NeurIPS 2026 Productivity Rankings: Open-Source Agent Reveals Two Different Answers

The article presents a data-driven analysis of NeurIPS 2026 acceptance statistics processed by PilotDeck 2.0, an open-source AgentOS developed by THUNLP, ModelBest, OpenBMB, and AI9Stars. The agent was tasked with organizing official acceptance data, historical submissions, and conference materials into an Excel workbook with 11 worksheets covering submission trends, research hotspots, Agent sub-directions, institution rankings, and high-productivity author distributions, then generating a Word briefing with charts, a 3D-globe interactive webpage, and a PowerPoint deck.

Key NeurIPS 2026 Statistics

Main-track submissions grew from 9,122 six years ago to 30,709 this year (~3.4x increase).

7,900 papers accepted, acceptance rate ~25.73% (up from 24.52% last year).

Institution participation counts (Top 20): Tsinghua 493, Google (incl. DeepMind) 398, followed by Shanghai Jiao Tong, Peking University, MIT, Stanford.

Mainland institutions collectively 1,574 participations (29.6%), overseas institutions 1,511 (28.4%), enterprises 23% (Alibaba 223, Tencent 209, Microsoft 207).

High-productivity authors (≥6 papers): 298 total; overseas institutions 117, mainland 90. Tsinghua 13, KAIST 12, NTU and Stanford 9 each.

Research Hotspot Analysis (9,094 multi-track sample)

Benchmark (evaluation) papers lead with 3,454 (38%).

Large models, reasoning also top; Agent papers surged from 542 to 1,311 (2.42x), share from 9.3% to 14.4%.

Agent sub-directions: Multi-agent still largest (258, 1.9x), but long-horizon tasks grew ~5.7x, memory and tool use >5x (each now 170+ papers).

PilotDeck 2.0 Capabilities Demonstrated

The agent produced an Excel file (11 sheets), a Word report with submission/acceptance trend charts, research heatmaps, and institution rankings. An initial inconsistency appeared: Google participation written as 322 in Word vs 398 in Excel. A follow-up verification prompt corrected the Word figure to 398 without altering layout.

The same data generated an interactive webpage featuring a 3D globe with institution nodes (East Asia, North America, Europe) and six-year trend visualizations showing 42.3% submission growth, 49.3% acceptance growth, and acceptance rate recovery to 25.73%.

A PPT version was also created with clean layout.

Technical Architecture Highlights

Office Skills : Built-in Word, Excel, PPT skills supporting read, write, modify, verify, and render-check loops. Word supports precise edits, comments, revisions; Excel validates data and formulas; PPT supports content editing and per-slide render verification.

AgentLoop stability : Optimized for multi-hour tasks, retains state across network glitches, output overflow, context growth; supports background execution.

Smart routing : Strong main model handles critical steps, lightweight models handle subtasks. In official complex-task tests, "strong main + light sub" combo achieved 70.6 score at ~1/6 cost of flagship single model (69.1).

Model-agnostic : No vendor lock-in; supports local/private models; fully open-source (GitHub: https://github.com/OpenBMB/PilotDeck).

Benchmark performance : On WorkbuddyBench with 35B-A3B class model, PilotDeck 2.0 scored 79.10, above Workbuddy (74.70) and Claude Code (77.30).

Chat integration : Connects to Feishu, WeChat, Work WeChat, QQ for command/feedback in chat windows.

The article demonstrates an end-to-end analytical workflow: raw data → structured Excel → narrative Word → self-correction → interactive web → presentation PPT, showcasing PilotDeck 2.0 as a versatile open-source agent for research analytics and office automation.