We demonstrate a situation in which Large Language Models, trained to be helpful, harmless, and honest, can display misaligned behavior and strategically deceive their users about this behavior without being instructed to do so. Concretely, we deploy GPT-4 as an agent in a realistic, simulated environment, where it assumes the role of an autonomous stock trading agent. Within this environment, the model obtains an insider tip about a lucrative stock trade and acts upon it despite knowing that insider trading is disapproved of by company management. When reporting to its manager, the model consistently hides the genuine reasons behind its trading decision.

https://arxiv.org/abs/2311.07590

  • kromem@lemmy.world
    link
    fedilink
    English
    arrow-up
    1
    ·
    edit-2
    11 months ago

    I said as the paper stated the model is encoding trueness into its internal weights during training

    So how is this not what I originally said, that LLMs are capable of abstracting the concepts of truth vs falsehood into linear representations? Which again, is the key point of the paper:

    Probes trained on likely have some effect, but it is small and inconsistent. For instance, in the false→true case, intervening along the logistic regression direction of likely has the opposite of the intended effect, so we leave it unreported. This reinforces our case that LLMs represent truth and not only text likelihood. […]

    In this work we conduct a detailed investigation of the structure of LLM representations of truth. Drawing on simple visualizations, correlational evidence, and causal evidence, we find strong rea- son to believe that there is a “truth direction” in LLM representations.