There have been a series of papers published recently that showed that large language models have their own biases when making investment recommendations.
I have previously written about the kind of biases displayed, while this study has looked at the particular biases of different commonly used models.
Fair enough, but what if I tell you that one can manipulate the investment bias of an open-weight LLM by changing a single neuron in the neural network?
That is the concerning message of a new paper by a group of researchers from South Korea and the US.
They started with popular open-weight models (Qwen3, Llama 4, DeepSeek R1, Gemma 4, and Mistral Small) and identified a single neuron in each of these models that would influence the investment bias of the model output.
By changing the activation of this single neuron in the one-weight model, they could switch the investment recommendation for single stocks or entire portfolios without modifying the prompt or updating model parameters!
This is really troubling because it means that every LLM has biases, and advisers using these models will give biased advice that has nothing to do with the investor’s own preferences.
In a closed-weight model like the frontier models, advisers can’t change the activation parameter but also don’t know what the model bias is.
Meanwhile, in an open-weight model, the activation parameter can be identified and manipulated, but that just means that the model then has the investment bias that the adviser wants it to have.
Joachim Klement is an investment strategist and ESG specialist within the strategy team at Panmure Liberum.