Siyuan Ma

New 'forked futures' method reveals reusable causal interfaces in LLMs

Researchers have introduced a new method called "forked futures" to identify reusable causal interfaces within language models. This approach compares states based on the response distributions they induce from sampled …