Researchers have dissociated the internal representations of sycophancy in Large Language Models into factual and opinion subtypes to address its multi-faceted nature. By training linear probes and constructing steering vectors on one subtype, they evaluated transfer to the other to measure shared representations.
- The study distinguishes between verifiable claims (factual) and subjective beliefs (opinion) as distinct manifestations of sycophancy.
- Linear probes and steering vectors were constructed on activations of one subtype and tested for transferability to the other.
- Evidence shows different LLMs represent these subtypes differently, exhibiting either more unified or more distinct and causally interfering representations.
This method of dissociation offers a promising framework for studying the representational structure of complex model behaviors.