🌊 Submission Tsunami at NeurIPS 2025: Is Peer Review About to Collapse?

[Joanne]

Headline number — NeurIPS 2025’s main track received roughly 25,000 submissions.

That's 60 times the submission volume of 2010, nearly triple since 2019, and a staggering compound annual growth rate (CAGR) of about 29% since 2017 (3,240 submissions).

---

Why It Matters

Pressure Point	What 25,000 submissions really mean…
Reviewer overload 🥵	30 minutes per paper = 12,500 person hours — over 6 years full-time. A 2-hour review? Try 50+ person-years.
Quality signal-to-noise	Reports of rushed literature reviews, LLM rewritten old ideas, shaky benchmarks, and duplicates are up. ~20,000 papers may be rejected.
Corporate gravity	Industry labs are driving submission growth, shifting focus toward product oriented, compute heavy research.
Carbon cost	Reproducibility runs + thousands of flights = a large environmental footprint. Calls grow for greener workflows.

---

If Growth Continues…

• The first 1 million submissions/year may arrive around 2040 (25,000 × 1.29¹⁵ ≈ 1.1M).
• By ~2075, NeurIPS could theoretically receive one paper per human on Earth.
• Maintaining today’s 6 paper per reviewer workload would require 20,000+ active reviewers.

---

️ Emerging Counter-Measures

Symptom	Experiments Underway
Crushing review load	Paid professional reviewers, AI-assisted triage, multi-stage screen-then-review pipelines
Reproducibility gaps	Shared GPU clusters for automatic verification, stronger code/data release mandates
Conference bloat	Rolling deadlines, and specialised sub-tracks
Carbon footprint	Regional hubs, default-virtual attendance, and energy disclosure in reviews

---

Quick Facts to Drop in Conversation

• 29% CAGR → ~5 million submissions/year by mid-2040s if nothing changes
• 6 full-time years just to skim this year’s submissions; 3× more for deep review
• NeurIPS was called “NIPS” until 2018
• Acceptance rate has remained steady at 20 to 25% despite the submission explosion
• Reviewer travel + GPU training now factor into AI’s carbon ledger

---

Bottom Line

The milestone of 25,000 submissions at NeurIPS 2025 is both thrilling and alarming. It highlights the global appetite for AI progress, but also exposes the strain on a volunteer-run, deadline-driven mega-conference. Unless we rethink publication and validation workflows, peer review may soon collapse under its own success.

[root]

Saw this picture, pretty fun

[Tari S]

Was this written by an LLM, lol? "30 minutes per paper = 12,500 person hours — over 6 years full-time. A 2-hour review? Try 50+ person-years."
30 minutes per paper = 12,500 person hours — 1,4 PERSON YEARS, AND 2-HOUR REVIEWS ARE X4, SO 5.7 YEARS, NOT 50

[root]

@Tari-S Thanks for pointing out the confusion.

I'd like to clarify the diff between calendar hours and full-time work hours (FTE). A person-year in reviewing is typically ~2,000–2,080 work hours, not 8,760 calendar hours.

Three scenario:

• Skim pass (30 min/paper):
  25,000 × 0.5 h = 12,500 h → 6.0–6.5 FTE-years (12,500 ÷ 2,080 ≈ 6.01; ÷ 2,000 = 6.25).

• Single in-depth review (2 h/paper):
  25,000 × 2 h = 50,000 h → 24.0–26.0 FTE-years (not 50+ yet).

• What would make it 50+?
  Typical programs use ≥3 reviews/paper.
  3 × 2 h = 6 h/paper ⇒ 25,000 × 6 h = 150,000 h → 72–78 FTE-years.
  Add meta-review/AC time (≈0.5–1 h/paper) and you’re at ~78–91 FTE-years.

So:

• “1.4 person-years” uses the wrong denominator (calendar hours).
• “50+ person-years” is too high for a single pass, but reasonable (and actually conservative) once you include multiple reviews and overhead.