Search papers, labs, and topics across Lattice.
1
0
3
2
By enforcing graph isomorphism across counterfactual inputs, UGID reveals that debiasing LLMs can be achieved by directly manipulating internal representations and attention mechanisms.
