ERM Fixes Causal Rung Collapse in LLMs

What Changed

Formalizes rung collapse as lack of gradient for P(Y|do(X)) vs P(Y|X)

Why It Matters

Addresses core reasoning flaws in LLMs, enabling better generalization and steerability. Could prevent entrenchment in production models, improving reliability across domains. Inverse scaling in steerability highlights need for targeted causal fixes.

What To Do Next

Prioritize whether this update affects your current workflow this week.