Stability of value-based preferences under systematic perturbations

Ascertain how value-based preferences in large language models behave under systematic perturbations, including whether such preferences remain stable across population-level variations induced by model dropout.

Background

For human-adjacent applications such as human-robot interaction, reliable and stable preferences are essential. The authors note uncertainty about whether any learned value-based preferences would persist under systematic perturbations (e.g., Monte Carlo dropout) applied to create model populations, which they use to assess brittleness.

References

Further, if a model has value-based preferences (VBPs), it is unclear how these preferences will fair under systematic perturbation.

— Do Large Language Models Learn Human-Like Strategic Preferences?  (2404.08710 - Roberts et al., 2024) in Section 3, Do LLMs Prefer Strategies Based on Value?, opening paragraph

These results open many future research directions into models' value stability. Examples of open questions include: "Which types of stability are more important for which types of usecases?", "Are different models specialized in different types of stabilty needed, or can one model by highly stable in all types?", "How can we explore these types of stability in more detail?", and many more.

— Large Language Models as Superpositions of Cultural Perspectives  (2307.07870 - Kovač et al., 2023) in Appendix Section 'Systematic comparison of models on different types of value stability (Mean-level, rank-order, ipsative),' paragraph following Table 4