ICLR went from 143 submissions to 62,301 submission IDs – what happened?

The dramatic increase in submission IDs at ICLR (International Conference on Learning Representations) from 143 to 62,301 reflects both the explosive growth of deep learning and a policy-driven shift in how submissions are handled. Here’s the full story:

🔍 The Core Issue: Submission Counting

ICLR’s use of OpenReview introduced a system where every initiated submission, including withdrawn, desk-rejected, or incomplete papers, is assigned a unique submission ID. This inflates the total count significantly, as the 62,301 figure includes all such IDs, not just reviewed papers. The actual number of papers reviewed annually is closer to ~7,000-11,000, with acceptance rates around 25-30%.

📊 Policy Change: Iterative Submissions

In 2024, ICLR adopted a policy allowing multiple versions of the same paper to be submitted under separate IDs. Each version—whether a major revision, minor update, or resubmission—received a new ID and was treated as a distinct submission. This aimed to encourage iterative improvement and reduce redundant reviews, but it also led to concerns about system gaming and artificial inflation of submission counts.

🧠 Growth Drivers

1. ·Deep learning dominance: From a niche subfield in 2013, representation learning became central to AI, transforming ICLR into one of the top ML conferences.
2. ·Global expansion: Increased participation from regions like China and India, alongside the rise of industry labs (DeepMind, Google, OpenAI, etc.), significantly boosted submissions.
3. ·Lower barriers: Tools like PyTorch and Hugging Face democratized research, enabling more submissions.
4. ·AI hype cycle: The release of large language models (e.g., ChatGPT) accelerated the influx of submissions.

⚠️ Consequences

The 62,301 figure drew widespread criticism for being misleading, as it didn’t reflect the actual number of reviewed papers. It also increased the reviewer workload and raised fairness concerns, with critics arguing that papers with multiple revisions had an unfair advantage.

🔄 ICLR’s Response

ICLR clarified that while multiple versions are allowed, only the final version counts toward acceptance, and reviewers assess the final submitted version. This addresses some concerns but hasn’t fully restored confidence in the submission process.

✅ Bottom Line

The jump in submission IDs is partly an artifact of OpenReview’s ID system and partly a result of ICLR’s iterative submission policy. While it captures the field’s growth, it also highlights how metrics can be misleading when incentivized by policy changes. The lesson: conferences must balance innovation with clarity to maintain trust in their processes.