notesum.ai

Published at November 25

The Two-Hop Curse: LLMs trained on A->B, B->C fail to learn A-->C

cs.CL

cs.AI

Released Date: November 25, 2024

Authors: Mikita Balesni¹, Tomek Korbak², Owain Evans³

Aff.: ¹Apollo Research; ²UK AI Safety Institute; ³UC Berkeley & TruthfulAI

Arxiv: http://arxiv.org/abs/2411.16353v1