Paper locates and steers Llama 3.1 'opportunity recognition' dial

Researchers built 636 matched scenario pairs to see whether a large language model contains anything resembling an "opportunity recognition dial," then turned that dial up and down and watched the model's judgments move. The preprint on arXiv, submitted September 14 by Christian Fisch, Angela Altmeier, Martin Obschonka, Michal Kosinski, and Pin Ni, gives the underlying line of work a name: "artificial entrepreneurial cognition, the functional organisation of entrepreneurship-relevant representations and computations inside artificial intelligence (AI) systems."

The primary target was Llama 3.1 8B-Instruct. Rather than infer the construct from outputs, the team recovered an OR direction in the model's internal representations and intervened on it directly. "To our knowledge, this is the first causal intervention on an internal representation of an entrepreneurship construct inside an LLM," they write.

There is a wrinkle. The opportunity recognition direction is "distinct from the opportunity evaluation and exploitation directions" on the geometry, but "steering it also shifts judgments about these neighbouring stages." Separable and leaky at the same time.

Recovery, signed steering, and geometric separation held across four additional LLMs spanning different scales and families, which the authors argue gives the contested distinction between opportunity recognition and evaluation "a concrete representational form inside AI systems." Two researchers on our Who's Who list posted the preprint.