/Tech6h ago

ICML awards honorable mention to Google DeepMind co-authored paper measuring exact language model memorization

Story Overview

A Google DeepMind co-authored paper earned an ICML 2026 Outstanding Paper Honorable Mention for a method that isolates unintended memorization of specific training points from broader generalization, yielding a capacity estimate of roughly 3.6 bits per parameter in GPT-style models.

452584.6K

#428

Original post

Kamalika Chaudhuri@kamalikac#428inTech

Honored that our paper on memorization with @jxmnop, @csitawarin @srush_nlp Chuan Guo, Narine Kokhlikyan, Edward Suh and Saeed Mahloujifar got an Outstanding Paper Honorable mention at ICML this year! Most of the credit goes to @jxmnop and Saeed. Paper: https://arxiv.org/abs/2505.24832

ICML Conference@icmlconf

🏆Announcing the #ICML2026 Awards! 🏆 Including Outstanding Papers (research paper & position paper, winner & honorable mentions) and the Test of Time Award!

Check out the blog post for all winners (or read on), laudatio, & description of the processes. https://blog.icml.cc/?p=1347

2:15 AM · Jul 6, 2026 · 4K Views

Capacity Insight

Models fill up on specifics before they generalize

Hundreds of transformers ranging from 500K to 1.5B parameters were trained on growing datasets; memorization climbed until capacity was reached, after which unintended retention dropped as generalization rose.

Methodology Advance

New split lets researchers count what the model truly retains

By subtracting generalization effects the approach produces a direct measure of model capacity and new scaling laws that tie size and data volume to membership inference risk.

Sentiment

Many users congratulated the memorization paper authors on earning the ICML Outstanding Paper Honorable Mention because they viewed the work as insightful and well deserved.

Pos

100.0%

Neg

0.0%

4 comments with sentiment.

Cluster Engagement

Digg Deeper

No Digg Deeper questions have been answered for this story yet.