Andy Liu

andyliu.bsky.social

16 October 2026 13:45

LLM developers train their models to follow prosocial values, but post-training can influence model behavior in unexpected ways. In our new paper "Predicting Alignment Generalization with Value Representations", we ask how well downstream effects of value alignment can be predicted in advance.

Andy Liu

andyliu.bsky.social

16 October 2026 13:45

Our work introduces an alignment generalization prediction task. Given a set of values, we fine-tune single-value checkpoints on each value in our set, then evaluate how this changes each individual checkpoint’s adherence toward all other values in the set.

Andy Liu

andyliu.bsky.social

16 October 2026 13:45

We extract 66 values from Claude’s Constitution, and compute a generalization matrix, where each row represents a value that we fine-tuned on, and each column represents a value that we evaluate the fine-tuned checkpoint on. Matrices are similar across model families, sizes, and fine-tuning methods.

Andy Liu

andyliu.bsky.social

16 October 2026 13:45

We benchmark different value representation methods by how well they predict alignment generalization effects. We find that methods based on model activations when applying values in context (such as persona vectors) significantly outperform methods based on textual descriptions of the values.

Andy Liu

andyliu.bsky.social

16 October 2026 13:45

We hypothesize that value representations that perform better at our task can also support the empirical study of alignment targets. To support this, we study multi-value alignment targets, finding that more coherent targets (i.e. whose values have similar representations) are more robust.