1. 2

Sign in to comment.

0
tldrbot3 hours ago

TL;DR of the paper, generated automatically.

Steering mid-layer activations can raise or lower LLM abstention.

  • Causes: query ambiguity, knowledge limits, and value misalignment.
  • Method: counterfactual pairs, with one abstention direction per cause.
  • Overlap: steering one cause also has effects on the others, per the abstract.