MODULATING REASONING IN LLMS

Summary from the TLDR AI newsletter

Researchers show that reasoning performance in LLMs can be adjusted via simple interventions in the residual stream using representation-based control vectors.