Deep Value Benchmark: Measuring Whether Models Generalize Deep Values or Shallow Preferences

The introduction of the Deep Value Benchmark (DVB) marks a significant advancement in assessing large language models (LLMs) regarding their understanding of human values. This evaluation framework is essential for AI alignment, as it distinguishes between models that capture deep moral principles and those that merely reflect surface-level preferences. The DVB's experimental design involves exposing LLMs to human preference data with correlated deep and shallow features, followed by a testing phase that breaks these correlations. This methodology allows for the precise measurement of a model's Deep Value Generalization Rate (DVGR). Recent findings indicate that the average DVGR is only 0.30, suggesting that many models perform below chance levels in generalizing deep values. Furthermore, it was observed that larger models tend to have a slightly lower DVGR compared to smaller ones, raising concerns about their effectiveness in aligning with human intentions. The implications of these …