Who’s the most “productive” at NeurIPS 2026? We ran them all with open‑source agents.

Which papers are the most “productive” at NeurIPS 2026? We ran an open‑source agent to find out

Submissions exceed 30,000 articles; which research areas are hottest at this year’s NeurIPS?

Tsinghua 493, Google 398.

In the count of institutional participation in accepted papers at NeurIPS 2026, Tsinghua ranks first, with Google close behind. Shanghai Jiao Tong University, Peking University, MIT, Stanford and other universities also rank among the top.

Among the top 20 institutions, mainland universities and institutes have the highest total participation count. However, when counting the number of prolific authors, overseas universities and institutes rank ahead.

This year, there are 298 prolific authors with six or more accepted papers. By institution type, overseas universities and institutions account for 117, and mainland universities and institutions account for 90.

In the past six NeurIPS conferences, the number of submissions to the main track grew from 9,122 papers to 30,709 papers, expanding to about 3.4 times. This year, 7,900 papers were accepted, with an acceptance rate still around one quarter.

This time, we handed the NeurIPS 2026 related materials to PilotDeck 2.0, an open-source AgentOS jointly developed by Tsinghua University THUNLP, Mianbi Intelligence, OpenBMB, and AI9Stars.

NeurIPS data, handed over to AgentOS

We give the official NeurIPS acceptance data, historical submission statistics, and related conference materials to PilotDeck 2.0, letting it first compile an Excel data sheet, then generate a Word briefing.

The task also specifically emphasizes that the primary track admission numbers must not be mixed with multi‑track analysis samples, and the key figures in Word must also match the underlying Excel table.

The generated Excel consists of 11 worksheets, covering submission trends, research hotspots, Agent sub‑domains, institution rankings, and high‑output author distribution.

The Word briefing summarizes the main statistics and includessubmission acceptance trends,research hotspotsandinstitution rankingsand other charts.

After the document was completed, we found that in the initial Word version, the number of Google (including DeepMind) participations was written as 322, which is inconsistent with the 398 in the Excel source table.

So, we added a verification request to the original conversation, please PilotDeck 2.0 Compare the two documents to check numbers, conclusions, and charts, only modify the problematic parts.

In the revised Word document, Google’s participation count has been corrected to 398, consistent with the Excel source table and the corresponding conclusions in the report, the original layout remains unchanged.

Beyond the documentation, there are interactive pages

We also used the same dataset to let PilotDeck 2.0 created a NeurIPS 2026 special webpage.

When the page loads, a 3D globe occupies the first screen. Nodes of institutions such as Tsinghua, Shanghai Jiao Tong, and Peking University are concentrated in East Asia. Switching to the North America and Europe view, you can also see nodes of local universities and tech companies.

The page also shows the submission and acceptance trends over the past six years.

This year the NeurIPS main tracksaw a 42.3% increase in submissions, and the acceptance volume grew by 49.3%. The acceptance rate also rose from last year’s 24.52% rebounded to 25.73%.

Among the 9,094 paper samples covering multiple tracks, based on keywords in titles and abstracts,Benchmark (benchmark)has the most related papers, totaling 3,454, about 38% of the sample.

Large models, inference and other areas also rank among the top, Agent growth is even more pronounced.

The number of agent‑related papers increased from 542 last year to 1,311, a 2.42‑fold rise..The share in the sample also rose from 9..3% to 14..4%.

Among the 14 subfields of Agent,Multi-agentstill leads with 258 papers, about 1.9 times last year,growth rate lower than the overall Agent.

Long‑term tasksThe number of related papers has reached about 5.7 times,memory and tool invocation also exceed 5 times. Last year these areas had only a few dozen related papers; this year they have already reached over 170 papers.

According to Top 20 institution categories summary, mainland universities and institutions participated a total of 1,574 times, while overseas universities and institutions participated 1,511 times, a difference of only 63 times. They account for 29.6% and 28.4%, overall very close.

Enterprises account for 23% of the participation count among the top 20 institutions. Alibaba 223 times, Tencent 209 times, close to Microsoft’s 207 times.

According to institutional statistics of prolific authors, Tsinghua has 13, KAIST has 12, and Nanyang Technological University and Stanford each have 9.

Core Capabilities of the Open-source AgentOS

Supporting these tasks is PilotDeck 2.0.This open‑source AgentOS.

As the little powerhouse in the Harness domain, it is responsible for coordinating models and tools so that data reading, file processing, result verification, and subsequent modifications can be carried out seamlessly.

PilotDeck 2.0’scode is completely open source, allowing users to view data transmission and processing details.

It does not bind to a specific model vendor or fixed model, allowing free integration of models from different vendors, and also supports local or private models. Developers can also modify and extend it as needed.

For complex tasks, beyond file generation and modification, we also need to consider runtime stability and model invocation costs.

PilotDeck 2.0 also optimizes these two aspects AgentLoop andIntelligent Routing, making it easier for tasks to continue when exceptions occur and allowing cost control through model-based division of labor.

Office File Processing Performance Optimization

PilotDeck 2.0 includes Office Skills for Word, Excel, and PowerPoint, supporting file reading, organizing, generating, modifying, and checking; files delivered can also continue to be edited.Office Skills

Among them, Word Skill supports precise editing, commenting, and revision; Excel Skill can check data and formulas; PPT Skill supports content editing and page‑by‑page rendering verification.

After the document is completed, the Agent can also perform checks using the text, document structure, calculation results, and rendered images, and continue to make corrections when issues are found.

We also had PilotDeck 2.0 turn the same batch of NeurIPS data into a PPT, with a relatively clean overall layout.

In the official WorkbuddyBench office test, using 35B–A3B level model, PilotDeck 2.0 achieved 79.10 points, higher than Workbuddy’s 74.70 points and Claude Code’s 77.30 points.

More Stable AgentLoop

Complex office tasks can sometimes take several hours. PilotDeck 2.0 has optimized AgentLoop, enabling better retention of task state when encountering network fluctuations, overly long outputs, or continuously growing context, and allowing continuation after exceptions, reducing the need to restart from scratch.

With the background task mechanism, the Agent can continuously process tasks, and users can come back later to view the results.

Intelligent Routing and Model Cost

PilotDeck 2.The intelligent routing of 0 lets the more capable main model handle critical steps, while some sub‑tasks are assigned to lightweight models to control model invocation costs.

In the official set of complex task tests, the “Strong‑main + Light‑lepton” model combination achieved about 1/6 of the cost and obtained a score of 70.6 points, slightly higher than the flagship single‑model solution’s 69.1 points.

Give PilotDeck 2.0 activities now can also be completed directly in common chat apps.

PilotDeck 2.0 now supports integration with Feishu, WeChat, Enterprise WeChat, and QQ, allowing commands to be sent and feedback received directly in the chat window.

Initial setup, file uploading, and result preview have also been optimized, and features such as historical conversation search and task branching make subsequent edits more convenient.

With the fully open‑source Harness and the ability to freely integrate different models, PilotDeck 2.0 becomes even more flexible in its usage. It can be used directly for everyday office work, and it also makes it easy for developers to modify and extend it according to their own needs.

Related open-source links

➤ GitHub: https://github.com/OpenBMB/PilotDeck

➤ Official website: https://pilotdeck.openbmb.cn/

·