior work to beat. The leaderboard tells you the current SOTA; the linked code lets you run it; the linked dataset gives you the test set.
Tracking subfield progress. If you want to know whether a specific eval has gotten easier (because models have caught up) or harder (because someone proposed a more rigorous version), the dated leaderboard is the trail of breadcrumbs.
For engineers, it's still useful, but more selectively. The leaderboards index research papers, which means closed-source production models often don't appear. GPT-5.5 might be the de-facto SOTA on a task; PwC's leaderboard might not list it because OpenAI didn't publish a paper.
Where it falls short
Closed-model coverage has thinned. When OpenAI, Anthropic, and Google were publishing detailed technical reports (GPT-3, GPT-4, the original Claude technical report), their numbers landed on the relevant leaderboards. As those labs have moved toward shorter, less-detailed system cards, fewer of their numbers make it onto PwC. The result: some leaderboards look outdated even when the field hasn't stopped moving.
Submission is community-driven. Hot subfields stay current; quieter ones don't. A 2024 paper might still sit at the top of a niche leaderboard not because nothing has improved but because no one has submitted the 2025 results.
No deployment metrics. PwC tells you which model scored 91.2% on MMLU. It doesn't tell you that the same model costs $50/M tokens or runs at 30 tok/s. For that, Artificial Analysis is the right reference.
Newer evals lag. Long-context benchmarks, multi-turn tool use, agent evals — the kind of thing that matters most for 2026 production AI — have less complete coverage on PwC than the established 2018-2022 vintage tasks.
How to use it
Two flows.
"What's SOTA on X?" Search for the task name (or the closest published equivalent). Open the leaderboard. Click the top row's paper. Done.
"What's been published in this subfield this year?" Browse by area (NLP / Computer Vision / RL / Audio), filter by year, sort by paper-citation count. The "trends" page is also useful for spotting subfields that are heating up.
The API is documented and stable; if you're building research infrastructure, you can pull leaderboards programmatically rather than scraping HTML.
Verdict
Papers With Code is essential if you do AI research. It is one of three benchmark references a serious AI engineer should know — PwC for academic tasks and SOTA tracking, LMArena for crowdsourced commercial-model preference, Artificial Analysis for deployment economics. None of them substitute for the others. PwC is the oldest, the most thorough, and the only one where the canonical link to the underlying research is one click away.
