/Tech4h ago

ICML Position Paper Calls for Studying AI Training Dynamics

2100114

Original post

Our ICML oral (Tuesday!) position paper argues that empirical science of DL and future model improvements both rely on understanding the training process, not just analyzing or manipulating a fully trained model.

Stella Biderman@BlancheMinerva

In film, "we'll fix it in post" is what you say when something went wrong on set and you don't want to redo it. AI research has made it our entire methodology: train the model, then patch whatever comes out. Our new ICML oral argues this can't be the basis of a science of AI. 🧵

2:59 PM · Jun 28, 2026 · 68 Views

Sentiment

Sentiment building, check back later.

Cluster Engagement

Digg Deeper

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

Posts from X

Most Activity

Naomi Saphra@nsaphra

Our ICML HiLD workshop paper shows that the reason why bigger models learn more complex tasks is because they are able to saturate the gradients for easier tasks; different tasks are competing for the same parameters and gradient mass.

4h91

REPLIES1

Naomi Saphra@nsaphra

Our ICML position paper argues that interpretability not only SHOULD be actionable, but CAN be---just follow our framework for targeting actionable insights.

Hadas Orgad@OrgadHadas

Our ICML 2025 workshop on Actionable Interpretability drew massive interest. But the same questions kept coming up: What does "actionable" mean? Is it achievable? How? We're ready to answer. 🧵

4h4600

Naomi Saphra@nsaphra

Our ICML mechinterp workshop paper demonstrates how feature geometry can lead to model failures, and analyzing that geometry can help us to efficiently build adversarial test sets based on concept combinations.

4h85